{ "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": 6, "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": 19400.285000000003, "pct_cuda_time": 97.63493415199191, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 22.911, "pct_cuda_time": 0.11530315025559092, "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": 22.911, "pct_cuda_time": 0.11530315025559092, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 19368.734000000004, "pct_cuda_time": 97.47614886572269, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 532.882, "pct_cuda_time": 2.6818110651870186, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 10.559, "pct_cuda_time": 0.05313980025091808, "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": 522.3229999999999, "pct_cuda_time": 2.6286712649361, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 4668.268000000001, "pct_cuda_time": 23.493780569916936, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 2196.005, "pct_cuda_time": 11.05173473340443, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.520000000000007, "pct_cuda_time": 0.11836803692599619, "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": 2172.4850000000006, "pct_cuda_time": 10.933366696478435, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 341.37500000000006, "pct_cuda_time": 1.718022474728399, "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": 341.37500000000006, "pct_cuda_time": 1.718022474728399, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 696.474, "pct_cuda_time": 3.505113101615487, "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": 168.57199999999995, "pct_cuda_time": 0.84836465649188, "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": 483.0059999999999, "pct_cuda_time": 2.430802382800922, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 44.89600000000002, "pct_cuda_time": 0.2259460623226839, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 1434.4139999999998, "pct_cuda_time": 7.218910260168615, "invocations": 32 }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 1434.4139999999998, "pct_cuda_time": 7.218910260168615, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 14167.583999999999, "pct_cuda_time": 71.30055723061871, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 8648.041000000001, "pct_cuda_time": 43.522603589520784, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.805000000000007, "pct_cuda_time": 0.11980234349589028, "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": 8624.236000000003, "pct_cuda_time": 43.4028012460249, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 1085.044, "pct_cuda_time": 5.460651711663715, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 1085.044, "pct_cuda_time": 5.460651711663715, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 4434.499, "pct_cuda_time": 22.317301929434226, "invocations": 32 }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 4434.499, "pct_cuda_time": 22.317301929434226, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 8.64, "pct_cuda_time": 0.04348213601363124, "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": 8.64, "pct_cuda_time": 0.04348213601363124, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 351.419, "pct_cuda_time": 1.768570457844245, "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": 5.023, "pct_cuda_time": 0.025279024212554364, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.737, "pct_cuda_time": 0.003709066463199794, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 345.659, "pct_cuda_time": 1.7395823671684907, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 118.525, "pct_cuda_time": 0.5964953901638476, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.3759999999999994, "pct_cuda_time": 0.027055551297370545, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 4.544, "pct_cuda_time": 0.022868382644206056, "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": 5.536, "pct_cuda_time": 0.02786077603836372, "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": 35.039, "pct_cuda_time": 0.17633918562287326, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 28.607, "pct_cuda_time": 0.14396915103494778, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 1.952, "pct_cuda_time": 0.009823741840116688, "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": 6.208, "pct_cuda_time": 0.031242719950535038, "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": 28.191, "pct_cuda_time": 0.14187556670836554, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 3.072, "pct_cuda_time": 0.015460315027068886, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 82958.251, "cuda_time_us": 19400.285000000003, "pct_cuda_time": 97.63493415199191, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 352.024, "cuda_time_us": 22.911, "pct_cuda_time": 0.11530315025559092, "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": 22.911, "pct_cuda_time": 0.11530315025559092, "trace": "index_select(bfloat16[128256, 4096], 0, int64[768]) <- embedding(bfloat16[128256, 4096], int64[768], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4209.524, "cuda_time_us": 611.061, "pct_cuda_time": 3.075258971600176, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 239.897, "cuda_time_us": 10.559, "pct_cuda_time": 0.05313980025091808, "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": 10.559, "pct_cuda_time": 0.05313980025091808, "trace": "_C::rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3099.865, "cuda_time_us": 145.052, "pct_cuda_time": 0.729996619565884, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 471.102, "cuda_time_us": 68.319, "pct_cuda_time": 0.34382593174945286, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.583, "pct_cuda_time": 0.34012189794088427, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1005.874, "cuda_time_us": 10.816, "pct_cuda_time": 0.05443319249113837, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.816, "pct_cuda_time": 0.05443319249113837, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1076.456, "cuda_time_us": 21.502, "pct_cuda_time": 0.10821213988021977, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.183, "pct_cuda_time": 0.026084248953547537, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.008, "pct_cuda_time": 0.07553008070515944, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.006597810221512796, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 294.837, "cuda_time_us": 44.415, "pct_cuda_time": 0.2235253554450731, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.415, "pct_cuda_time": 0.2235253554450731, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 121.791, "cuda_time_us": 8.448, "pct_cuda_time": 0.04251586632443944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.448, "pct_cuda_time": 0.04251586632443944, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 628.502, "cuda_time_us": 447.002, "pct_cuda_time": 2.249606685458934, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 243.479, "cuda_time_us": 270.525, "pct_cuda_time": 1.36145889410736, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.789, "pct_cuda_time": 1.3577548602987914, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 138.734, "cuda_time_us": 33.951, "pct_cuda_time": 0.1708636573841197, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.951, "pct_cuda_time": 0.1708636573841197, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 164.811, "cuda_time_us": 142.526, "pct_cuda_time": 0.7172841339674545, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 142.526, "pct_cuda_time": 0.7172841339674545, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2558.609, "cuda_time_us": 604.345, "pct_cuda_time": 3.0414596630969877, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.87, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1845.043, "cuda_time_us": 146.495, "pct_cuda_time": 0.7372587401987163, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.235, "cuda_time_us": 69.15100000000001, "pct_cuda_time": 0.3480131004026174, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.415, "pct_cuda_time": 0.3443090665940488, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 535.987, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 795.887, "cuda_time_us": 21.888, "pct_cuda_time": 0.11015474456786581, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.312, "pct_cuda_time": 0.02673346140097328, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.07601321554975535, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.563, "cuda_time_us": 44.832, "pct_cuda_time": 0.22562397242628657, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.832, "pct_cuda_time": 0.22562397242628657, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.225, "cuda_time_us": 8.448, "pct_cuda_time": 0.04251586632443944, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.448, "pct_cuda_time": 0.04251586632443944, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.487, "cuda_time_us": 441.274, "pct_cuda_time": 2.220779639731379, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.811, "cuda_time_us": 269.02, "pct_cuda_time": 1.353884748887393, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.284, "pct_cuda_time": 1.3501807150788243, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.694, "cuda_time_us": 33.92, "pct_cuda_time": 0.17070764509055228, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.92, "pct_cuda_time": 0.17070764509055228, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.556, "cuda_time_us": 138.334, "pct_cuda_time": 0.6961872457534334, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.334, "pct_cuda_time": 0.6961872457534334, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2647.601, "cuda_time_us": 602.584, "pct_cuda_time": 3.032597158291431, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.438, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1872.395, "cuda_time_us": 145.375, "pct_cuda_time": 0.7316221670117641, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.237, "cuda_time_us": 67.935, "pct_cuda_time": 0.34189339237106925, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.199, "pct_cuda_time": 0.33818935856250065, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.434, "cuda_time_us": 10.688, "pct_cuda_time": 0.05378901269834384, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.688, "pct_cuda_time": 0.05378901269834384, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 800.507, "cuda_time_us": 22.080000000000002, "pct_cuda_time": 0.11112101425705763, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.408, "pct_cuda_time": 0.027216596245569186, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.232, "pct_cuda_time": 0.0766573953425499, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007247022668938539, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 210.182, "cuda_time_us": 44.672, "pct_cuda_time": 0.22481874768529334, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.672, "pct_cuda_time": 0.22481874768529334, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.176, "cuda_time_us": 8.319, "pct_cuda_time": 0.041866653877013695, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.319, "pct_cuda_time": 0.041866653877013695, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 506.356, "cuda_time_us": 440.66599999999994, "pct_cuda_time": 2.2177197857156044, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 175.237, "cuda_time_us": 269.085, "pct_cuda_time": 1.3542118714384215, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.317, "pct_cuda_time": 1.3503467926816544, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.304, "cuda_time_us": 33.503, "pct_cuda_time": 0.16860902810933884, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.503, "pct_cuda_time": 0.16860902810933884, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 167.847, "cuda_time_us": 138.078, "pct_cuda_time": 0.6948988861678442, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.078, "pct_cuda_time": 0.6948988861678442, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2514.525, "cuda_time_us": 603.45, "pct_cuda_time": 3.036955437202057, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.141, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1812.187, "cuda_time_us": 145.536, "pct_cuda_time": 0.7324324244073884, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.689, "cuda_time_us": 68.447, "pct_cuda_time": 0.34447011154224744, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.711, "pct_cuda_time": 0.3407660777336788, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 536.545, "cuda_time_us": 10.56, "pct_cuda_time": 0.05314483290554929, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.56, "pct_cuda_time": 0.05314483290554929, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 748.856, "cuda_time_us": 21.889, "pct_cuda_time": 0.11015977722249701, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.408, "pct_cuda_time": 0.027216596245569186, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.07585217060155672, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.409, "pct_cuda_time": 0.0070910103753711136, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.267, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.858, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 470.682, "cuda_time_us": 441.37, "pct_cuda_time": 2.2212627745759748, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 177.263, "cuda_time_us": 269.757, "pct_cuda_time": 1.3575938153505929, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.021, "pct_cuda_time": 1.3538897815420246, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.634, "cuda_time_us": 33.727, "pct_cuda_time": 0.16973634274672925, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.727, "pct_cuda_time": 0.16973634274672925, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.14, "cuda_time_us": 137.886, "pct_cuda_time": 0.6939326164786525, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.886, "pct_cuda_time": 0.6939326164786525, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2376.593, "cuda_time_us": 605.304, "pct_cuda_time": 3.046285978888315, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.445, "cuda_time_us": 8.16, "pct_cuda_time": 0.04106646179065173, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.16, "pct_cuda_time": 0.04106646179065173, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1689.821, "cuda_time_us": 146.78199999999998, "pct_cuda_time": 0.7387031120778726, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.036, "cuda_time_us": 69.375, "pct_cuda_time": 0.3491404150400078, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.639, "pct_cuda_time": 0.34543638123143916, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.77, "cuda_time_us": 10.848, "pct_cuda_time": 0.054594237439337, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.848, "pct_cuda_time": 0.054594237439337, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 710.893, "cuda_time_us": 21.568, "pct_cuda_time": 0.10854429508587947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.28, "pct_cuda_time": 0.026572416452774646, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.008, "pct_cuda_time": 0.07553008070515944, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.69, "cuda_time_us": 44.991, "pct_cuda_time": 0.2264241645126485, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.991, "pct_cuda_time": 0.2264241645126485, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 93.199, "cuda_time_us": 8.288, "pct_cuda_time": 0.04171064158344626, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.288, "pct_cuda_time": 0.04171064158344626, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.427, "cuda_time_us": 442.074, "pct_cuda_time": 2.2248057634363447, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.94, "cuda_time_us": 270.524, "pct_cuda_time": 1.361453861452729, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.788, "pct_cuda_time": 1.3577498276441604, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.59, "cuda_time_us": 33.408, "pct_cuda_time": 0.16813092591937412, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.408, "pct_cuda_time": 0.16813092591937412, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.195, "cuda_time_us": 138.142, "pct_cuda_time": 0.6952209760642415, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.142, "pct_cuda_time": 0.6952209760642415, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2754.525, "cuda_time_us": 606.232, "pct_cuda_time": 3.0509562823860756, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.322, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2034.263, "cuda_time_us": 145.375, "pct_cuda_time": 0.7316221670117641, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.694, "cuda_time_us": 68.415, "pct_cuda_time": 0.3443090665940488, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.679, "pct_cuda_time": 0.34060503278548016, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 728.387, "cuda_time_us": 10.496, "pct_cuda_time": 0.052822743009152025, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.496, "pct_cuda_time": 0.052822743009152025, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 784.681, "cuda_time_us": 21.824, "pct_cuda_time": 0.10983265467146854, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.376, "pct_cuda_time": 0.027055551297370552, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.944, "pct_cuda_time": 0.07520799080876218, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007569112565335809, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.367, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.983, "cuda_time_us": 8.159, "pct_cuda_time": 0.041061429136020525, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.159, "pct_cuda_time": 0.041061429136020525, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.813, "cuda_time_us": 444.28200000000004, "pct_cuda_time": 2.2359178648620506, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 180.166, "cuda_time_us": 272.06, "pct_cuda_time": 1.3691840189662634, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 271.324, "pct_cuda_time": 1.3654799851576949, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.872, "cuda_time_us": 33.92, "pct_cuda_time": 0.17070764509055228, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.92, "pct_cuda_time": 0.17070764509055228, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.268, "cuda_time_us": 138.302, "pct_cuda_time": 0.6960262008052347, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.302, "pct_cuda_time": 0.6960262008052347, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2372.49, "cuda_time_us": 604.823, "pct_cuda_time": 3.0438652720107044, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.542, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1700.088, "cuda_time_us": 146.07800000000003, "pct_cuda_time": 0.7351601232175029, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.744, "cuda_time_us": 69.43900000000001, "pct_cuda_time": 0.34946250493640507, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.703, "pct_cuda_time": 0.3457584711278365, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 488.011, "cuda_time_us": 10.592, "pct_cuda_time": 0.05330587785374793, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.592, "pct_cuda_time": 0.05330587785374793, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 739.384, "cuda_time_us": 21.728, "pct_cuda_time": 0.10934951982687265, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.184, "pct_cuda_time": 0.026089281608178746, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.07569112565335809, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007569112565335809, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 175.288, "cuda_time_us": 44.319, "pct_cuda_time": 0.2230422206004772, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.319, "pct_cuda_time": 0.2230422206004772, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.577, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 461.631, "cuda_time_us": 442.26599999999996, "pct_cuda_time": 2.225772033125536, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.93, "cuda_time_us": 269.88399999999996, "pct_cuda_time": 1.3582329624887561, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.116, "pct_cuda_time": 1.3543678837319888, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.689, "cuda_time_us": 33.952, "pct_cuda_time": 0.1708686900387509, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.952, "pct_cuda_time": 0.1708686900387509, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.791, "cuda_time_us": 138.43, "pct_cuda_time": 0.6966703805980292, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.43, "pct_cuda_time": 0.6966703805980292, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2406.608, "cuda_time_us": 604.216, "pct_cuda_time": 3.0408104506495617, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 109.515, "cuda_time_us": 8.544, "pct_cuda_time": 0.042999001169035336, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.544, "pct_cuda_time": 0.042999001169035336, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1677.156, "cuda_time_us": 145.663, "pct_cuda_time": 0.7330715715455518, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.086, "cuda_time_us": 68.574, "pct_cuda_time": 0.3451092586804107, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.839, "pct_cuda_time": 0.34141025752647336, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.332, "cuda_time_us": 10.528, "pct_cuda_time": 0.05298378795735066, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.528, "pct_cuda_time": 0.05298378795735066, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 722.888, "cuda_time_us": 21.857, "pct_cuda_time": 0.10999873227429838, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.408, "pct_cuda_time": 0.027216596245569186, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.07601321554975535, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.345, "pct_cuda_time": 0.0067689204789738435, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.99, "cuda_time_us": 44.704, "pct_cuda_time": 0.22497979263349202, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.704, "pct_cuda_time": 0.22497979263349202, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.67, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.294, "cuda_time_us": 441.753, "pct_cuda_time": 2.223190281299727, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.317, "cuda_time_us": 269.628, "pct_cuda_time": 1.356944602903167, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.893, "pct_cuda_time": 1.3532456017492296, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.805, "cuda_time_us": 34.111, "pct_cuda_time": 0.17166888212511286, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.111, "pct_cuda_time": 0.17166888212511286, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.03, "cuda_time_us": 138.014, "pct_cuda_time": 0.694576796271447, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.014, "pct_cuda_time": 0.694576796271447, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2535.457, "cuda_time_us": 608.664, "pct_cuda_time": 3.0631956984491717, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.437, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1784.31, "cuda_time_us": 144.958, "pct_cuda_time": 0.7295235500305506, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 180.869, "cuda_time_us": 68.447, "pct_cuda_time": 0.34447011154224744, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.711, "pct_cuda_time": 0.3407660777336788, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.218, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.361, "cuda_time_us": 21.44, "pct_cuda_time": 0.10790011529308494, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.088, "pct_cuda_time": 0.025606146763582843, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.07585217060155672, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 183.483, "cuda_time_us": 44.447, "pct_cuda_time": 0.22368640039327173, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.447, "pct_cuda_time": 0.22368640039327173, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.434, "cuda_time_us": 8.319, "pct_cuda_time": 0.041866653877013695, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.319, "pct_cuda_time": 0.041866653877013695, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 527.601, "cuda_time_us": 447.19499999999994, "pct_cuda_time": 2.250577987802757, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.883, "cuda_time_us": 270.52399999999994, "pct_cuda_time": 1.3614538614527285, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.756, "pct_cuda_time": 1.3575887826959616, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 110.731, "cuda_time_us": 34.304, "pct_cuda_time": 0.17264018446893592, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.304, "pct_cuda_time": 0.17264018446893592, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 203.146, "cuda_time_us": 142.367, "pct_cuda_time": 0.7164839418810924, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 142.367, "pct_cuda_time": 0.7164839418810924, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2376.277, "cuda_time_us": 604.537, "pct_cuda_time": 3.0424259327861796, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.074, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1713.086, "cuda_time_us": 146.27100000000002, "pct_cuda_time": 0.7361314255613258, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.533, "cuda_time_us": 68.63900000000001, "pct_cuda_time": 0.34543638123143927, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.903, "pct_cuda_time": 0.3417323474228707, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 487.292, "cuda_time_us": 10.56, "pct_cuda_time": 0.05314483290554929, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.56, "pct_cuda_time": 0.05314483290554929, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 757.714, "cuda_time_us": 21.953, "pct_cuda_time": 0.11048186711889428, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.441, "pct_cuda_time": 0.027382673848399024, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.07569112565335809, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.549, "cuda_time_us": 45.119, "pct_cuda_time": 0.22706834430544304, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.119, "pct_cuda_time": 0.22706834430544304, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.63, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.433, "cuda_time_us": 441.946, "pct_cuda_time": 2.22416158364355, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.85, "cuda_time_us": 269.82, "pct_cuda_time": 1.357910872592359, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.085, "pct_cuda_time": 1.3542118714384215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.962, "cuda_time_us": 34.24, "pct_cuda_time": 0.17231809457253863, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.24, "pct_cuda_time": 0.17231809457253863, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.588, "cuda_time_us": 137.886, "pct_cuda_time": 0.6939326164786525, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.886, "pct_cuda_time": 0.6939326164786525, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2516.478, "cuda_time_us": 603.959, "pct_cuda_time": 3.0395170584093414, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.131, "cuda_time_us": 8.32, "pct_cuda_time": 0.0418716865316449, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.32, "pct_cuda_time": 0.0418716865316449, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1809.87, "cuda_time_us": 146.526, "pct_cuda_time": 0.7374147524922837, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.509, "cuda_time_us": 68.64, "pct_cuda_time": 0.3454414138860704, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.904, "pct_cuda_time": 0.3417373800775018, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 521.416, "cuda_time_us": 10.976, "pct_cuda_time": 0.05523841723213155, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.976, "pct_cuda_time": 0.05523841723213155, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 770.884, "cuda_time_us": 22.175, "pct_cuda_time": 0.11159911644702232, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.28, "pct_cuda_time": 0.026572416452774646, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.264, "pct_cuda_time": 0.07681844029074852, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.631, "pct_cuda_time": 0.008208259703499138, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.177, "cuda_time_us": 44.735, "pct_cuda_time": 0.22513580492705942, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.735, "pct_cuda_time": 0.22513580492705942, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.266, "cuda_time_us": 8.064, "pct_cuda_time": 0.04058332694605582, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.064, "pct_cuda_time": 0.04058332694605582, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.339, "cuda_time_us": 441.049, "pct_cuda_time": 2.219647292439357, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.848, "cuda_time_us": 269.532, "pct_cuda_time": 1.3564614680585712, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.796, "pct_cuda_time": 1.3527574342500026, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.551, "cuda_time_us": 33.631, "pct_cuda_time": 0.16925320790213336, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.631, "pct_cuda_time": 0.16925320790213336, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.981, "cuda_time_us": 137.886, "pct_cuda_time": 0.6939326164786525, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.886, "pct_cuda_time": 0.6939326164786525, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2460.829, "cuda_time_us": 605.5900000000001, "pct_cuda_time": 3.047725318112841, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.134, "cuda_time_us": 8.479, "pct_cuda_time": 0.04267187861800686, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.479, "pct_cuda_time": 0.04267187861800686, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1757.05, "cuda_time_us": 146.174, "pct_cuda_time": 0.7356432580620987, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 159.1, "cuda_time_us": 68.863, "pct_cuda_time": 0.3465636958688296, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.127, "pct_cuda_time": 0.34285966206026103, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.587, "cuda_time_us": 10.752, "pct_cuda_time": 0.054111102594741105, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.752, "pct_cuda_time": 0.054111102594741105, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 712.169, "cuda_time_us": 21.727, "pct_cuda_time": 0.10934448717224143, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.311, "pct_cuda_time": 0.026728428746342077, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.07601321554975535, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006602842876144003, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 180.025, "cuda_time_us": 44.832, "pct_cuda_time": 0.22562397242628657, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.832, "pct_cuda_time": 0.22562397242628657, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.222, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.514, "cuda_time_us": 442.71400000000006, "pct_cuda_time": 2.2280266624003175, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 180.082, "cuda_time_us": 270.908, "pct_cuda_time": 1.3633864008311125, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 270.172, "pct_cuda_time": 1.359682367022544, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.777, "cuda_time_us": 33.792, "pct_cuda_time": 0.17006346529775776, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.792, "pct_cuda_time": 0.17006346529775776, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.017, "cuda_time_us": 138.014, "pct_cuda_time": 0.694576796271447, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.014, "pct_cuda_time": 0.694576796271447, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2630.49, "cuda_time_us": 604.952, "pct_cuda_time": 3.0445144844581304, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.153, "cuda_time_us": 8.32, "pct_cuda_time": 0.0418716865316449, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.32, "pct_cuda_time": 0.0418716865316449, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1931.061, "cuda_time_us": 146.238, "pct_cuda_time": 0.735965347958496, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.901, "cuda_time_us": 69.087, "pct_cuda_time": 0.3476910105062201, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.351, "pct_cuda_time": 0.3439869766976515, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.707, "cuda_time_us": 10.848, "pct_cuda_time": 0.054594237439337, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.848, "pct_cuda_time": 0.054594237439337, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 963.053, "cuda_time_us": 21.6, "pct_cuda_time": 0.10870534003407811, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.12, "pct_cuda_time": 0.025767191711781476, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.168, "pct_cuda_time": 0.07633530544615262, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006602842876144003, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.564, "cuda_time_us": 44.703, "pct_cuda_time": 0.22497475997886085, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.703, "pct_cuda_time": 0.22497475997886085, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.221, "cuda_time_us": 8.288, "pct_cuda_time": 0.04171064158344626, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.288, "pct_cuda_time": 0.04171064158344626, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.354, "cuda_time_us": 442.106, "pct_cuda_time": 2.224966808384543, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.748, "cuda_time_us": 269.725, "pct_cuda_time": 1.3574327704023943, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.003709066463199794, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.988, "pct_cuda_time": 1.3537237039391945, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.605, "cuda_time_us": 34.111, "pct_cuda_time": 0.17166888212511286, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.111, "pct_cuda_time": 0.17166888212511286, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.324, "cuda_time_us": 138.27, "pct_cuda_time": 0.6958651558570361, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.27, "pct_cuda_time": 0.6958651558570361, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2465.988, "cuda_time_us": 605.178, "pct_cuda_time": 3.045651864404783, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.731, "cuda_time_us": 8.319, "pct_cuda_time": 0.041866653877013695, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.319, "pct_cuda_time": 0.041866653877013695, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1803.448, "cuda_time_us": 145.50400000000002, "pct_cuda_time": 0.7322713794591899, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.672, "cuda_time_us": 68.32, "pct_cuda_time": 0.343830964404084, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.003709066463199794, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.583, "pct_cuda_time": 0.34012189794088427, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 557.335, "cuda_time_us": 10.432, "pct_cuda_time": 0.05250065311275476, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.432, "pct_cuda_time": 0.05250065311275476, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 748.725, "cuda_time_us": 22.112000000000002, "pct_cuda_time": 0.11128205920525626, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.503, "pct_cuda_time": 0.02769469843553388, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.041, "pct_cuda_time": 0.0756961583079893, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.007891202461733077, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.102, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.952, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 446.464, "cuda_time_us": 443.227, "pct_cuda_time": 2.2306084142261264, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.567, "cuda_time_us": 270.877, "pct_cuda_time": 1.3632303885375452, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 270.141, "pct_cuda_time": 1.3595263547289766, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.162, "cuda_time_us": 34.176, "pct_cuda_time": 0.17199600467614135, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.176, "pct_cuda_time": 0.17199600467614135, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.263, "cuda_time_us": 138.174, "pct_cuda_time": 0.6953820210124402, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.174, "pct_cuda_time": 0.6953820210124402, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2395.617, "cuda_time_us": 605.495, "pct_cuda_time": 3.047247215922876, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.976, "cuda_time_us": 8.287, "pct_cuda_time": 0.041705608928815065, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.287, "pct_cuda_time": 0.041705608928815065, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1689.493, "cuda_time_us": 147.453, "pct_cuda_time": 0.7420800233354129, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.084, "cuda_time_us": 68.894, "pct_cuda_time": 0.3467197081623971, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.159, "pct_cuda_time": 0.3430207070084597, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 501.993, "cuda_time_us": 11.296, "pct_cuda_time": 0.05684886671411788, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 11.296, "pct_cuda_time": 0.05684886671411788, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.12, "cuda_time_us": 22.048000000000002, "pct_cuda_time": 0.110959969308859, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.248, "pct_cuda_time": 0.026411371504576012, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.136, "pct_cuda_time": 0.07617426049795398, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.008374337306328979, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 177.173, "cuda_time_us": 45.215, "pct_cuda_time": 0.22755147915003898, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.215, "pct_cuda_time": 0.22755147915003898, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.198, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 483.329, "cuda_time_us": 441.49899999999997, "pct_cuda_time": 2.2219119870234003, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.328, "cuda_time_us": 269.405, "pct_cuda_time": 1.3558223209204077, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.637, "pct_cuda_time": 1.3519572421636408, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.939, "cuda_time_us": 33.791, "pct_cuda_time": 0.17005843264312653, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.791, "pct_cuda_time": 0.17005843264312653, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.007, "cuda_time_us": 138.303, "pct_cuda_time": 0.696031233459866, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.303, "pct_cuda_time": 0.696031233459866, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2431.163, "cuda_time_us": 602.872, "pct_cuda_time": 3.034046562825219, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.24, "cuda_time_us": 8.511, "pct_cuda_time": 0.042832923566205494, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.511, "pct_cuda_time": 0.042832923566205494, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1762.434, "cuda_time_us": 145.02300000000002, "pct_cuda_time": 0.7298506725815792, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 160.456, "cuda_time_us": 68.287, "pct_cuda_time": 0.34366488680125423, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.551, "pct_cuda_time": 0.33996085299268564, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.07, "cuda_time_us": 10.561, "pct_cuda_time": 0.0531498655601805, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.561, "pct_cuda_time": 0.0531498655601805, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.731, "cuda_time_us": 21.568, "pct_cuda_time": 0.10854429508587947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.184, "pct_cuda_time": 0.026089281608178746, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.07601321554975535, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.666, "cuda_time_us": 44.607, "pct_cuda_time": 0.2244916251342649, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.607, "pct_cuda_time": 0.2244916251342649, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.191, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 448.008, "cuda_time_us": 441.11499999999995, "pct_cuda_time": 2.2199794476450165, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.911, "cuda_time_us": 269.30899999999997, "pct_cuda_time": 1.355339186075812, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.541, "pct_cuda_time": 1.3514741073190446, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.133, "cuda_time_us": 33.984, "pct_cuda_time": 0.17102973498694954, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.984, "pct_cuda_time": 0.17102973498694954, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.06, "cuda_time_us": 137.822, "pct_cuda_time": 0.6936105265822552, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.822, "pct_cuda_time": 0.6936105265822552, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2438.707, "cuda_time_us": 602.9670000000001, "pct_cuda_time": 3.034524665015184, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.394, "cuda_time_us": 8.512, "pct_cuda_time": 0.042837956220836707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.512, "pct_cuda_time": 0.042837956220836707, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1769.558, "cuda_time_us": 145.43800000000002, "pct_cuda_time": 0.7319392242535302, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.082, "cuda_time_us": 68.799, "pct_cuda_time": 0.3462416059724324, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.063, "pct_cuda_time": 0.3425375721638638, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.61, "cuda_time_us": 10.432, "pct_cuda_time": 0.05250065311275476, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.432, "pct_cuda_time": 0.05250065311275476, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 784.68, "cuda_time_us": 21.568, "pct_cuda_time": 0.10854429508587947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.344, "pct_cuda_time": 0.02689450634917192, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.944, "pct_cuda_time": 0.07520799080876218, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.675, "cuda_time_us": 44.639, "pct_cuda_time": 0.22465267008246353, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.639, "pct_cuda_time": 0.22465267008246353, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.482, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.713, "cuda_time_us": 440.761, "pct_cuda_time": 2.2181978879055695, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.564, "cuda_time_us": 269.019, "pct_cuda_time": 1.3538797162327618, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.284, "pct_cuda_time": 1.3501807150788243, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.254, "cuda_time_us": 34.144, "pct_cuda_time": 0.17183495972794272, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.144, "pct_cuda_time": 0.17183495972794272, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.103, "cuda_time_us": 137.598, "pct_cuda_time": 0.6924832119448648, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.598, "pct_cuda_time": 0.6924832119448648, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2328.873, "cuda_time_us": 604.921, "pct_cuda_time": 3.0443584721645633, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.921, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1642.292, "cuda_time_us": 145.791, "pct_cuda_time": 0.7337157513383463, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.132, "cuda_time_us": 68.703, "pct_cuda_time": 0.3457584711278365, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.967, "pct_cuda_time": 0.3420544373192679, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 466.443, "cuda_time_us": 10.784, "pct_cuda_time": 0.054272147542939735, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.784, "pct_cuda_time": 0.054272147542939735, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 702.598, "cuda_time_us": 21.76, "pct_cuda_time": 0.10951056477507128, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.024, "pct_cuda_time": 0.025284056867185576, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.232, "pct_cuda_time": 0.0766573953425499, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007569112565335809, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 179.289, "cuda_time_us": 44.544, "pct_cuda_time": 0.22417456789249882, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.544, "pct_cuda_time": 0.22417456789249882, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.163, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.075, "cuda_time_us": 442.522, "pct_cuda_time": 2.227060392711125, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.875, "cuda_time_us": 270.332, "pct_cuda_time": 1.360487591763537, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.596, "pct_cuda_time": 1.3567835579549685, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.153, "cuda_time_us": 33.664, "pct_cuda_time": 0.1694192855049632, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.664, "pct_cuda_time": 0.1694192855049632, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.228, "cuda_time_us": 138.526, "pct_cuda_time": 0.6971535154426253, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.526, "pct_cuda_time": 0.6971535154426253, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2453.531, "cuda_time_us": 602.839, "pct_cuda_time": 3.0338804852223897, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.25, "cuda_time_us": 8.096, "pct_cuda_time": 0.040744371894254464, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.096, "pct_cuda_time": 0.040744371894254464, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1767.45, "cuda_time_us": 145.534, "pct_cuda_time": 0.732422359098126, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 176.923, "cuda_time_us": 68.416, "pct_cuda_time": 0.34431409924867995, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.648, "pct_cuda_time": 0.3404490204919127, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 548.726, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 715.21, "cuda_time_us": 21.791, "pct_cuda_time": 0.10966657706863871, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.28, "pct_cuda_time": 0.026572416452774646, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.039, "pct_cuda_time": 0.07568609299872688, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.139, "cuda_time_us": 44.703, "pct_cuda_time": 0.22497475997886085, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.703, "pct_cuda_time": 0.22497475997886085, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.197, "cuda_time_us": 8.287, "pct_cuda_time": 0.041705608928815065, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.287, "pct_cuda_time": 0.041705608928815065, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.473, "cuda_time_us": 440.922, "pct_cuda_time": 2.219008145301194, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 174.788, "cuda_time_us": 269.565, "pct_cuda_time": 1.3566275456614012, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.797, "pct_cuda_time": 1.3527624669046339, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.932, "cuda_time_us": 33.663, "pct_cuda_time": 0.16941425285033196, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.663, "pct_cuda_time": 0.16941425285033196, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.214, "cuda_time_us": 137.694, "pct_cuda_time": 0.6929663467894606, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.694, "pct_cuda_time": 0.6929663467894606, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2624.326, "cuda_time_us": 604.504, "pct_cuda_time": 3.0422598551833495, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.425, "cuda_time_us": 8.576, "pct_cuda_time": 0.04316004611723398, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.04316004611723398, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1938.872, "cuda_time_us": 144.893, "pct_cuda_time": 0.7291964274795222, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.204, "cuda_time_us": 68.223, "pct_cuda_time": 0.3433427969048569, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.487, "pct_cuda_time": 0.3396387630962883, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.21, "cuda_time_us": 10.527, "pct_cuda_time": 0.05297875530271944, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.527, "pct_cuda_time": 0.05297875530271944, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 972.729, "cuda_time_us": 21.536, "pct_cuda_time": 0.10838325013768084, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.28, "pct_cuda_time": 0.026572416452774646, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_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": 0.07536903575696083, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.151, "cuda_time_us": 44.607, "pct_cuda_time": 0.2244916251342649, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.607, "pct_cuda_time": 0.2244916251342649, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 92.238, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.452, "cuda_time_us": 442.779, "pct_cuda_time": 2.228353784951346, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.055, "cuda_time_us": 269.726, "pct_cuda_time": 1.3574378030570255, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.705, "pct_cuda_time": 0.0035480215150011604, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.021, "pct_cuda_time": 1.3538897815420246, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.403, "cuda_time_us": 34.175, "pct_cuda_time": 0.17199097202151012, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.175, "pct_cuda_time": 0.17199097202151012, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.85, "cuda_time_us": 138.878, "pct_cuda_time": 0.69892500987281, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.878, "pct_cuda_time": 0.69892500987281, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2393.463, "cuda_time_us": 604.92, "pct_cuda_time": 3.0443534395099316, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.033, "cuda_time_us": 8.255, "pct_cuda_time": 0.04154456398061643, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.255, "pct_cuda_time": 0.04154456398061643, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1713.1, "cuda_time_us": 144.287, "pct_cuda_time": 0.7261466387730106, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.06, "cuda_time_us": 67.87100000000001, "pct_cuda_time": 0.34157130247467204, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.135, "pct_cuda_time": 0.3378672686661034, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 482.703, "cuda_time_us": 10.368, "pct_cuda_time": 0.05217856321635749, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.368, "pct_cuda_time": 0.05217856321635749, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.406, "cuda_time_us": 21.408, "pct_cuda_time": 0.1077390703448863, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.12, "pct_cuda_time": 0.025767191711781476, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.008, "pct_cuda_time": 0.07553008070515944, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.07, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.64, "pct_cuda_time": 0.22465770273709476, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.067, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.379, "cuda_time_us": 444.154, "pct_cuda_time": 2.235273685069256, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.968, "cuda_time_us": 271.932, "pct_cuda_time": 1.3685398391734689, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 271.197, "pct_cuda_time": 1.3648408380195316, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.577, "cuda_time_us": 33.824, "pct_cuda_time": 0.17022451024595636, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.824, "pct_cuda_time": 0.17022451024595636, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.928, "cuda_time_us": 138.398, "pct_cuda_time": 0.6965093356498306, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.398, "pct_cuda_time": 0.6965093356498306, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2419.005, "cuda_time_us": 604.44, "pct_cuda_time": 3.0419377652869524, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.988, "cuda_time_us": 8.48, "pct_cuda_time": 0.04267691127263807, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.48, "pct_cuda_time": 0.04267691127263807, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1717.3, "cuda_time_us": 146.079, "pct_cuda_time": 0.735165155872134, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 169.992, "cuda_time_us": 68.51100000000001, "pct_cuda_time": 0.34479220143864475, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.775, "pct_cuda_time": 0.3410881676300761, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 497.468, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.624, "pct_cuda_time": 0.05346692280194656, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 725.269, "cuda_time_us": 21.857, "pct_cuda_time": 0.10999873227429838, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.472, "pct_cuda_time": 0.027538686141966452, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.07488590091236491, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007574145219967016, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.91, "cuda_time_us": 45.087, "pct_cuda_time": 0.22690729935724444, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.087, "pct_cuda_time": 0.22690729935724444, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.386, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.026, "cuda_time_us": 441.65700000000004, "pct_cuda_time": 2.2227071464551313, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.134, "cuda_time_us": 270.748, "pct_cuda_time": 1.3625811760901192, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.98, "pct_cuda_time": 1.3587160973333523, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 123.5, "cuda_time_us": 33.535, "pct_cuda_time": 0.16877007305753744, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.535, "pct_cuda_time": 0.16877007305753744, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.879, "cuda_time_us": 137.374, "pct_cuda_time": 0.6913558973074743, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.374, "pct_cuda_time": 0.6913558973074743, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2304.474, "cuda_time_us": 604.024, "pct_cuda_time": 3.03984418096037, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.314, "cuda_time_us": 8.064, "pct_cuda_time": 0.04058332694605582, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.064, "pct_cuda_time": 0.04058332694605582, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1649.186, "cuda_time_us": 145.43800000000002, "pct_cuda_time": 0.7319392242535302, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.485, "cuda_time_us": 68.063, "pct_cuda_time": 0.3425375721638638, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.327, "pct_cuda_time": 0.3388335383552952, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.659, "cuda_time_us": 10.592, "pct_cuda_time": 0.05330587785374793, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.592, "pct_cuda_time": 0.05330587785374793, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 706.545, "cuda_time_us": 21.888, "pct_cuda_time": 0.11015474456786581, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.088, "pct_cuda_time": 0.025606146763582843, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.0771405301871458, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 175.399, "cuda_time_us": 44.895, "pct_cuda_time": 0.22594102966805263, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.895, "pct_cuda_time": 0.22594102966805263, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.342, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.224, "pct_cuda_time": 0.041388551687049, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 439.768, "cuda_time_us": 442.298, "pct_cuda_time": 2.225933078073735, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.132, "cuda_time_us": 270.46099999999996, "pct_cuda_time": 1.3611368042109626, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.693, "pct_cuda_time": 1.3572717254541955, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.599, "cuda_time_us": 33.855, "pct_cuda_time": 0.1703805225395238, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.855, "pct_cuda_time": 0.1703805225395238, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 133.978, "cuda_time_us": 137.982, "pct_cuda_time": 0.6944157513232484, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.982, "pct_cuda_time": 0.6944157513232484, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2465.223, "cuda_time_us": 605.207, "pct_cuda_time": 3.0457978113890882, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 98.288, "cuda_time_us": 8.639, "pct_cuda_time": 0.04347710335900003, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.639, "pct_cuda_time": 0.04347710335900003, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1768.289, "cuda_time_us": 145.502, "pct_cuda_time": 0.7322613141499275, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.296, "cuda_time_us": 68.19099999999999, "pct_cuda_time": 0.3431817519566583, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0035429888603699528, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.487, "pct_cuda_time": 0.3396387630962883, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 513.295, "cuda_time_us": 10.752, "pct_cuda_time": 0.054111102594741105, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.752, "pct_cuda_time": 0.054111102594741105, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.751, "cuda_time_us": 21.407000000000004, "pct_cuda_time": 0.10773403769025511, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.151, "pct_cuda_time": 0.025923204005348904, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_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": 0.07536903575696083, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 184.242, "cuda_time_us": 45.152, "pct_cuda_time": 0.2272344219082729, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.152, "pct_cuda_time": 0.2272344219082729, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.266, "cuda_time_us": 8.321, "pct_cuda_time": 0.041876719186276105, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.321, "pct_cuda_time": 0.041876719186276105, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 444.0, "cuda_time_us": 442.745, "pct_cuda_time": 2.2281826746938846, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.551, "cuda_time_us": 270.555, "pct_cuda_time": 1.3616098737462965, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0036990011539373797, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.82, "pct_cuda_time": 1.357910872592359, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.305, "cuda_time_us": 34.336, "pct_cuda_time": 0.17280122941713452, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.336, "pct_cuda_time": 0.17280122941713452, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.849, "cuda_time_us": 137.854, "pct_cuda_time": 0.6937715715304539, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.854, "pct_cuda_time": 0.6937715715304539, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2460.4, "cuda_time_us": 603.511, "pct_cuda_time": 3.0372624291345605, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.865, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1780.963, "cuda_time_us": 145.117, "pct_cuda_time": 0.7303237421169125, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.773, "cuda_time_us": 68.415, "pct_cuda_time": 0.3443090665940488, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.679, "pct_cuda_time": 0.34060503278548016, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 490.624, "cuda_time_us": 10.496, "pct_cuda_time": 0.052822743009152025, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.496, "pct_cuda_time": 0.052822743009152025, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 788.77, "cuda_time_us": 21.407, "pct_cuda_time": 0.10773403769025508, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.248, "pct_cuda_time": 0.026411371504576012, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 14.879, "pct_cuda_time": 0.0748808682577337, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.214, "cuda_time_us": 44.799, "pct_cuda_time": 0.22545789482345668, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.799, "pct_cuda_time": 0.22545789482345668, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.705, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.192, "pct_cuda_time": 0.04122750673885036, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.271, "cuda_time_us": 441.9459999999999, "pct_cuda_time": 2.2241615836435495, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.251, "cuda_time_us": 269.59599999999995, "pct_cuda_time": 1.3567835579549683, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 268.828, "pct_cuda_time": 1.3529184791982012, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.149, "cuda_time_us": 34.592, "pct_cuda_time": 0.1740895890027236, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.592, "pct_cuda_time": 0.1740895890027236, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.749, "cuda_time_us": 137.758, "pct_cuda_time": 0.6932884366858579, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.758, "pct_cuda_time": 0.6932884366858579, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2369.297, "cuda_time_us": 602.967, "pct_cuda_time": 3.0345246650151836, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.607, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.128, "pct_cuda_time": 0.04090541684245309, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1697.747, "cuda_time_us": 146.04500000000002, "pct_cuda_time": 0.7349940456146731, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.072, "cuda_time_us": 68.863, "pct_cuda_time": 0.3465636958688296, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.127, "pct_cuda_time": 0.34285966206026103, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.792, "cuda_time_us": 10.464, "pct_cuda_time": 0.052661698060953395, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.464, "pct_cuda_time": 0.052661698060953395, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 750.684, "cuda_time_us": 21.855, "pct_cuda_time": 0.10998866696503597, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.184, "pct_cuda_time": 0.026089281608178746, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.168, "pct_cuda_time": 0.07633530544615262, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007564079910704601, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.811, "cuda_time_us": 44.863, "pct_cuda_time": 0.22577998471985397, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.863, "pct_cuda_time": 0.22577998471985397, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.472, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 455.419, "cuda_time_us": 440.538, "pct_cuda_time": 2.21707560592281, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.859, "cuda_time_us": 268.7, "pct_cuda_time": 1.3522742994054067, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 267.964, "pct_cuda_time": 1.3485702655968381, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.65, "cuda_time_us": 34.112, "pct_cuda_time": 0.17167391477974409, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.112, "pct_cuda_time": 0.17167391477974409, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.432, "cuda_time_us": 137.726, "pct_cuda_time": 0.6931273917376592, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 137.726, "pct_cuda_time": 0.6931273917376592, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2639.138, "cuda_time_us": 605.69, "pct_cuda_time": 3.0482285835759617, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.119, "cuda_time_us": 8.096, "pct_cuda_time": 0.040744371894254464, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.096, "pct_cuda_time": 0.040744371894254464, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1954.434, "cuda_time_us": 146.112, "pct_cuda_time": 0.7353312334749639, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.619, "cuda_time_us": 68.512, "pct_cuda_time": 0.3447972340932759, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.776, "pct_cuda_time": 0.34109320028470724, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 515.994, "cuda_time_us": 10.913, "pct_cuda_time": 0.05492135999036547, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.913, "pct_cuda_time": 0.05492135999036547, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 939.118, "cuda_time_us": 21.887, "pct_cuda_time": 0.1101497119132346, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.215, "pct_cuda_time": 0.02624529390174617, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.2, "pct_cuda_time": 0.07649635039435125, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.694, "cuda_time_us": 44.8, "pct_cuda_time": 0.22546292747808788, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.8, "pct_cuda_time": 0.22546292747808788, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.996, "cuda_time_us": 8.352, "pct_cuda_time": 0.04203273147984353, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.352, "pct_cuda_time": 0.04203273147984353, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.417, "cuda_time_us": 443.13, "pct_cuda_time": 2.2301202467268992, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.927, "cuda_time_us": 270.652, "pct_cuda_time": 1.3620980412455235, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.916, "pct_cuda_time": 1.3583940074369547, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.995, "cuda_time_us": 33.6, "pct_cuda_time": 0.16909719560856595, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.6, "pct_cuda_time": 0.16909719560856595, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.339, "cuda_time_us": 138.878, "pct_cuda_time": 0.69892500987281, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 138.878, "pct_cuda_time": 0.69892500987281, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2441.899, "cuda_time_us": 607.1289999999999, "pct_cuda_time": 3.055470573590268, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.257, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.0423548213762408, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1730.66, "cuda_time_us": 146.367, "pct_cuda_time": 0.7366145604059217, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.223, "cuda_time_us": 68.576, "pct_cuda_time": 0.3451193239896731, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.0035429888603699528, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.872, "pct_cuda_time": 0.3415763351293032, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 527.63, "cuda_time_us": 11.007, "pct_cuda_time": 0.055394429525698965, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 11.007, "pct_cuda_time": 0.055394429525698965, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 728.953, "cuda_time_us": 21.568, "pct_cuda_time": 0.10854429508587947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.248, "pct_cuda_time": 0.026411371504576012, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.07569112565335809, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.247, "cuda_time_us": 45.216, "pct_cuda_time": 0.22755651180467015, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.216, "pct_cuda_time": 0.22755651180467015, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 98.91, "cuda_time_us": 8.287, "pct_cuda_time": 0.041705608928815065, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.287, "pct_cuda_time": 0.041705608928815065, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.768, "cuda_time_us": 444.05899999999997, "pct_cuda_time": 2.2347955828792907, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.104, "cuda_time_us": 270.876, "pct_cuda_time": 1.3632253558829137, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 270.14, "pct_cuda_time": 1.3595213220743452, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.609, "cuda_time_us": 33.664, "pct_cuda_time": 0.1694192855049632, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.664, "pct_cuda_time": 0.1694192855049632, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.701, "cuda_time_us": 139.519, "pct_cuda_time": 0.702150941491414, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 139.519, "pct_cuda_time": 0.702150941491414, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2397.654, "cuda_time_us": 607.513, "pct_cuda_time": 3.0574031129686525, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.442, "cuda_time_us": 8.352, "pct_cuda_time": 0.04203273147984353, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.352, "pct_cuda_time": 0.04203273147984353, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1719.478, "cuda_time_us": 146.751, "pct_cuda_time": 0.7385470997843053, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.307, "cuda_time_us": 69.60000000000001, "pct_cuda_time": 0.3502727623320295, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.864, "pct_cuda_time": 0.3465687285234609, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 511.489, "cuda_time_us": 10.559, "pct_cuda_time": 0.05313980025091808, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.559, "pct_cuda_time": 0.05313980025091808, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.115, "cuda_time_us": 21.472, "pct_cuda_time": 0.10806116024128358, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.152, "pct_cuda_time": 0.02592823665998011, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.07569112565335809, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.146, "cuda_time_us": 45.12, "pct_cuda_time": 0.22707337696007426, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.12, "pct_cuda_time": 0.22707337696007426, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.634, "cuda_time_us": 8.16, "pct_cuda_time": 0.04106646179065173, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.16, "pct_cuda_time": 0.04106646179065173, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 454.236, "cuda_time_us": 444.25, "pct_cuda_time": 2.235756819913852, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.123, "cuda_time_us": 270.3, "pct_cuda_time": 1.3603265468153385, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 269.564, "pct_cuda_time": 1.35662251300677, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.639, "cuda_time_us": 34.24, "pct_cuda_time": 0.17231809457253863, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.24, "pct_cuda_time": 0.17231809457253863, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.955, "cuda_time_us": 139.71, "pct_cuda_time": 0.7031121785259746, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 139.71, "pct_cuda_time": 0.7031121785259746, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2573.032, "cuda_time_us": 610.038, "pct_cuda_time": 3.070110565912451, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.776, "cuda_time_us": 8.512, "pct_cuda_time": 0.042837956220836707, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.512, "pct_cuda_time": 0.042837956220836707, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1865.002, "cuda_time_us": 147.229, "pct_cuda_time": 0.7409527086980225, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.6, "cuda_time_us": 69.15100000000001, "pct_cuda_time": 0.3480131004026174, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.415, "pct_cuda_time": 0.3443090665940488, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 527.007, "cuda_time_us": 10.56, "pct_cuda_time": 0.05314483290554929, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.56, "pct_cuda_time": 0.05314483290554929, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 829.244, "cuda_time_us": 22.239, "pct_cuda_time": 0.1119212063434196, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.536, "pct_cuda_time": 0.02786077603836372, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.231, "pct_cuda_time": 0.07665236268791868, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.551, "cuda_time_us": 45.279, "pct_cuda_time": 0.22787356904643624, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.279, "pct_cuda_time": 0.22787356904643624, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.023, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.256, "pct_cuda_time": 0.04154959663524763, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 482.797, "cuda_time_us": 446.041, "pct_cuda_time": 2.244770304358344, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.675, "cuda_time_us": 272.348, "pct_cuda_time": 1.370633423500051, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 271.612, "pct_cuda_time": 1.3669293896914827, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.837, "cuda_time_us": 33.599, "pct_cuda_time": 0.1690921629539347, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.599, "pct_cuda_time": 0.1690921629539347, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.928, "cuda_time_us": 140.094, "pct_cuda_time": 0.7050447179043582, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 140.094, "pct_cuda_time": 0.7050447179043582, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2389.084, "cuda_time_us": 606.5859999999999, "pct_cuda_time": 3.0527378421255227, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.327, "cuda_time_us": 8.033, "pct_cuda_time": 0.040427314652488396, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.033, "pct_cuda_time": 0.040427314652488396, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1709.914, "cuda_time_us": 146.335, "pct_cuda_time": 0.7364535154577232, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.023, "cuda_time_us": 68.384, "pct_cuda_time": 0.3441530543004813, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.003709066463199794, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 67.647, "pct_cuda_time": 0.3404439878372816, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.675, "cuda_time_us": 10.944, "pct_cuda_time": 0.055077372283932904, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.944, "pct_cuda_time": 0.055077372283932904, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.653, "cuda_time_us": 21.92, "pct_cuda_time": 0.11031578951606447, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.248, "pct_cuda_time": 0.026411371504576012, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.2, "pct_cuda_time": 0.07649635039435125, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.007408067617137174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.862, "cuda_time_us": 45.087, "pct_cuda_time": 0.22690729935724444, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.087, "pct_cuda_time": 0.22690729935724444, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.873, "cuda_time_us": 8.48, "pct_cuda_time": 0.04267691127263807, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.48, "pct_cuda_time": 0.04267691127263807, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 455.271, "cuda_time_us": 443.73799999999994, "pct_cuda_time": 2.2331801007426733, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.814, "cuda_time_us": 270.876, "pct_cuda_time": 1.3632253558829137, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 270.14, "pct_cuda_time": 1.3595213220743452, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.9, "cuda_time_us": 33.856, "pct_cuda_time": 0.17038555519415502, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.856, "pct_cuda_time": 0.17038555519415502, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.294, "cuda_time_us": 139.006, "pct_cuda_time": 0.6995691896656047, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 139.006, "pct_cuda_time": 0.6995691896656047, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2480.019, "cuda_time_us": 608.216, "pct_cuda_time": 3.0609410691743912, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.583, "cuda_time_us": 8.64, "pct_cuda_time": 0.04348213601363124, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.04348213601363124, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1760.307, "cuda_time_us": 146.847, "pct_cuda_time": 0.7390302346289013, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.264, "cuda_time_us": 68.89500000000001, "pct_cuda_time": 0.3467247408170283, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 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": 68.159, "pct_cuda_time": 0.3430207070084597, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[768, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 518.342, "cuda_time_us": 10.528, "pct_cuda_time": 0.05298378795735066, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 10.528, "pct_cuda_time": 0.05298378795735066, "trace": "_C::rotary_embedding(int64[768], bfloat16[768, 4096], bfloat16[768, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 735.498, "cuda_time_us": 21.952, "pct_cuda_time": 0.1104768344642631, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 5.248, "pct_cuda_time": 0.026411371504576012, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[768], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.07762366503174169, "trace": "_vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.006441797927945369, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], None, None, bfloat16[768, 32, 128], int32[7], int32[7], None, None, None, 128, 128, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[768, 32, 128], bfloat16[768, 8, 128], bfloat16[768, 8, 128], bfloat16[768, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.526, "cuda_time_us": 45.472, "pct_cuda_time": 0.22884487139025922, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 45.472, "pct_cuda_time": 0.22884487139025922, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[768, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 106.508, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.223, "pct_cuda_time": 0.04138351903241779, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 473.075, "cuda_time_us": 444.506, "pct_cuda_time": 2.237045179499441, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.863, "cuda_time_us": 271.772, "pct_cuda_time": 1.3677346144324756, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 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": 271.036, "pct_cuda_time": 1.364030580623907, "trace": "mm(bfloat16[768, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[768, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[768, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.021, "cuda_time_us": 33.664, "pct_cuda_time": 0.1694192855049632, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 33.664, "pct_cuda_time": 0.1694192855049632, "trace": "_C::silu_and_mul(bfloat16[768, 14336], bfloat16[768, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.175, "cuda_time_us": 139.07, "pct_cuda_time": 0.6998912795620019, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 139.07, "pct_cuda_time": 0.6998912795620019, "trace": "mm(bfloat16[768, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[768, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[768, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.089, "cuda_time_us": 8.64, "pct_cuda_time": 0.04348213601363124, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.04348213601363124, "trace": "_C::fused_add_rms_norm(bfloat16[768, 4096], bfloat16[768, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 459.467, "cuda_time_us": 351.419, "pct_cuda_time": 1.768570457844245, "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": 5.023, "pct_cuda_time": 0.025279024212554364, "trace": "index_select(bfloat16[768, 4096], 0, int64[6])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.003709066463199794, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 345.659, "pct_cuda_time": 1.7395823671684907, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3499.821, "cuda_time_us": 118.525, "pct_cuda_time": 0.5964953901638476, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.003704033808568587, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 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": 0.003704033808568587, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 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": 0.0038650787567672215, "trace": "copy_(int32[6], int32[6], True) <- _to_copy(int32[6], 3, 0, None, None, True, None) <- to(int32[6], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0038650787567672215, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 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": 0.0038650787567672215, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 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": 0.004026123704965856, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 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": 0.004026123704965856, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.544, "pct_cuda_time": 0.022868382644206056, "trace": "copy_(float32[6, 128256], bfloat16[6, 128256], False) <- _to_copy(bfloat16[6, 128256], 6, None, None, None, False, None) <- to(bfloat16[6, 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": 5.536, "pct_cuda_time": 0.02786077603836372, "trace": "div_(float32[6, 128256], bfloat16[6, 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": 35.039, "pct_cuda_time": 0.17633918562287326, "trace": "_softmax(float32[6, 128256], -1, False) <- softmax(float32[6, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 28.607, "pct_cuda_time": 0.14396915103494778, "trace": "_log_softmax(float32[6, 128256], -1, False) <- log_softmax(float32[6, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.952, "pct_cuda_time": 0.009823741840116688, "trace": "copy_(int64[6], int32[6], False) <- _to_copy(int32[6], 4, None, None, None, False, None) <- to(int32[6], 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": 6.208, "pct_cuda_time": 0.031242719950535038, "trace": "index(float32[6, 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": 28.191, "pct_cuda_time": 0.14187556670836554, "trace": "argmax(float32[6, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 3.072, "pct_cuda_time": 0.015460315027068886, "trace": "copy_(int64[6], int64[6], False) <- _to_copy(int64[6], 4, 0, None, None, False, None) <- to(int64[6], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 6 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6389.298, "pct_cuda_time": 93.07273406358541, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 6.048, "pct_cuda_time": 0.08810105517328581, "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": 6.048, "pct_cuda_time": 0.08810105517328581, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6380.209999999999, "pct_cuda_time": 92.94034940925094, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 198.788, "pct_cuda_time": 2.89573950988544, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.256, "pct_cuda_time": 0.06199703882564557, "invocations": 1 }, "children": [] }, { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 194.53199999999998, "pct_cuda_time": 2.833742471059794, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 1834.019, "pct_cuda_time": 26.716106003282814, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 665.1420000000002, "pct_cuda_time": 9.689105826731097, "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": 665.1420000000002, "pct_cuda_time": 9.689105826731097, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 118.36500000000004, "pct_cuda_time": 1.7242198074712265, "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": 118.36500000000004, "pct_cuda_time": 1.7242198074712265, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 481.52700000000004, "pct_cuda_time": 7.014391004369509, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cuda_time_us": 80.893, "pct_cuda_time": 1.178366179916106, "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": 358.7809999999999, "pct_cuda_time": 5.226353286396601, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 41.853, "pct_cuda_time": 0.6096715380568007, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 568.985, "pct_cuda_time": 8.288389364710982, "invocations": 32 }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cuda_time_us": 501.178, "pct_cuda_time": 7.300646598815646, "invocations": 32 }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cuda_time_us": 67.80700000000003, "pct_cuda_time": 0.9877427658953362, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4347.403, "pct_cuda_time": 63.32850389608271, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2664.4130000000005, "pct_cuda_time": 38.81243331967923, "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": 2664.4130000000005, "pct_cuda_time": 38.81243331967923, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 289.72599999999994, "pct_cuda_time": 4.220430937687731, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 289.72599999999994, "pct_cuda_time": 4.220430937687731, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1393.264, "pct_cuda_time": 20.295639638715752, "invocations": 32 }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 1393.264, "pct_cuda_time": 20.295639638715752, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "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": 3.04, "pct_cuda_time": 0.0442835991611754, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 350.619, "pct_cuda_time": 5.107457649438211, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 4.768, "pct_cuda_time": 0.06945532921068563, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.768, "pct_cuda_time": 0.011187435577560101, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 345.083, "pct_cuda_time": 5.026814884649965, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 124.92699999999999, "pct_cuda_time": 1.819808286976368, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 13.184, "pct_cuda_time": 0.19205097741478175, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 4.416, "pct_cuda_time": 0.06432775457097059, "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": 5.504, "pct_cuda_time": 0.08017662163918073, "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": 34.943, "pct_cuda_time": 0.5090137518055763, "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": 28.384, "pct_cuda_time": 0.41346897322065873, "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": 0.026570159496705242, "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": 6.208, "pct_cuda_time": 0.09043177091861082, "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": 27.776, "pct_cuda_time": 0.40461225338842366, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.688, "pct_cuda_time": 0.03915602452146036, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 84543.208, "cuda_time_us": 6389.298, "pct_cuda_time": 93.07273406358541, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 412.114, "cuda_time_us": 6.048, "pct_cuda_time": 0.08810105517328581, "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": 6.048, "pct_cuda_time": 0.08810105517328581, "trace": "index_select(bfloat16[128256, 4096], 0, int64[6]) <- embedding(bfloat16[128256, 4096], int64[6], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 5617.006, "cuda_time_us": 206.52300000000002, "pct_cuda_time": 3.0084150491984962, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 405.089, "cuda_time_us": 4.256, "pct_cuda_time": 0.06199703882564557, "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": 4.256, "pct_cuda_time": 0.06199703882564557, "trace": "_C::rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 4166.876, "cuda_time_us": 62.269, "pct_cuda_time": 0.9070708671602734, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 850.843, "cuda_time_us": 25.055, "pct_cuda_time": 0.36497551874449, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 25.055, "pct_cuda_time": 0.36497551874449, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1204.019, "cuda_time_us": 3.583, "pct_cuda_time": 0.052193465721872195, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.052193465721872195, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1363.045, "cuda_time_us": 14.847, "pct_cuda_time": 0.2162758541927537, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03449459303081031, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.167, "pct_cuda_time": 0.1626693920502782, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 360.603, "cuda_time_us": 18.784, "pct_cuda_time": 0.27362602850115747, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 16.704, "pct_cuda_time": 0.2433267238119322, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.030299304689225277, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 151.749, "cuda_time_us": 3.103, "pct_cuda_time": 0.04520131848589713, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04520131848589713, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 741.937, "cuda_time_us": 136.895, "pct_cuda_time": 1.99414582472668, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 267.14, "cuda_time_us": 83.551, "pct_cuda_time": 1.2170851952353179, "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": 83.551, "pct_cuda_time": 1.2170851952353179, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 207.194, "cuda_time_us": 9.024, "pct_cuda_time": 0.13145236803633117, "trace": "" }, "children": [ { "entry": { "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": 0.13145236803633117, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 174.569, "cuda_time_us": 44.32, "pct_cuda_time": 0.6456082614550308, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.32, "pct_cuda_time": 0.6456082614550308, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2799.497, "cuda_time_us": 198.75, "pct_cuda_time": 2.8951859648959246, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.621, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2037.352, "cuda_time_us": 57.6, "pct_cuda_time": 0.8390576683170077, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.159, "cuda_time_us": 21.408, "pct_cuda_time": 0.3118497667244879, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3118497667244879, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 582.658, "cuda_time_us": 3.68, "pct_cuda_time": 0.05360646214247549, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05360646214247549, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 935.956, "cuda_time_us": 15.008, "pct_cuda_time": 0.21862113691148696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.683, "cuda_time_us": 17.503999999999998, "pct_cuda_time": 0.2549803025385573, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.22421485470026703, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.159, "cuda_time_us": 3.264, "pct_cuda_time": 0.04754660120463043, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04754660120463043, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 501.274, "cuda_time_us": 134.814, "pct_cuda_time": 1.9638319530640462, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 178.437, "cuda_time_us": 82.687, "pct_cuda_time": 1.2044993302105627, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.687, "pct_cuda_time": 1.2044993302105627, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 109.761, "cuda_time_us": 8.959, "pct_cuda_time": 0.1305055147647929, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.959, "pct_cuda_time": 0.1305055147647929, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.617, "cuda_time_us": 43.168, "pct_cuda_time": 0.6288271080886907, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.168, "pct_cuda_time": 0.6288271080886907, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2466.437, "cuda_time_us": 199.038, "pct_cuda_time": 2.89938125323751, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.822, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1733.674, "cuda_time_us": 56.607, "pct_cuda_time": 0.8245926637225842, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.946, "cuda_time_us": 20.384, "pct_cuda_time": 0.2969331859544077, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.384, "pct_cuda_time": 0.2969331859544077, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.338, "cuda_time_us": 3.679, "pct_cuda_time": 0.053591895169067205, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.679, "pct_cuda_time": 0.053591895169067205, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 742.643, "cuda_time_us": 14.975999999999999, "pct_cuda_time": 0.21815499376242198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.168, "pct_cuda_time": 0.1626839590236865, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.295, "cuda_time_us": 17.567999999999998, "pct_cuda_time": 0.25591258883668727, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.22514714099839706, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.037, "cuda_time_us": 3.105, "pct_cuda_time": 0.045230452432713696, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045230452432713696, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 494.051, "cuda_time_us": 136.286, "pct_cuda_time": 1.9852745379210366, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.436, "cuda_time_us": 83.103, "pct_cuda_time": 1.2105591911484077, "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": 1.2105591911484077, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.566, "cuda_time_us": 9.056, "pct_cuda_time": 0.1319185111853962, "trace": "" }, "children": [ { "entry": { "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": 0.1319185111853962, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 167.744, "cuda_time_us": 44.127, "pct_cuda_time": 0.6427968355872327, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.127, "pct_cuda_time": 0.6427968355872327, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2495.568, "cuda_time_us": 198.749, "pct_cuda_time": 2.895171397922516, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.829, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1773.174, "cuda_time_us": 56.672, "pct_cuda_time": 0.8255395169941224, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.47, "cuda_time_us": 20.416, "pct_cuda_time": 0.2973993291034727, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.416, "pct_cuda_time": 0.2973993291034727, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 519.722, "cuda_time_us": 3.712, "pct_cuda_time": 0.054072605291540496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054072605291540496, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 774.855, "cuda_time_us": 14.975999999999999, "pct_cuda_time": 0.21815499376242198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03635916562707033, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 164.792, "cuda_time_us": 17.567999999999998, "pct_cuda_time": 0.25591258883668727, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.22514714099839706, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.585, "cuda_time_us": 3.104, "pct_cuda_time": 0.04521588545930541, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04521588545930541, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.347, "cuda_time_us": 135.933, "pct_cuda_time": 1.980132396307913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.735, "cuda_time_us": 83.774, "pct_cuda_time": 1.2203336303053645, "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": 83.774, "pct_cuda_time": 1.2203336303053645, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.825, "cuda_time_us": 8.864, "pct_cuda_time": 0.12912165229100617, "trace": "" }, "children": [ { "entry": { "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": 0.12912165229100617, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.907, "cuda_time_us": 43.295, "pct_cuda_time": 0.6306771137115424, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.295, "pct_cuda_time": 0.6306771137115424, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2407.614, "cuda_time_us": 199.39100000000002, "pct_cuda_time": 2.9045233948506333, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.132, "cuda_time_us": 3.008, "pct_cuda_time": 0.0438174560121104, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0438174560121104, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1718.45, "cuda_time_us": 57.44, "pct_cuda_time": 0.8367269525716826, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.936, "cuda_time_us": 20.608, "pct_cuda_time": 0.30019618799786274, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30019618799786274, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.606, "cuda_time_us": 3.648, "pct_cuda_time": 0.053140318993410485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053140318993410485, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 774.824, "cuda_time_us": 14.911999999999999, "pct_cuda_time": 0.21722270746429195, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.104, "pct_cuda_time": 0.16175167272555646, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.244, "cuda_time_us": 18.272, "pct_cuda_time": 0.26616773811611744, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 16.16, "pct_cuda_time": 0.23540229027782714, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.63, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.791, "cuda_time_us": 135.80700000000002, "pct_cuda_time": 1.9782969576584701, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.072, "cuda_time_us": 82.911, "pct_cuda_time": 1.2077623322540179, "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": 82.911, "pct_cuda_time": 1.2077623322540179, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.511, "cuda_time_us": 8.96, "pct_cuda_time": 0.1305200817382012, "trace": "" }, "children": [ { "entry": { "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": 0.1305200817382012, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.227, "cuda_time_us": 43.936, "pct_cuda_time": 0.6400145436662509, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.936, "pct_cuda_time": 0.6400145436662509, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2566.144, "cuda_time_us": 198.844, "pct_cuda_time": 2.896555260396303, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.867, "cuda_time_us": 2.976, "pct_cuda_time": 0.043351312863045395, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043351312863045395, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1844.153, "cuda_time_us": 57.053, "pct_cuda_time": 0.8310895338626776, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.178, "cuda_time_us": 20.416, "pct_cuda_time": 0.2973993291034727, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.416, "pct_cuda_time": 0.2973993291034727, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 483.374, "cuda_time_us": 3.807, "pct_cuda_time": 0.05545646776532723, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.807, "pct_cuda_time": 0.05545646776532723, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 900.974, "cuda_time_us": 15.166999999999998, "pct_cuda_time": 0.2209372856834037, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.687, "pct_cuda_time": 0.039141457548052074, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 175.059, "cuda_time_us": 17.663, "pct_cuda_time": 0.2572964513104741, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.52, "pct_cuda_time": 0.22607942729652705, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.143, "pct_cuda_time": 0.031217024013947, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.501, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.572, "cuda_time_us": 135.679, "pct_cuda_time": 1.9764323850622096, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.959, "cuda_time_us": 82.815, "pct_cuda_time": 1.2063639028068225, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.815, "pct_cuda_time": 1.2063639028068225, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.199, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.773, "cuda_time_us": 43.744, "pct_cuda_time": 0.6372176847718608, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.744, "pct_cuda_time": 0.6372176847718608, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2466.917, "cuda_time_us": 199.486, "pct_cuda_time": 2.9059072573244196, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.967, "cuda_time_us": 3.104, "pct_cuda_time": 0.04521588545930541, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04521588545930541, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1777.985, "cuda_time_us": 57.791, "pct_cuda_time": 0.8418399602379892, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.588, "cuda_time_us": 20.543, "pct_cuda_time": 0.2992493347263244, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.543, "pct_cuda_time": 0.2992493347263244, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.506, "cuda_time_us": 3.648, "pct_cuda_time": 0.053140318993410485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053140318993410485, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 762.815, "cuda_time_us": 15.136, "pct_cuda_time": 0.22048570950774699, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.328, "pct_cuda_time": 0.1650146747690115, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 163.766, "cuda_time_us": 18.464, "pct_cuda_time": 0.26896459701050746, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 16.32, "pct_cuda_time": 0.23773300602315217, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.144, "pct_cuda_time": 0.031231590987355288, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.265, "cuda_time_us": 3.168, "pct_cuda_time": 0.04614817175743542, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04614817175743542, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.971, "cuda_time_us": 135.423, "pct_cuda_time": 1.9727032398696898, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.728, "cuda_time_us": 82.4, "pct_cuda_time": 1.2003186088423858, "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": 82.4, "pct_cuda_time": 1.2003186088423858, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.468, "cuda_time_us": 9.024, "pct_cuda_time": 0.13145236803633117, "trace": "" }, "children": [ { "entry": { "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": 0.13145236803633117, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.296, "cuda_time_us": 43.999, "pct_cuda_time": 0.6409322629909726, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6409322629909726, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2379.871, "cuda_time_us": 198.07899999999998, "pct_cuda_time": 2.8854115257389674, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.302, "cuda_time_us": 3.104, "pct_cuda_time": 0.04521588545930541, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04521588545930541, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1666.146, "cuda_time_us": 56.833, "pct_cuda_time": 0.8278847997128558, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.244, "cuda_time_us": 20.543, "pct_cuda_time": 0.2992493347263244, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.543, "pct_cuda_time": 0.2992493347263244, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.732, "cuda_time_us": 3.776, "pct_cuda_time": 0.0550048915896705, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0550048915896705, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 738.98, "cuda_time_us": 14.976999999999999, "pct_cuda_time": 0.21816956073583024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.497, "pct_cuda_time": 0.03637373260047861, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 155.588, "cuda_time_us": 17.537, "pct_cuda_time": 0.25546101266103055, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.22468099784933204, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030780014811698564, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 100.855, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 466.415, "cuda_time_us": 135.10199999999998, "pct_cuda_time": 1.9680272414056308, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.168, "cuda_time_us": 82.815, "pct_cuda_time": 1.2063639028068225, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.815, "pct_cuda_time": 1.2063639028068225, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.401, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 150.025, "cuda_time_us": 43.199, "pct_cuda_time": 0.6292786842643474, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.199, "pct_cuda_time": 0.6292786842643474, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2281.26, "cuda_time_us": 200.414, "pct_cuda_time": 2.919425408647305, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.154, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1609.567, "cuda_time_us": 57.344, "pct_cuda_time": 0.8353285231244876, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.569, "cuda_time_us": 20.896, "pct_cuda_time": 0.3043914763394478, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3043914763394478, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 462.379, "cuda_time_us": 3.744, "pct_cuda_time": 0.0545387484406055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0545387484406055, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 716.644, "cuda_time_us": 14.975999999999999, "pct_cuda_time": 0.21815499376242198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03635916562707033, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 150.088, "cuda_time_us": 17.728, "pct_cuda_time": 0.25824330458201233, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.22794399989278707, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.030299304689225277, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.091, "cuda_time_us": 3.232, "pct_cuda_time": 0.047080458055565426, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047080458055565426, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.527, "cuda_time_us": 136.798, "pct_cuda_time": 1.9927328283060768, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 150.788, "cuda_time_us": 83.519, "pct_cuda_time": 1.216619052086253, "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": 1.216619052086253, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.049, "cuda_time_us": 9.216, "pct_cuda_time": 0.13424922693072122, "trace": "" }, "children": [ { "entry": { "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": 0.13424922693072122, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.724, "cuda_time_us": 44.063, "pct_cuda_time": 0.6418645492891026, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.063, "pct_cuda_time": 0.6418645492891026, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2499.41, "cuda_time_us": 199.489, "pct_cuda_time": 2.9059509582446448, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.224, "cuda_time_us": 3.105, "pct_cuda_time": 0.045230452432713696, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045230452432713696, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1815.47, "cuda_time_us": 56.896, "pct_cuda_time": 0.8288025190375775, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 212.443, "cuda_time_us": 20.352, "pct_cuda_time": 0.2964670428053427, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.352, "pct_cuda_time": 0.2964670428053427, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.342, "cuda_time_us": 3.744, "pct_cuda_time": 0.0545387484406055, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0545387484406055, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 771.524, "cuda_time_us": 15.008, "pct_cuda_time": 0.21862113691148696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.329, "cuda_time_us": 17.792, "pct_cuda_time": 0.25917559088014236, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.68, "pct_cuda_time": 0.2284101430418521, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.082, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 454.093, "cuda_time_us": 136.448, "pct_cuda_time": 1.9876343876131783, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.472, "cuda_time_us": 83.615, "pct_cuda_time": 1.2180174815334477, "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": 1.2180174815334477, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.364, "cuda_time_us": 9.665, "pct_cuda_time": 0.14078979799103955, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.665, "pct_cuda_time": 0.14078979799103955, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.795, "cuda_time_us": 43.168, "pct_cuda_time": 0.6288271080886907, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.168, "pct_cuda_time": 0.6288271080886907, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2330.797, "cuda_time_us": 198.687, "pct_cuda_time": 2.894268245571203, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.503, "cuda_time_us": 3.041, "pct_cuda_time": 0.044298166134583684, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044298166134583684, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1662.34, "cuda_time_us": 57.056, "pct_cuda_time": 0.8311332347829026, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.5, "cuda_time_us": 20.608, "pct_cuda_time": 0.30019618799786274, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30019618799786274, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.112, "cuda_time_us": 3.647, "pct_cuda_time": 0.0531257520200022, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.647, "pct_cuda_time": 0.0531257520200022, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.213, "cuda_time_us": 14.977, "pct_cuda_time": 0.21816956073583024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.169, "pct_cuda_time": 0.16269852599709475, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 151.172, "cuda_time_us": 17.823999999999998, "pct_cuda_time": 0.2596417340292073, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.552, "pct_cuda_time": 0.22654557044559206, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.0330961635836153, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.137, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.582, "cuda_time_us": 135.518, "pct_cuda_time": 1.9740871023434765, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.068, "cuda_time_us": 83.199, "pct_cuda_time": 1.2119576205956026, "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": 83.199, "pct_cuda_time": 1.2119576205956026, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.453, "cuda_time_us": 8.864, "pct_cuda_time": 0.12912165229100617, "trace": "" }, "children": [ { "entry": { "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": 0.12912165229100617, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.031, "cuda_time_us": 43.455, "pct_cuda_time": 0.6330078294568674, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6330078294568674, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2470.545, "cuda_time_us": 199.26000000000002, "pct_cuda_time": 2.902615121334149, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.202, "cuda_time_us": 3.104, "pct_cuda_time": 0.04521588545930541, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04521588545930541, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1767.752, "cuda_time_us": 57.342, "pct_cuda_time": 0.8352993891776711, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.88, "cuda_time_us": 20.512, "pct_cuda_time": 0.2987977585506677, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.512, "pct_cuda_time": 0.2987977585506677, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 464.977, "cuda_time_us": 3.712, "pct_cuda_time": 0.054072605291540496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054072605291540496, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.145, "cuda_time_us": 15.103, "pct_cuda_time": 0.2200049993852737, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.263, "pct_cuda_time": 0.1640678214974732, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.312, "cuda_time_us": 18.015, "pct_cuda_time": 0.2624240259501891, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.903, "pct_cuda_time": 0.23165857811189883, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.493, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.982, "cuda_time_us": 135.74200000000002, "pct_cuda_time": 1.9773501043869317, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.33, "cuda_time_us": 82.943, "pct_cuda_time": 1.2082284754030825, "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": 1.2082284754030825, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.868, "cuda_time_us": 8.992, "pct_cuda_time": 0.1309862248872662, "trace": "" }, "children": [ { "entry": { "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": 0.1309862248872662, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.368, "cuda_time_us": 43.807, "pct_cuda_time": 0.6381354040965825, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.807, "pct_cuda_time": 0.6381354040965825, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2665.113, "cuda_time_us": 198.781, "pct_cuda_time": 2.8956375410715816, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.352, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1933.883, "cuda_time_us": 57.053999999999995, "pct_cuda_time": 0.8311041008360859, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.378, "cuda_time_us": 20.512, "pct_cuda_time": 0.2987977585506677, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.512, "pct_cuda_time": 0.2987977585506677, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 515.005, "cuda_time_us": 3.648, "pct_cuda_time": 0.053140318993410485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053140318993410485, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 939.022, "cuda_time_us": 14.975999999999999, "pct_cuda_time": 0.21815499376242198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.136, "pct_cuda_time": 0.16221781587462147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.93, "cuda_time_us": 17.918, "pct_cuda_time": 0.2610110295295858, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.807, "pct_cuda_time": 0.23026014866470382, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030750880864882, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.769, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 481.962, "cuda_time_us": 135.64700000000002, "pct_cuda_time": 1.975966241913145, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.933, "cuda_time_us": 83.775, "pct_cuda_time": 1.2203481972787729, "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": 1.2203481972787729, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.114, "cuda_time_us": 8.96, "pct_cuda_time": 0.1305200817382012, "trace": "" }, "children": [ { "entry": { "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": 0.1305200817382012, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.68, "cuda_time_us": 42.912, "pct_cuda_time": 0.6250979628961707, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.912, "pct_cuda_time": 0.6250979628961707, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2350.863, "cuda_time_us": 198.94, "pct_cuda_time": 2.897953689843498, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.938, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1692.745, "cuda_time_us": 56.989, "pct_cuda_time": 0.8301572475645477, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.954, "cuda_time_us": 20.639, "pct_cuda_time": 0.30064776417351946, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.639, "pct_cuda_time": 0.30064776417351946, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.907, "cuda_time_us": 3.648, "pct_cuda_time": 0.053140318993410485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053140318993410485, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 759.559, "cuda_time_us": 15.071, "pct_cuda_time": 0.2195388562362087, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03635916562707033, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.263, "pct_cuda_time": 0.1640678214974732, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 153.368, "cuda_time_us": 17.631, "pct_cuda_time": 0.25683030816140906, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.52, "pct_cuda_time": 0.22607942729652705, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030750880864882, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.44, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 445.613, "cuda_time_us": 135.839, "pct_cuda_time": 1.9787631008075346, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.883, "cuda_time_us": 83.871, "pct_cuda_time": 1.2217466267259678, "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": 1.2217466267259678, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.78, "cuda_time_us": 8.992, "pct_cuda_time": 0.1309862248872662, "trace": "" }, "children": [ { "entry": { "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": 0.1309862248872662, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.268, "cuda_time_us": 42.976, "pct_cuda_time": 0.6260302491943007, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.976, "pct_cuda_time": 0.6260302491943007, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2536.969, "cuda_time_us": 199.036, "pct_cuda_time": 2.8993521192906933, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.255, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1825.257, "cuda_time_us": 56.991, "pct_cuda_time": 0.8301863815113643, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 172.241, "cuda_time_us": 20.607, "pct_cuda_time": 0.30018162102445445, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.607, "pct_cuda_time": 0.30018162102445445, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.326, "cuda_time_us": 3.648, "pct_cuda_time": 0.053140318993410485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053140318993410485, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 796.856, "cuda_time_us": 15.04, "pct_cuda_time": 0.21908728006055198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.232, "pct_cuda_time": 0.16361624532181648, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.916, "cuda_time_us": 17.695999999999998, "pct_cuda_time": 0.2577771614329473, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.584, "pct_cuda_time": 0.22701171359465708, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.733, "cuda_time_us": 3.2, "pct_cuda_time": 0.04661431490650043, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04661431490650043, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.89, "cuda_time_us": 135.805, "pct_cuda_time": 1.9782678237116533, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.433, "cuda_time_us": 83.07, "pct_cuda_time": 1.2100784810259344, "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": 83.07, "pct_cuda_time": 1.2100784810259344, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.827, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.725, "cuda_time_us": 43.615, "pct_cuda_time": 0.6353385452021926, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.615, "pct_cuda_time": 0.6353385452021926, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2367.036, "cuda_time_us": 199.23, "pct_cuda_time": 2.9021781121318995, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.202, "cuda_time_us": 3.008, "pct_cuda_time": 0.0438174560121104, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0438174560121104, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1655.114, "cuda_time_us": 57.248, "pct_cuda_time": 0.8339300936772925, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.836, "cuda_time_us": 20.832, "pct_cuda_time": 0.3034591900413178, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.832, "pct_cuda_time": 0.3034591900413178, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.476, "cuda_time_us": 3.68, "pct_cuda_time": 0.05360646214247549, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05360646214247549, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 727.114, "cuda_time_us": 15.008, "pct_cuda_time": 0.21862113691148696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 153.782, "cuda_time_us": 17.728, "pct_cuda_time": 0.25824330458201233, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.2274778567437221, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.554, "cuda_time_us": 3.168, "pct_cuda_time": 0.04614817175743542, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04614817175743542, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.23, "cuda_time_us": 135.80599999999998, "pct_cuda_time": 1.9782823906850613, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.595, "cuda_time_us": 83.039, "pct_cuda_time": 1.209626904850278, "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": 83.039, "pct_cuda_time": 1.209626904850278, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.907, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.811, "cuda_time_us": 43.679, "pct_cuda_time": 0.6362708315003226, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.679, "pct_cuda_time": 0.6362708315003226, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2424.643, "cuda_time_us": 199.29299999999998, "pct_cuda_time": 2.9030958314566213, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.082, "cuda_time_us": 3.168, "pct_cuda_time": 0.04614817175743542, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04614817175743542, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1699.211, "cuda_time_us": 56.992, "pct_cuda_time": 0.8302009484847725, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.467, "cuda_time_us": 20.448, "pct_cuda_time": 0.2978654722525377, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.448, "pct_cuda_time": 0.2978654722525377, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 495.342, "cuda_time_us": 3.681, "pct_cuda_time": 0.05362102911588377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.681, "pct_cuda_time": 0.05362102911588377, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 757.919, "cuda_time_us": 15.103, "pct_cuda_time": 0.2200049993852737, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.264, "pct_cuda_time": 0.1640823884708815, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.019097302138256892, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 165.15, "cuda_time_us": 17.759999999999998, "pct_cuda_time": 0.25870944773107735, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.22794399989278707, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.644, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 486.867, "cuda_time_us": 136.06099999999998, "pct_cuda_time": 1.981996968904173, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.411, "cuda_time_us": 83.71, "pct_cuda_time": 1.2194013440072344, "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": 83.71, "pct_cuda_time": 1.2194013440072344, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.337, "cuda_time_us": 9.056, "pct_cuda_time": 0.1319185111853962, "trace": "" }, "children": [ { "entry": { "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": 0.1319185111853962, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.996, "cuda_time_us": 43.295, "pct_cuda_time": 0.6306771137115424, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.295, "pct_cuda_time": 0.6306771137115424, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2385.86, "cuda_time_us": 198.36300000000003, "pct_cuda_time": 2.8895485461869206, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.426, "cuda_time_us": 3.135, "pct_cuda_time": 0.04566746163496213, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04566746163496213, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1694.155, "cuda_time_us": 56.862, "pct_cuda_time": 0.828307241941696, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.172, "cuda_time_us": 20.608, "pct_cuda_time": 0.30019618799786274, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30019618799786274, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.554, "cuda_time_us": 3.679, "pct_cuda_time": 0.053591895169067205, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.679, "pct_cuda_time": 0.053591895169067205, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 749.591, "cuda_time_us": 15.039, "pct_cuda_time": 0.21907271308714368, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.232, "pct_cuda_time": 0.16361624532181648, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.279, "pct_cuda_time": 0.018631158989191886, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 157.244, "cuda_time_us": 17.536, "pct_cuda_time": 0.2554464456876223, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.22468099784933204, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.212, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.987, "cuda_time_us": 135.32600000000002, "pct_cuda_time": 1.9712902434490869, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.728, "cuda_time_us": 82.527, "pct_cuda_time": 1.2021686144652377, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.527, "pct_cuda_time": 1.2021686144652377, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.46, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.312, "cuda_time_us": 43.679, "pct_cuda_time": 0.6362708315003226, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.679, "pct_cuda_time": 0.6362708315003226, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2495.336, "cuda_time_us": 198.973, "pct_cuda_time": 2.898434399965972, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.685, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1772.15, "cuda_time_us": 56.926, "pct_cuda_time": 0.829239528239826, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.535, "cuda_time_us": 20.512, "pct_cuda_time": 0.2987977585506677, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.512, "pct_cuda_time": 0.2987977585506677, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.265, "cuda_time_us": 3.84, "pct_cuda_time": 0.05593717788780051, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05593717788780051, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 773.572, "cuda_time_us": 14.975, "pct_cuda_time": 0.21814042678901369, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.036344598653662054, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.168, "pct_cuda_time": 0.1626839590236865, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.502, "cuda_time_us": 17.599, "pct_cuda_time": 0.25636416501234405, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.519, "pct_cuda_time": 0.2260648603231188, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.030299304689225277, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.161, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.016, "cuda_time_us": 135.839, "pct_cuda_time": 1.9787631008075346, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.787, "cuda_time_us": 83.327, "pct_cuda_time": 1.2138221931918627, "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": 1.2138221931918627, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.629, "cuda_time_us": 8.992, "pct_cuda_time": 0.1309862248872662, "trace": "" }, "children": [ { "entry": { "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": 0.1309862248872662, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.06, "cuda_time_us": 43.52, "pct_cuda_time": 0.6339546827284058, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.52, "pct_cuda_time": 0.6339546827284058, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2721.419, "cuda_time_us": 199.35999999999999, "pct_cuda_time": 2.9040718186749763, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.062, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1811.053, "cuda_time_us": 57.248, "pct_cuda_time": 0.8339300936772925, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 167.021, "cuda_time_us": 20.576, "pct_cuda_time": 0.2997300448487977, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.576, "pct_cuda_time": 0.2997300448487977, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.855, "cuda_time_us": 3.712, "pct_cuda_time": 0.054072605291540496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054072605291540496, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 808.751, "cuda_time_us": 15.232, "pct_cuda_time": 0.22188413895494202, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.424, "pct_cuda_time": 0.1664131042162065, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 159.379, "cuda_time_us": 17.728, "pct_cuda_time": 0.25824330458201233, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.2274778567437221, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.326, "cuda_time_us": 3.073, "pct_cuda_time": 0.04476430928364869, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04476430928364869, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 681.562, "cuda_time_us": 135.999, "pct_cuda_time": 1.9810938165528598, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.829, "cuda_time_us": 83.231, "pct_cuda_time": 1.2124237637446678, "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": 83.231, "pct_cuda_time": 1.2124237637446678, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.537, "cuda_time_us": 8.96, "pct_cuda_time": 0.1305200817382012, "trace": "" }, "children": [ { "entry": { "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": 0.1305200817382012, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 359.714, "cuda_time_us": 43.808, "pct_cuda_time": 0.6381499710699908, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.808, "pct_cuda_time": 0.6381499710699908, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2469.643, "cuda_time_us": 199.804, "pct_cuda_time": 2.910539554868253, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.729, "cuda_time_us": 3.008, "pct_cuda_time": 0.0438174560121104, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0438174560121104, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1764.788, "cuda_time_us": 56.861999999999995, "pct_cuda_time": 0.8283072419416958, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.436, "cuda_time_us": 20.351, "pct_cuda_time": 0.2964524758319344, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.351, "pct_cuda_time": 0.2964524758319344, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.881, "cuda_time_us": 3.712, "pct_cuda_time": 0.054072605291540496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054072605291540496, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 777.73, "cuda_time_us": 15.008, "pct_cuda_time": 0.21862113691148696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.168, "pct_cuda_time": 0.1626839590236865, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.888, "cuda_time_us": 17.791, "pct_cuda_time": 0.25916102390673407, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.679, "pct_cuda_time": 0.2283955760684438, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.727, "cuda_time_us": 3.232, "pct_cuda_time": 0.047080458055565426, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047080458055565426, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.961, "cuda_time_us": 136.702, "pct_cuda_time": 1.9913343988588814, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.117, "cuda_time_us": 83.103, "pct_cuda_time": 1.2105591911484077, "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": 1.2105591911484077, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.857, "cuda_time_us": 9.024, "pct_cuda_time": 0.13145236803633117, "trace": "" }, "children": [ { "entry": { "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": 0.13145236803633117, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.333, "cuda_time_us": 44.575, "pct_cuda_time": 0.6493228396741426, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6493228396741426, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2398.708, "cuda_time_us": 199.293, "pct_cuda_time": 2.9030958314566218, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.042, "cuda_time_us": 2.976, "pct_cuda_time": 0.043351312863045395, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043351312863045395, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1710.257, "cuda_time_us": 57.951, "pct_cuda_time": 0.8441706759833144, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.796, "cuda_time_us": 21.184, "pct_cuda_time": 0.30858676468103285, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.184, "pct_cuda_time": 0.30858676468103285, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 522.349, "cuda_time_us": 3.808, "pct_cuda_time": 0.055471034738735506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.055471034738735506, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 725.488, "cuda_time_us": 15.039, "pct_cuda_time": 0.21907271308714368, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.311, "pct_cuda_time": 0.019097302138256892, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 161.192, "cuda_time_us": 17.92, "pct_cuda_time": 0.2610401634764024, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.808, "pct_cuda_time": 0.2302747156381121, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.876, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 461.766, "cuda_time_us": 135.23, "pct_cuda_time": 1.9698918140018913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.156, "cuda_time_us": 82.847, "pct_cuda_time": 1.2068300459558876, "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": 1.2068300459558876, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.803, "cuda_time_us": 8.96, "pct_cuda_time": 0.1305200817382012, "trace": "" }, "children": [ { "entry": { "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": 0.1305200817382012, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.521, "cuda_time_us": 43.423, "pct_cuda_time": 0.6325416863078025, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.423, "pct_cuda_time": 0.6325416863078025, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2418.429, "cuda_time_us": 200.35, "pct_cuda_time": 2.918493122349175, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.838, "cuda_time_us": 3.168, "pct_cuda_time": 0.04614817175743542, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04614817175743542, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1744.66, "cuda_time_us": 57.086, "pct_cuda_time": 0.8315702439851511, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.091, "cuda_time_us": 20.607, "pct_cuda_time": 0.30018162102445445, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.607, "pct_cuda_time": 0.30018162102445445, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.88, "cuda_time_us": 3.647, "pct_cuda_time": 0.0531257520200022, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.647, "pct_cuda_time": 0.0531257520200022, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 745.006, "cuda_time_us": 15.296, "pct_cuda_time": 0.22281642525307202, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.72, "pct_cuda_time": 0.039622167670525364, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.168, "pct_cuda_time": 0.1626839590236865, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020510298558860187, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.142, "cuda_time_us": 17.536, "pct_cuda_time": 0.2554464456876223, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.22468099784933204, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.602, "cuda_time_us": 3.137, "pct_cuda_time": 0.0456965955817787, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.137, "pct_cuda_time": 0.0456965955817787, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.527, "cuda_time_us": 136.959, "pct_cuda_time": 1.99507811102481, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.059, "cuda_time_us": 84.287, "pct_cuda_time": 1.227806487663813, "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": 1.227806487663813, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.737, "cuda_time_us": 9.248, "pct_cuda_time": 0.13471537007978623, "trace": "" }, "children": [ { "entry": { "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": 0.13471537007978623, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.249, "cuda_time_us": 43.424, "pct_cuda_time": 0.6325562532812108, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.424, "pct_cuda_time": 0.6325562532812108, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2468.735, "cuda_time_us": 199.422, "pct_cuda_time": 2.9049749710262898, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.167, "cuda_time_us": 3.008, "pct_cuda_time": 0.0438174560121104, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0438174560121104, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1759.998, "cuda_time_us": 57.37599999999999, "pct_cuda_time": 0.8357946662735525, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.034, "cuda_time_us": 20.415, "pct_cuda_time": 0.29738476213006443, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.415, "pct_cuda_time": 0.29738476213006443, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.285, "cuda_time_us": 3.68, "pct_cuda_time": 0.05360646214247549, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05360646214247549, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.243, "cuda_time_us": 15.008, "pct_cuda_time": 0.21862113691148696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.518, "cuda_time_us": 18.273, "pct_cuda_time": 0.2661823050895257, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 16.128, "pct_cuda_time": 0.23493614712876212, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.145, "pct_cuda_time": 0.031246157960763563, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.306, "cuda_time_us": 3.168, "pct_cuda_time": 0.04614817175743542, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04614817175743542, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.235, "cuda_time_us": 135.87, "pct_cuda_time": 1.9792146769831913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.753, "cuda_time_us": 83.903, "pct_cuda_time": 1.222212769875033, "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": 1.222212769875033, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.757, "cuda_time_us": 8.863, "pct_cuda_time": 0.1291070853175979, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.863, "pct_cuda_time": 0.1291070853175979, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.467, "cuda_time_us": 43.104, "pct_cuda_time": 0.6278948217905607, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.104, "pct_cuda_time": 0.6278948217905607, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2363.001, "cuda_time_us": 199.007, "pct_cuda_time": 2.898929677061853, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.111, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1687.92, "cuda_time_us": 56.70399999999999, "pct_cuda_time": 0.8260056601431874, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.256, "cuda_time_us": 20.576, "pct_cuda_time": 0.2997300448487977, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.576, "pct_cuda_time": 0.2997300448487977, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 482.232, "cuda_time_us": 3.68, "pct_cuda_time": 0.05360646214247549, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05360646214247549, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.343, "cuda_time_us": 15.008, "pct_cuda_time": 0.21862113691148696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.936, "cuda_time_us": 17.439999999999998, "pct_cuda_time": 0.2540480162404273, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.22328256840213703, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.921, "cuda_time_us": 3.2, "pct_cuda_time": 0.04661431490650043, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04661431490650043, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.935, "cuda_time_us": 136.06300000000002, "pct_cuda_time": 1.9820261028509902, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.474, "cuda_time_us": 83.903, "pct_cuda_time": 1.222212769875033, "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": 1.222212769875033, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.624, "cuda_time_us": 9.024, "pct_cuda_time": 0.13145236803633117, "trace": "" }, "children": [ { "entry": { "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": 0.13145236803633117, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.349, "cuda_time_us": 43.136, "pct_cuda_time": 0.6283609649396257, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.136, "pct_cuda_time": 0.6283609649396257, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2420.999, "cuda_time_us": 198.752, "pct_cuda_time": 2.8952150988427414, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.258, "cuda_time_us": 2.976, "pct_cuda_time": 0.043351312863045395, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.043351312863045395, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1710.752, "cuda_time_us": 56.801, "pct_cuda_time": 0.8274186565637909, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.844, "cuda_time_us": 20.352, "pct_cuda_time": 0.2964670428053427, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.352, "pct_cuda_time": 0.2964670428053427, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 471.927, "cuda_time_us": 3.68, "pct_cuda_time": 0.05360646214247549, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05360646214247549, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 767.429, "cuda_time_us": 15.072, "pct_cuda_time": 0.219553423209617, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.232, "pct_cuda_time": 0.16361624532181648, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.361, "cuda_time_us": 17.697, "pct_cuda_time": 0.2577917284063556, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.584, "pct_cuda_time": 0.22701171359465708, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030780014811698564, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.665, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 480.866, "cuda_time_us": 135.839, "pct_cuda_time": 1.9787631008075346, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 172.652, "cuda_time_us": 82.911, "pct_cuda_time": 1.2077623322540179, "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": 82.911, "pct_cuda_time": 1.2077623322540179, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.925, "cuda_time_us": 9.184, "pct_cuda_time": 0.1337830837816562, "trace": "" }, "children": [ { "entry": { "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": 0.1337830837816562, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.205, "cuda_time_us": 43.744, "pct_cuda_time": 0.6372176847718608, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.744, "pct_cuda_time": 0.6372176847718608, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2390.622, "cuda_time_us": 198.684, "pct_cuda_time": 2.8942245446509784, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.554, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1708.79, "cuda_time_us": 56.926, "pct_cuda_time": 0.829239528239826, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.897, "cuda_time_us": 20.383, "pct_cuda_time": 0.2969186189809994, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.383, "pct_cuda_time": 0.2969186189809994, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 531.967, "cuda_time_us": 3.808, "pct_cuda_time": 0.055471034738735506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.055471034738735506, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.046, "cuda_time_us": 14.975999999999999, "pct_cuda_time": 0.21815499376242198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03635916562707033, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 154.214, "cuda_time_us": 17.759, "pct_cuda_time": 0.25869488075766905, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.615, "pct_cuda_time": 0.2274632897703138, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.144, "pct_cuda_time": 0.031231590987355288, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.708, "cuda_time_us": 3.072, "pct_cuda_time": 0.044749742310240405, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044749742310240405, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.507, "cuda_time_us": 135.646, "pct_cuda_time": 1.9759516749397361, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.266, "cuda_time_us": 83.551, "pct_cuda_time": 1.2170851952353179, "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": 83.551, "pct_cuda_time": 1.2170851952353179, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.679, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.46, "cuda_time_us": 43.007, "pct_cuda_time": 0.6264818253699573, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.007, "pct_cuda_time": 0.6264818253699573, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2706.337, "cuda_time_us": 199.356, "pct_cuda_time": 2.904013550781343, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 90.642, "cuda_time_us": 3.008, "pct_cuda_time": 0.0438174560121104, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0438174560121104, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1972.876, "cuda_time_us": 57.437, "pct_cuda_time": 0.8366832516514577, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 163.861, "cuda_time_us": 21.215, "pct_cuda_time": 0.3090383408566895, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3090383408566895, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 496.079, "cuda_time_us": 3.68, "pct_cuda_time": 0.05360646214247549, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05360646214247549, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 980.941, "cuda_time_us": 15.039, "pct_cuda_time": 0.21907271308714368, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.527, "pct_cuda_time": 0.03681074180272706, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019111869111665174, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.471, "cuda_time_us": 17.503, "pct_cuda_time": 0.254965735565149, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.391, "pct_cuda_time": 0.22420028772685874, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.030765447838290282, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.62, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 486.044, "cuda_time_us": 135.87099999999998, "pct_cuda_time": 1.9792292439565995, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 174.559, "cuda_time_us": 83.455, "pct_cuda_time": 1.2156867657881227, "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": 83.455, "pct_cuda_time": 1.2156867657881227, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.604, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1323846543344612, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.548, "cuda_time_us": 43.328, "pct_cuda_time": 0.6311578238340158, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6311578238340158, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2458.767, "cuda_time_us": 200.99000000000004, "pct_cuda_time": 2.9278159853304753, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.974, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1766.867, "cuda_time_us": 58.24, "pct_cuda_time": 0.8483805312983077, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.808, "cuda_time_us": 21.696, "pct_cuda_time": 0.3160450550660729, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3160450550660729, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 526.076, "cuda_time_us": 3.681, "pct_cuda_time": 0.05362102911588377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.681, "pct_cuda_time": 0.05362102911588377, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 777.063, "cuda_time_us": 15.04, "pct_cuda_time": 0.21908728006055198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.232, "pct_cuda_time": 0.16361624532181648, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.288, "cuda_time_us": 17.823, "pct_cuda_time": 0.2596271670557991, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.2288762861909171, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030750880864882, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.734, "cuda_time_us": 3.168, "pct_cuda_time": 0.04614817175743542, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04614817175743542, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.636, "cuda_time_us": 136.44600000000003, "pct_cuda_time": 1.9876052536663618, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.9, "cuda_time_us": 84.031, "pct_cuda_time": 1.224077342471293, "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": 1.224077342471293, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.112, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.12, "pct_cuda_time": 0.1328507974835262, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.664, "cuda_time_us": 43.295, "pct_cuda_time": 0.6306771137115424, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.295, "pct_cuda_time": 0.6306771137115424, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2458.723, "cuda_time_us": 199.294, "pct_cuda_time": 2.90311039843003, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.489, "cuda_time_us": 2.977, "pct_cuda_time": 0.04336587983645367, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 2.977, "pct_cuda_time": 0.04336587983645367, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1776.138, "cuda_time_us": 57.662, "pct_cuda_time": 0.839960820668321, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.378, "cuda_time_us": 20.864, "pct_cuda_time": 0.3039253331903828, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.864, "pct_cuda_time": 0.3039253331903828, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 513.372, "cuda_time_us": 3.648, "pct_cuda_time": 0.053140318993410485, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053140318993410485, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.747, "cuda_time_us": 15.261999999999999, "pct_cuda_time": 0.22232114815719045, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.527, "pct_cuda_time": 0.03681074180272706, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.231, "pct_cuda_time": 0.1636016783484082, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0219087280060552, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 218.598, "cuda_time_us": 17.887999999999998, "pct_cuda_time": 0.26057402032733734, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.744, "pct_cuda_time": 0.2293424293399821, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.144, "pct_cuda_time": 0.031231590987355288, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.382, "cuda_time_us": 3.136, "pct_cuda_time": 0.045682028608370416, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.045682028608370416, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.882, "cuda_time_us": 135.519, "pct_cuda_time": 1.974101669316885, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.15, "cuda_time_us": 83.199, "pct_cuda_time": 1.2119576205956026, "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": 83.199, "pct_cuda_time": 1.2119576205956026, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.296, "cuda_time_us": 9.024, "pct_cuda_time": 0.13145236803633117, "trace": "" }, "children": [ { "entry": { "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": 0.13145236803633117, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.281, "cuda_time_us": 43.296, "pct_cuda_time": 0.6306916806849507, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.296, "pct_cuda_time": 0.6306916806849507, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2487.864, "cuda_time_us": 199.294, "pct_cuda_time": 2.90311039843003, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.098, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1771.315, "cuda_time_us": 57.025000000000006, "pct_cuda_time": 0.830681658607246, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.46, "cuda_time_us": 20.704, "pct_cuda_time": 0.3015946174450578, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3015946174450578, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 485.487, "cuda_time_us": 3.712, "pct_cuda_time": 0.054072605291540496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054072605291540496, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 792.687, "cuda_time_us": 15.04, "pct_cuda_time": 0.21908728006055198, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.232, "pct_cuda_time": 0.16361624532181648, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018645725962600168, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.994, "cuda_time_us": 17.569, "pct_cuda_time": 0.25592715581009556, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.22514714099839706, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030780014811698564, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.994, "cuda_time_us": 3.2, "pct_cuda_time": 0.04661431490650043, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04661431490650043, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 486.193, "cuda_time_us": 136.029, "pct_cuda_time": 1.981530825755108, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 172.044, "cuda_time_us": 82.751, "pct_cuda_time": 1.2054316165086927, "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": 82.751, "pct_cuda_time": 1.2054316165086927, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.417, "cuda_time_us": 9.055, "pct_cuda_time": 0.1319039442119879, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.055, "pct_cuda_time": 0.1319039442119879, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.636, "cuda_time_us": 44.223, "pct_cuda_time": 0.6441952650344276, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.223, "pct_cuda_time": 0.6441952650344276, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2485.63, "cuda_time_us": 197.278, "pct_cuda_time": 2.8737433800389347, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.711, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1775.014, "cuda_time_us": 56.736000000000004, "pct_cuda_time": 0.8264718032922527, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.863, "cuda_time_us": 20.32, "pct_cuda_time": 0.2960008996562777, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.32, "pct_cuda_time": 0.2960008996562777, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 527.937, "cuda_time_us": 3.713, "pct_cuda_time": 0.054087172264948774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.713, "pct_cuda_time": 0.054087172264948774, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 753.173, "cuda_time_us": 15.232, "pct_cuda_time": 0.22188413895494202, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03682530877613533, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.2, "pct_cuda_time": 0.16315010217275147, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0219087280060552, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 9], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 192.681, "cuda_time_us": 17.471, "pct_cuda_time": 0.25449959241608405, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.36, "pct_cuda_time": 0.22374871155120205, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030750880864882, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.743, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.457, "cuda_time_us": 134.462, "pct_cuda_time": 1.9587043784243312, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.141, "cuda_time_us": 82.59, "pct_cuda_time": 1.2030863337899595, "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": 82.59, "pct_cuda_time": 1.2030863337899595, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.041, "cuda_time_us": 8.928, "pct_cuda_time": 0.1300539385891362, "trace": "" }, "children": [ { "entry": { "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": 0.1300539385891362, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.205, "cuda_time_us": 42.944, "pct_cuda_time": 0.6255641060452357, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.944, "pct_cuda_time": 0.6255641060452357, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.31, "cuda_time_us": 3.04, "pct_cuda_time": 0.0442835991611754, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0442835991611754, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 516.635, "cuda_time_us": 350.619, "pct_cuda_time": 5.107457649438211, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 4.768, "pct_cuda_time": 0.06945532921068563, "trace": "index_select(bfloat16[6, 4096], 0, int64[6])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011187435577560101, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 345.083, "pct_cuda_time": 5.026814884649965, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3902.634, "cuda_time_us": 124.92699999999999, "pct_cuda_time": 1.819808286976368, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.026570159496705242, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.856, "pct_cuda_time": 0.027036302645770248, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.888, "pct_cuda_time": 0.02750244579483525, "trace": "copy_(int32[6], int32[6], True) <- _to_copy(int32[6], 3, 0, None, None, True, None) <- to(int32[6], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.856, "pct_cuda_time": 0.027036302645770248, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.952, "pct_cuda_time": 0.02843473209296526, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.888, "pct_cuda_time": 0.02750244579483525, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 1.92, "pct_cuda_time": 0.027968588943900256, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.416, "pct_cuda_time": 0.06432775457097059, "trace": "copy_(float32[6, 128256], bfloat16[6, 128256], False) <- _to_copy(bfloat16[6, 128256], 6, None, None, None, False, None) <- to(bfloat16[6, 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": 5.504, "pct_cuda_time": 0.08017662163918073, "trace": "div_(float32[6, 128256], bfloat16[6, 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": 34.943, "pct_cuda_time": 0.5090137518055763, "trace": "_softmax(float32[6, 128256], -1, False) <- softmax(float32[6, 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": 28.384, "pct_cuda_time": 0.41346897322065873, "trace": "_log_softmax(float32[6, 128256], -1, False) <- log_softmax(float32[6, 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": 0.026570159496705242, "trace": "copy_(int64[6], int32[6], False) <- _to_copy(int32[6], 4, None, None, None, False, None) <- to(int32[6], 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": 6.208, "pct_cuda_time": 0.09043177091861082, "trace": "index(float32[6, 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": 27.776, "pct_cuda_time": 0.40461225338842366, "trace": "argmax(float32[6, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03915602452146036, "trace": "copy_(int64[6], int64[6], False) <- _to_copy(int64[6], 4, 0, None, None, False, None) <- to(int64[6], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }