{ "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": 8, "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": 88903.76299999999, "pct_cuda_time": 99.4323491994139, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 112.319, "pct_cuda_time": 0.12562057727217882, "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": 112.319, "pct_cuda_time": 0.12562057727217882, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 88746.99699999999, "pct_cuda_time": 99.25701790714233, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 2834.8100000000004, "pct_cuda_time": 3.170527414390666, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 64.351, "pct_cuda_time": 0.07197188158763858, "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": 2770.4590000000003, "pct_cuda_time": 3.0985555328030276, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 19827.230999999992, "pct_cuda_time": 22.17530608293199, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 8815.206999999999, "pct_cuda_time": 9.859163561941891, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 34.59000000000001, "pct_cuda_time": 0.03868638224917125, "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": 8780.617, "pct_cuda_time": 9.820477179692723, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 1754.22, "pct_cuda_time": 1.9619666224094008, "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": 1754.22, "pct_cuda_time": 1.9619666224094008, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 3092.6339999999996, "pct_cuda_time": 3.4588846799879573, "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": 726.5510000000002, "pct_cuda_time": 0.8125940939438457, "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": 2317.157, "pct_cuda_time": 2.5915704375062996, "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": 48.92600000000001, "pct_cuda_time": 0.054720148537813026, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 6165.169999999998, "pct_cuda_time": 6.895291218592743, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 35.839000000000006, "pct_cuda_time": 0.04008329729482649, "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": 6129.331, "pct_cuda_time": 6.855207921297919, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 66084.95599999999, "pct_cuda_time": 73.91118440981967, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 41573.184, "pct_cuda_time": 46.49656223009916, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 36.19200000000001, "pct_cuda_time": 0.04047810194744163, "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": 41536.992, "pct_cuda_time": 46.456084128151716, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 5631.893, "pct_cuda_time": 6.298859941729742, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 5631.893, "pct_cuda_time": 6.298859941729742, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 18879.879, "pct_cuda_time": 21.115762237990776, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 33.18200000000001, "pct_cuda_time": 0.037111637345822505, "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": 18846.697, "pct_cuda_time": 21.078650600644952, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 44.447, "pct_cuda_time": 0.049710714999390414, "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": 44.447, "pct_cuda_time": 0.049710714999390414, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 365.59499999999997, "pct_cuda_time": 0.40889123788336973, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 7.232, "pct_cuda_time": 0.008088462458109466, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 1.312, "pct_cuda_time": 0.0014673759326658768, "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": 357.051, "pct_cuda_time": 0.39933539949259445, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 141.94899999999998, "pct_cuda_time": 0.15875956270273514, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 17.790000000000003, "pct_cuda_time": 0.019896812379669167, "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": 5.215, "pct_cuda_time": 0.00583259564699127, "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": 6.848, "pct_cuda_time": 0.00765898657537799, "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": 37.888, "pct_cuda_time": 0.042374953762838966, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 30.783, "pct_cuda_time": 0.03442853150552872, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 2.113, "pct_cuda_time": 0.0023632357818010653, "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": 7.904, "pct_cuda_time": 0.008840045252889549, "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": 30.08, "pct_cuda_time": 0.03364227748063229, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 3.328, "pct_cuda_time": 0.0037221243170061256, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 81886.48, "cuda_time_us": 88903.76299999999, "pct_cuda_time": 99.4323491994139, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 309.494, "cuda_time_us": 112.319, "pct_cuda_time": 0.12562057727217882, "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": 112.319, "pct_cuda_time": 0.12562057727217882, "trace": "index_select(bfloat16[128256, 4096], 0, int64[4096]) <- embedding(bfloat16[128256, 4096], int64[4096], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3988.076, "cuda_time_us": 2718.8109999999997, "pct_cuda_time": 3.0407910265756435, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 235.674, "cuda_time_us": 64.351, "pct_cuda_time": 0.07197188158763858, "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": 64.351, "pct_cuda_time": 0.07197188158763858, "trace": "_C::rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2951.947, "cuda_time_us": 602.1679999999999, "pct_cuda_time": 0.6734808160225193, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 412.992, "cuda_time_us": 268.828, "pct_cuda_time": 0.30066443386181574, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 268.092, "pct_cuda_time": 0.2998412717532471, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 944.829, "cuda_time_us": 52.799, "pct_cuda_time": 0.05905181544880001, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 52.799, "pct_cuda_time": 0.05905181544880001, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1053.452, "cuda_time_us": 90.84700000000001, "pct_cuda_time": 0.10160571749611046, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.92, "pct_cuda_time": 0.024515914972588426, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 67.647, "pct_cuda_time": 0.07565821624775043, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0014315862757715869, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 294.757, "cuda_time_us": 189.694, "pct_cuda_time": 0.21215884921579325, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.000824280535346609, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 188.957, "pct_cuda_time": 0.21133456868044664, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 111.56, "cuda_time_us": 43.263, "pct_cuda_time": 0.04838649769430169, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.263, "pct_cuda_time": 0.04838649769430169, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 574.058, "cuda_time_us": 2009.029, "pct_cuda_time": 2.246951831271184, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 195.478, "cuda_time_us": 1259.3110000000001, "pct_cuda_time": 1.408447144162651, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1258.575, "pct_cuda_time": 1.4076239820540821, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 129.256, "cuda_time_us": 172.35, "pct_cuda_time": 0.19276085517908828, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.35, "pct_cuda_time": 0.19276085517908828, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 174.675, "cuda_time_us": 577.368, "pct_cuda_time": 0.6457438319294451, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 576.6, "pct_cuda_time": 0.644884880163982, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2695.125, "cuda_time_us": 2698.1079999999997, "pct_cuda_time": 3.017636236991816, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.989, "cuda_time_us": 43.712, "pct_cuda_time": 0.0488886713175997, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.712, "pct_cuda_time": 0.0488886713175997, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1939.172, "cuda_time_us": 595.0, "pct_cuda_time": 0.6654639328781986, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.659, "cuda_time_us": 266.108, "pct_cuda_time": 0.29762231302580117, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.372, "pct_cuda_time": 0.2967991509172325, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 541.279, "cuda_time_us": 53.631, "pct_cuda_time": 0.059982346528051546, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.631, "pct_cuda_time": 0.059982346528051546, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 816.348, "cuda_time_us": 91.295, "pct_cuda_time": 0.1021067726926305, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.432, "pct_cuda_time": 0.02508854948289706, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 67.135, "pct_cuda_time": 0.0750855817374418, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 231.385, "cuda_time_us": 183.96599999999998, "pct_cuda_time": 0.2057525006317154, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.23, "pct_cuda_time": 0.20492933852314676, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 90.434, "cuda_time_us": 43.551, "pct_cuda_time": 0.0487086046063503, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.551, "pct_cuda_time": 0.0487086046063503, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 495.61, "cuda_time_us": 2015.8449999999998, "pct_cuda_time": 2.2545750281896675, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 173.341, "cuda_time_us": 1265.135, "pct_cuda_time": 1.4149608617174114, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1264.399, "pct_cuda_time": 1.4141376996088426, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.933, "cuda_time_us": 172.318, "pct_cuda_time": 0.192725065522194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.318, "pct_cuda_time": 0.192725065522194, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.208, "cuda_time_us": 578.3919999999999, "pct_cuda_time": 0.6468891009500621, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 577.656, "pct_cuda_time": 0.6460659388414935, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2527.601, "cuda_time_us": 2692.191, "pct_cuda_time": 3.0110185057467063, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.52, "cuda_time_us": 43.648, "pct_cuda_time": 0.04881709200381112, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.648, "pct_cuda_time": 0.04881709200381112, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1801.7, "cuda_time_us": 593.721, "pct_cuda_time": 0.664033465029205, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 158.851, "cuda_time_us": 266.012, "pct_cuda_time": 0.29751494405511825, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.276, "pct_cuda_time": 0.2966917819465496, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.775, "cuda_time_us": 53.088, "pct_cuda_time": 0.05937504078762657, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.088, "pct_cuda_time": 0.05937504078762657, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 751.704, "cuda_time_us": 91.455, "pct_cuda_time": 0.10228572097710194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.824, "pct_cuda_time": 0.02440854600190556, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 68.191, "pct_cuda_time": 0.07626664041495335, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016105345602430354, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.192, "cuda_time_us": 183.166, "pct_cuda_time": 0.20485775920935817, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 182.398, "pct_cuda_time": 0.20399880744389523, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.49, "cuda_time_us": 42.783, "pct_cuda_time": 0.047849652840887345, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 42.783, "pct_cuda_time": 0.047849652840887345, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.798, "cuda_time_us": 2012.039, "pct_cuda_time": 2.250318295872803, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.661, "cuda_time_us": 1261.776, "pct_cuda_time": 1.411204066170289, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1261.04, "pct_cuda_time": 1.4103809040617203, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.239, "cuda_time_us": 172.318, "pct_cuda_time": 0.192725065522194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.318, "pct_cuda_time": 0.192725065522194, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.123, "cuda_time_us": 577.9449999999999, "pct_cuda_time": 0.6463891641803201, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 577.209, "pct_cuda_time": 0.6455660020717515, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2423.704, "cuda_time_us": 2694.173, "pct_cuda_time": 3.013235227620596, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.94, "cuda_time_us": 43.839, "pct_cuda_time": 0.049030711518398905, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.049030711518398905, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1737.459, "cuda_time_us": 595.193, "pct_cuda_time": 0.6656797892463423, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.485, "cuda_time_us": 266.493, "pct_cuda_time": 0.2980529073353106, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.757, "pct_cuda_time": 0.2972297452267419, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 485.227, "cuda_time_us": 53.376, "pct_cuda_time": 0.05969714769967517, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.376, "pct_cuda_time": 0.05969714769967517, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 752.077, "cuda_time_us": 91.038, "pct_cuda_time": 0.10181933701069823, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.463, "pct_cuda_time": 0.025123220713013407, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 67.327, "pct_cuda_time": 0.07530031967880753, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 198.658, "cuda_time_us": 184.286, "pct_cuda_time": 0.20611039720065832, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.55, "pct_cuda_time": 0.20528723509208968, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.85, "cuda_time_us": 44.096, "pct_cuda_time": 0.04931814720033116, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.096, "pct_cuda_time": 0.04931814720033116, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 469.116, "cuda_time_us": 2011.0449999999998, "pct_cuda_time": 2.249206579655524, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.111, "cuda_time_us": 1260.655, "pct_cuda_time": 1.4099503097522108, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1259.92, "pct_cuda_time": 1.40912826607042, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.837, "cuda_time_us": 172.157, "pct_cuda_time": 0.1925449988109446, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.157, "pct_cuda_time": 0.1925449988109446, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.154, "cuda_time_us": 578.233, "pct_cuda_time": 0.6467112710923687, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 577.497, "pct_cuda_time": 0.6458881089838, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2589.491, "cuda_time_us": 2696.9260000000004, "pct_cuda_time": 3.016314256540284, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.825, "cuda_time_us": 43.455, "pct_cuda_time": 0.04860123563566743, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.455, "pct_cuda_time": 0.04860123563566743, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1862.428, "cuda_time_us": 595.385, "pct_cuda_time": 0.665894527187708, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.281, "cuda_time_us": 266.493, "pct_cuda_time": 0.2980529073353106, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.245, "pct_cuda_time": 0.2966571107164333, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 522.313, "cuda_time_us": 53.919, "pct_cuda_time": 0.06030445344010015, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.919, "pct_cuda_time": 0.06030445344010015, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 787.218, "cuda_time_us": 90.592, "pct_cuda_time": 0.10132051866773406, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.952, "pct_cuda_time": 0.02455170462948272, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 67.392, "pct_cuda_time": 0.07537301741937405, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 228.713, "cuda_time_us": 184.381, "pct_cuda_time": 0.20621664774456328, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.645, "pct_cuda_time": 0.2053934856359946, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.358, "cuda_time_us": 43.488, "pct_cuda_time": 0.048638143719339666, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.488, "pct_cuda_time": 0.048638143719339666, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 494.908, "cuda_time_us": 2014.5980000000002, "pct_cuda_time": 2.2531803499975687, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.114, "cuda_time_us": 1264.689, "pct_cuda_time": 1.4144620433744473, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.000824280535346609, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1263.952, "pct_cuda_time": 1.4136377628391006, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.997, "cuda_time_us": 171.806, "pct_cuda_time": 0.19215243101188537, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 171.806, "pct_cuda_time": 0.19215243101188537, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 161.162, "cuda_time_us": 578.1030000000001, "pct_cuda_time": 0.6465658756112358, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 577.368, "pct_cuda_time": 0.6457438319294451, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2568.15, "cuda_time_us": 2695.933, "pct_cuda_time": 3.0152036587497824, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.611, "cuda_time_us": 43.904, "pct_cuda_time": 0.04910340925896544, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.904, "pct_cuda_time": 0.04910340925896544, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1806.793, "cuda_time_us": 595.3199999999999, "pct_cuda_time": 0.6658218294471415, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.983, "cuda_time_us": 266.301, "pct_cuda_time": 0.2978381693939448, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.565, "pct_cuda_time": 0.29701500728537616, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 545.981, "cuda_time_us": 53.503, "pct_cuda_time": 0.059839187900474385, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.503, "pct_cuda_time": 0.059839187900474385, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 745.789, "cuda_time_us": 91.615, "pct_cuda_time": 0.10246466926157338, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.856, "pct_cuda_time": 0.02444433565879985, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 68.287, "pct_cuda_time": 0.07637400938563622, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.001646324217137325, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.62, "cuda_time_us": 183.90099999999998, "pct_cuda_time": 0.2056798028911489, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.165, "pct_cuda_time": 0.20485664078258026, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 120.625, "cuda_time_us": 43.008, "pct_cuda_time": 0.04810129886592532, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.008, "pct_cuda_time": 0.04810129886592532, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 480.791, "cuda_time_us": 2013.701, "pct_cuda_time": 2.25217712117775, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.799, "cuda_time_us": 1263.2150000000001, "pct_cuda_time": 1.4128134823037541, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1262.479, "pct_cuda_time": 1.4119903201951853, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.257, "cuda_time_us": 171.326, "pct_cuda_time": 0.19161558615847102, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 171.326, "pct_cuda_time": 0.19161558615847102, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.338, "cuda_time_us": 579.16, "pct_cuda_time": 0.6477480527155252, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 578.424, "pct_cuda_time": 0.6469248906069566, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2359.353, "cuda_time_us": 2700.538, "pct_cuda_time": 3.0203540140622263, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.613, "cuda_time_us": 44.767, "pct_cuda_time": 0.05006861156833331, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.767, "pct_cuda_time": 0.05006861156833331, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1683.622, "cuda_time_us": 596.247, "pct_cuda_time": 0.6668586110702979, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.975, "cuda_time_us": 266.10799999999995, "pct_cuda_time": 0.29762231302580106, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.34, "pct_cuda_time": 0.29676336126033814, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 469.921, "cuda_time_us": 53.663, "pct_cuda_time": 0.060018136184945836, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.663, "pct_cuda_time": 0.060018136184945836, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.858, "cuda_time_us": 91.74300000000001, "pct_cuda_time": 0.10260782788915056, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.984, "pct_cuda_time": 0.024587494286377007, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 68.191, "pct_cuda_time": 0.07626664041495335, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.001753693187820194, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.023, "cuda_time_us": 184.733, "pct_cuda_time": 0.20661033397040046, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.997, "pct_cuda_time": 0.2057871718618318, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.734, "cuda_time_us": 42.687, "pct_cuda_time": 0.047742283870204474, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 42.687, "pct_cuda_time": 0.047742283870204474, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.287, "cuda_time_us": 2016.837, "pct_cuda_time": 2.255684507553391, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.57, "cuda_time_us": 1263.375, "pct_cuda_time": 1.4129924305882255, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1262.639, "pct_cuda_time": 1.4121692684796567, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.178, "cuda_time_us": 172.478, "pct_cuda_time": 0.19290401380666547, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.478, "pct_cuda_time": 0.19290401380666547, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.554, "cuda_time_us": 580.984, "pct_cuda_time": 0.6497880631584998, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 579.736, "pct_cuda_time": 0.6483922665396223, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2474.226, "cuda_time_us": 2696.573, "pct_cuda_time": 3.015919451887668, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.691, "cuda_time_us": 44.192, "pct_cuda_time": 0.04942551617101404, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.192, "pct_cuda_time": 0.04942551617101404, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1778.514, "cuda_time_us": 594.744, "pct_cuda_time": 0.6651776156230443, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.817, "cuda_time_us": 267.164, "pct_cuda_time": 0.29880337170331267, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 266.396, "pct_cuda_time": 0.29794441993784976, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.802, "cuda_time_us": 52.831, "pct_cuda_time": 0.059087605105694316, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 52.831, "pct_cuda_time": 0.059087605105694316, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 779.669, "cuda_time_us": 89.887, "pct_cuda_time": 0.10053202778928173, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.728, "pct_cuda_time": 0.02430117703122269, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 66.719, "pct_cuda_time": 0.07462031619781602, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016105345602430354, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.865, "cuda_time_us": 184.862, "pct_cuda_time": 0.20675461102475554, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.134, "pct_cuda_time": 0.20482196955246387, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.367, "cuda_time_us": 42.879, "pct_cuda_time": 0.04795702181157021, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 42.879, "pct_cuda_time": 0.04795702181157021, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.021, "cuda_time_us": 2014.7579999999998, "pct_cuda_time": 2.2533592982820396, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.995, "cuda_time_us": 1262.928, "pct_cuda_time": 1.4124924938184835, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1262.192, "pct_cuda_time": 1.4116693317099147, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.775, "cuda_time_us": 172.542, "pct_cuda_time": 0.19297559312045404, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.542, "pct_cuda_time": 0.19297559312045404, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.079, "cuda_time_us": 579.2879999999999, "pct_cuda_time": 0.6478912113431022, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 577.56, "pct_cuda_time": 0.6459585698708107, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2419.011, "cuda_time_us": 2697.179, "pct_cuda_time": 3.0165972185151038, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.868, "cuda_time_us": 44.096, "pct_cuda_time": 0.04931814720033116, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.096, "pct_cuda_time": 0.04931814720033116, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1735.108, "cuda_time_us": 594.551, "pct_cuda_time": 0.6649617592549006, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.704, "cuda_time_us": 266.58799999999997, "pct_cuda_time": 0.29815915787921543, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 265.852, "pct_cuda_time": 0.2973359957706468, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.268, "cuda_time_us": 53.12, "pct_cuda_time": 0.05941083044452085, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.12, "pct_cuda_time": 0.05941083044452085, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.818, "cuda_time_us": 90.81400000000001, "pct_cuda_time": 0.10156880941243822, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.432, "pct_cuda_time": 0.02508854948289706, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 66.911, "pct_cuda_time": 0.07483505413918176, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.0016452057903593787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.918, "cuda_time_us": 184.02900000000002, "pct_cuda_time": 0.2058229615187261, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.294, "pct_cuda_time": 0.2050009178369354, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.92, "cuda_time_us": 43.359, "pct_cuda_time": 0.048493866664984564, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.359, "pct_cuda_time": 0.048493866664984564, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.353, "cuda_time_us": 2015.1729999999998, "pct_cuda_time": 2.2538234453948873, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.713, "cuda_time_us": 1263.695, "pct_cuda_time": 1.4133503271571684, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0014673759326658768, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1262.383, "pct_cuda_time": 1.4118829512245026, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.001, "cuda_time_us": 172.509, "pct_cuda_time": 0.19293868503678177, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 172.509, "pct_cuda_time": 0.19293868503678177, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 163.845, "cuda_time_us": 578.9689999999999, "pct_cuda_time": 0.6475344332009373, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 578.233, "pct_cuda_time": 0.6467112710923687, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2461.22, "cuda_time_us": 2705.727, "pct_cuda_time": 3.026157530612991, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.349, "cuda_time_us": 44.191, "pct_cuda_time": 0.0494243977442361, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.191, "pct_cuda_time": 0.0494243977442361, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1725.55, "cuda_time_us": 597.722, "pct_cuda_time": 0.6685082905677691, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.21, "cuda_time_us": 267.549, "pct_cuda_time": 0.2992339660128221, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0015031655895601662, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 266.205, "pct_cuda_time": 0.29773080042326194, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.7, "cuda_time_us": 53.824, "pct_cuda_time": 0.06019820289619522, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.824, "pct_cuda_time": 0.06019820289619522, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.408, "cuda_time_us": 91.2, "pct_cuda_time": 0.10200052214872557, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.208, "pct_cuda_time": 0.024838021884637032, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 67.36, "pct_cuda_time": 0.07533722776247975, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0018252725016087731, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.35, "cuda_time_us": 185.149, "pct_cuda_time": 0.20707559951002621, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.375, "pct_cuda_time": 0.0015378368196765093, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 183.774, "pct_cuda_time": 0.2055377626903497, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.022, "cuda_time_us": 42.783, "pct_cuda_time": 0.047849652840887345, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 42.783, "pct_cuda_time": 0.047849652840887345, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 482.827, "cuda_time_us": 2021.031, "pct_cuda_time": 2.2603751894600985, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.427, "cuda_time_us": 1268.88, "pct_cuda_time": 1.4191493700008213, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1267.152, "pct_cuda_time": 1.4172167285285298, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.854, "cuda_time_us": 171.774, "pct_cuda_time": 0.19211664135499107, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 171.774, "pct_cuda_time": 0.19211664135499107, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.565, "cuda_time_us": 580.3770000000001, "pct_cuda_time": 0.6491091781042863, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0015031655895601662, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 579.033, "pct_cuda_time": 0.647606012514726, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2447.687, "cuda_time_us": 2707.102, "pct_cuda_time": 3.0276953674326674, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.131, "cuda_time_us": 44.704, "pct_cuda_time": 0.04999815068132267, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.704, "pct_cuda_time": 0.04999815068132267, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1758.98, "cuda_time_us": 596.856, "pct_cuda_time": 0.6675397329780673, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 157.571, "cuda_time_us": 267.99699999999996, "pct_cuda_time": 0.2997350212093421, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.984, "pct_cuda_time": 0.0022189587274459594, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 266.013, "pct_cuda_time": 0.2975160624818962, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 483.922, "cuda_time_us": 53.568, "pct_cuda_time": 0.05991188564104091, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.568, "pct_cuda_time": 0.05991188564104091, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 761.714, "cuda_time_us": 90.749, "pct_cuda_time": 0.10149611167187167, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 21.855, "pct_cuda_time": 0.0244432172320219, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 67.359, "pct_cuda_time": 0.07533610933570181, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.535, "pct_cuda_time": 0.0017167851041479578, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 198.27, "cuda_time_us": 184.542, "pct_cuda_time": 0.20639671445581267, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001968431129185932, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 182.782, "pct_cuda_time": 0.20442828332662674, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.031, "cuda_time_us": 42.943, "pct_cuda_time": 0.04802860112535879, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 42.943, "pct_cuda_time": 0.04802860112535879, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.862, "cuda_time_us": 2022.5990000000002, "pct_cuda_time": 2.262128882647919, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.308, "cuda_time_us": 1265.4560000000001, "pct_cuda_time": 1.4153198767131325, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.92, "pct_cuda_time": 0.00214737941365738, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1263.536, "pct_cuda_time": 1.4131724972994748, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.166, "cuda_time_us": 171.998, "pct_cuda_time": 0.19236716895325107, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 171.998, "pct_cuda_time": 0.19236716895325107, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.921, "cuda_time_us": 585.1450000000001, "pct_cuda_time": 0.6544418369815355, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0015031655895601662, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 583.801, "pct_cuda_time": 0.6529386713919753, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2528.052, "cuda_time_us": 2727.5170000000003, "pct_cuda_time": 3.0505280501044467, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.093, "cuda_time_us": 43.999, "pct_cuda_time": 0.04920965980287035, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.999, "pct_cuda_time": 0.04920965980287035, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1810.221, "cuda_time_us": 608.538, "pct_cuda_time": 0.680605194598039, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.928, "cuda_time_us": 271.93199999999996, "pct_cuda_time": 0.3041360305805618, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 271.164, "pct_cuda_time": 0.3032770788150989, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.546, "cuda_time_us": 53.792, "pct_cuda_time": 0.060162413239300945, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 53.792, "pct_cuda_time": 0.060162413239300945, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 768.177, "cuda_time_us": 93.44, "pct_cuda_time": 0.10450579813132584, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.176, "pct_cuda_time": 0.024802232227742742, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 69.696, "pct_cuda_time": 0.0779498727157629, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.001753693187820194, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 210.788, "cuda_time_us": 189.37400000000002, "pct_cuda_time": 0.21180095264685042, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.0020400104429745118, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 187.55, "pct_cuda_time": 0.2097609422038759, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.782, "cuda_time_us": 43.711, "pct_cuda_time": 0.048887552890821744, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.711, "pct_cuda_time": 0.048887552890821744, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.771, "cuda_time_us": 2031.2690000000002, "pct_cuda_time": 2.2718256428127153, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.534, "cuda_time_us": 1267.824, "pct_cuda_time": 1.41796831132331, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1267.088, "pct_cuda_time": 1.4171451492147409, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.529, "cuda_time_us": 175.613, "pct_cuda_time": 0.19641028175552788, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 175.613, "pct_cuda_time": 0.19641028175552788, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.61, "cuda_time_us": 587.832, "pct_cuda_time": 0.6574470497338777, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0017179035309259042, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 586.296, "pct_cuda_time": 0.6557291462029519, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2516.877, "cuda_time_us": 2747.0370000000003, "pct_cuda_time": 3.0723597408099637, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.003, "cuda_time_us": 44.352, "pct_cuda_time": 0.049604464455485484, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.352, "pct_cuda_time": 0.049604464455485484, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1822.869, "cuda_time_us": 620.471, "pct_cuda_time": 0.6939513813392753, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.352, "cuda_time_us": 276.54, "pct_cuda_time": 0.30928974117333957, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.216, "pct_cuda_time": 0.0013600069619830076, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 275.324, "pct_cuda_time": 0.30792973421135655, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.619, "cuda_time_us": 54.847, "pct_cuda_time": 0.06134235349003456, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 54.847, "pct_cuda_time": 0.06134235349003456, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 812.91, "cuda_time_us": 97.087, "pct_cuda_time": 0.10858470059049692, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.688, "pct_cuda_time": 0.02537486673805138, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 72.895, "pct_cuda_time": 0.08152771997841392, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016821138740316145, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.857, "cuda_time_us": 191.99699999999999, "pct_cuda_time": 0.2147345860854042, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 191.261, "pct_cuda_time": 0.21391142397683555, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.658, "cuda_time_us": 43.103, "pct_cuda_time": 0.04820754940983025, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.103, "pct_cuda_time": 0.04820754940983025, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 466.971, "cuda_time_us": 2039.111, "pct_cuda_time": 2.280596345605372, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.772, "cuda_time_us": 1275.055, "pct_cuda_time": 1.4260556553546413, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1273.807, "pct_cuda_time": 1.424659858735764, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.088, "cuda_time_us": 176.094, "pct_cuda_time": 0.1969482450357202, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 176.094, "pct_cuda_time": 0.1969482450357202, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.889, "cuda_time_us": 587.962, "pct_cuda_time": 0.6575924452150108, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.000824280535346609, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 587.225, "pct_cuda_time": 0.6567681646796641, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2361.393, "cuda_time_us": 2742.3979999999997, "pct_cuda_time": 3.0671713589870686, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.738, "cuda_time_us": 44.639, "pct_cuda_time": 0.04992545294075615, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.639, "pct_cuda_time": 0.04992545294075615, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1662.873, "cuda_time_us": 617.531, "pct_cuda_time": 0.6906632066121123, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.042, "cuda_time_us": 275.453, "pct_cuda_time": 0.30807401126571166, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 274.717, "pct_cuda_time": 0.30725084915714296, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.527, "cuda_time_us": 54.272, "pct_cuda_time": 0.060699258092715286, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 54.272, "pct_cuda_time": 0.060699258092715286, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 708.892, "cuda_time_us": 95.807, "pct_cuda_time": 0.10715311431472532, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.56, "pct_cuda_time": 0.02523170811047422, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 71.712, "pct_cuda_time": 0.08020462110010315, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.535, "pct_cuda_time": 0.0017167851041479578, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 185.675, "cuda_time_us": 191.999, "pct_cuda_time": 0.21473682293896007, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.249, "pct_cuda_time": 0.001396915045655244, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 190.75, "pct_cuda_time": 0.21333990789330481, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.334, "cuda_time_us": 43.071, "pct_cuda_time": 0.04817175975293595, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.071, "pct_cuda_time": 0.04817175975293595, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.119, "cuda_time_us": 2037.1569999999997, "pct_cuda_time": 2.2784109396812644, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 177.891, "cuda_time_us": 1274.0629999999999, "pct_cuda_time": 1.424946175990918, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0018968518153973527, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1272.367, "pct_cuda_time": 1.4230493241755209, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.282, "cuda_time_us": 175.646, "pct_cuda_time": 0.19644718983920012, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 175.646, "pct_cuda_time": 0.19644718983920012, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.875, "cuda_time_us": 587.448, "pct_cuda_time": 0.6570175738511462, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 586.712, "pct_cuda_time": 0.6561944117425775, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2416.504, "cuda_time_us": 2755.52, "pct_cuda_time": 3.081847355167284, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.904, "cuda_time_us": 45.184, "pct_cuda_time": 0.05053499553473701, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.184, "pct_cuda_time": 0.05053499553473701, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1735.447, "cuda_time_us": 621.656, "pct_cuda_time": 0.6952767170711418, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.597, "cuda_time_us": 277.212, "pct_cuda_time": 0.3100413239681197, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.856, "pct_cuda_time": 0.0020758000998688008, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 275.356, "pct_cuda_time": 0.3079655238682508, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 495.563, "cuda_time_us": 54.655, "pct_cuda_time": 0.06112761554866882, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 54.655, "pct_cuda_time": 0.06112761554866882, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.908, "cuda_time_us": 96.89500000000001, "pct_cuda_time": 0.1083699626491312, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.592, "pct_cuda_time": 0.025267497767368505, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 72.799, "pct_cuda_time": 0.08142035100773107, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016821138740316145, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.177, "cuda_time_us": 192.89399999999998, "pct_cuda_time": 0.2157378149052222, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 192.158, "pct_cuda_time": 0.21491465279665356, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.942, "cuda_time_us": 43.552, "pct_cuda_time": 0.04870972303312825, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.552, "pct_cuda_time": 0.04870972303312825, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.991, "cuda_time_us": 2045.128, "pct_cuda_time": 2.2873259195282767, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.545, "cuda_time_us": 1273.872, "pct_cuda_time": 1.4247325564763305, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.0016105345602430354, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1272.432, "pct_cuda_time": 1.4231220219160874, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.053, "cuda_time_us": 176.158, "pct_cuda_time": 0.19701982434950877, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 176.158, "pct_cuda_time": 0.19701982434950877, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.969, "cuda_time_us": 595.098, "pct_cuda_time": 0.6655735387024373, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.000824280535346609, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 594.361, "pct_cuda_time": 0.6647492581670907, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2412.612, "cuda_time_us": 2838.4579999999996, "pct_cuda_time": 3.1746074352766143, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.246, "cuda_time_us": 45.375, "pct_cuda_time": 0.05074861504932481, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.375, "pct_cuda_time": 0.05074861504932481, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1713.569, "cuda_time_us": 636.184, "pct_cuda_time": 0.7115252213011494, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.421, "cuda_time_us": 281.372, "pct_cuda_time": 0.3146939793643773, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 280.636, "pct_cuda_time": 0.3138708172558087, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.399, "cuda_time_us": 55.967, "pct_cuda_time": 0.0625949914813347, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.967, "pct_cuda_time": 0.0625949914813347, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 754.596, "cuda_time_us": 100.319, "pct_cuda_time": 0.11219945593682018, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.976, "pct_cuda_time": 0.025696973650099986, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.839, "pct_cuda_time": 0.08482036841268857, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016821138740316145, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 198.915, "cuda_time_us": 198.526, "pct_cuda_time": 0.22203679451861724, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0018968518153973527, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.83, "pct_cuda_time": 0.22013994270321988, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.533, "cuda_time_us": 44.319, "pct_cuda_time": 0.04956755637181325, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.319, "pct_cuda_time": 0.04956755637181325, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.318, "cuda_time_us": 2112.58, "pct_cuda_time": 2.3627660425543273, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.654, "cuda_time_us": 1333.582, "pct_cuda_time": 1.4915138193875193, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1332.814, "pct_cuda_time": 1.4906548676220563, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.738, "cuda_time_us": 178.75, "pct_cuda_time": 0.19991878655794623, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.75, "pct_cuda_time": 0.19991878655794623, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.118, "cuda_time_us": 600.2479999999999, "pct_cuda_time": 0.671333436608862, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0018252725016087731, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 598.616, "pct_cuda_time": 0.6695081641072533, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2430.839, "cuda_time_us": 2832.058, "pct_cuda_time": 3.167449503897757, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.598, "cuda_time_us": 44.287, "pct_cuda_time": 0.04953176671491896, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.287, "pct_cuda_time": 0.04953176671491896, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1753.637, "cuda_time_us": 634.4879999999999, "pct_cuda_time": 0.7096283694857519, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.257, "cuda_time_us": 280.477, "pct_cuda_time": 0.3136929873981151, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.741, "pct_cuda_time": 0.3128698252895465, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 515.37, "cuda_time_us": 55.871, "pct_cuda_time": 0.062487622510651826, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.871, "pct_cuda_time": 0.062487622510651826, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 751.708, "cuda_time_us": 99.615, "pct_cuda_time": 0.11141208348514581, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.135, "pct_cuda_time": 0.02587480350779349, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 74.975, "pct_cuda_time": 0.08385404767654275, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.001683232300809561, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 187.01, "cuda_time_us": 198.525, "pct_cuda_time": 0.2220356760918393, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 197.79, "pct_cuda_time": 0.22121363241004854, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.729, "cuda_time_us": 43.103, "pct_cuda_time": 0.04820754940983025, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.103, "pct_cuda_time": 0.04820754940983025, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.983, "cuda_time_us": 2110.18, "pct_cuda_time": 2.360081818287256, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.497, "cuda_time_us": 1335.023, "pct_cuda_time": 1.49312547237454, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0014315862757715869, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1333.743, "pct_cuda_time": 1.4916938860987683, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.588, "cuda_time_us": 177.853, "pct_cuda_time": 0.1989155577381282, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 177.853, "pct_cuda_time": 0.1989155577381282, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.389, "cuda_time_us": 597.304, "pct_cuda_time": 0.6680407881745875, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 596.536, "pct_cuda_time": 0.6671818364091244, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2389.922, "cuda_time_us": 2830.008, "pct_cuda_time": 3.1651567290029665, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.382, "cuda_time_us": 45.311, "pct_cuda_time": 0.05067703573553623, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.311, "pct_cuda_time": 0.05067703573553623, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1725.841, "cuda_time_us": 634.998, "pct_cuda_time": 0.7101987671425048, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 184.035, "cuda_time_us": 280.571, "pct_cuda_time": 0.3137981195152421, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.836, "pct_cuda_time": 0.31297607583345144, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 490.306, "cuda_time_us": 55.423, "pct_cuda_time": 0.06198656731413177, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.423, "pct_cuda_time": 0.06198656731413177, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 725.42, "cuda_time_us": 100.79899999999999, "pct_cuda_time": 0.11273630079023451, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.36, "pct_cuda_time": 0.02612644953283146, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.711, "pct_cuda_time": 0.08467720978511142, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 184.255, "cuda_time_us": 198.20499999999998, "pct_cuda_time": 0.22167777952289638, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.957, "pct_cuda_time": 0.2202819829040191, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.44, "cuda_time_us": 44.256, "pct_cuda_time": 0.04949709548480261, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.256, "pct_cuda_time": 0.04949709548480261, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.835, "cuda_time_us": 2105.4429999999998, "pct_cuda_time": 2.3547838306401223, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.235, "cuda_time_us": 1329.55, "pct_cuda_time": 1.4870043226188385, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.952, "pct_cuda_time": 0.00218316907055167, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1327.598, "pct_cuda_time": 1.484821153548287, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.184, "cuda_time_us": 178.846, "pct_cuda_time": 0.2000261555286291, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.846, "pct_cuda_time": 0.2000261555286291, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.256, "cuda_time_us": 597.0469999999999, "pct_cuda_time": 0.6677533524926551, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.343, "pct_cuda_time": 0.0015020471627822196, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.704, "pct_cuda_time": 0.666251305329873, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2467.044, "cuda_time_us": 2832.571, "pct_cuda_time": 3.1680232568348434, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.501, "cuda_time_us": 44.479, "pct_cuda_time": 0.0497465046562847, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.479, "pct_cuda_time": 0.0497465046562847, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1749.555, "cuda_time_us": 635.8639999999999, "pct_cuda_time": 0.7111673247322065, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.037, "cuda_time_us": 281.916, "pct_cuda_time": 0.31530240353158023, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0017179035309259042, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 280.38, "pct_cuda_time": 0.31358450000065435, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.559, "cuda_time_us": 55.871, "pct_cuda_time": 0.062487622510651826, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.871, "pct_cuda_time": 0.062487622510651826, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 724.384, "cuda_time_us": 100.83099999999999, "pct_cuda_time": 0.1127720904471288, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.136, "pct_cuda_time": 0.025875921934571434, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.871, "pct_cuda_time": 0.08485615806958287, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.0020400104429745118, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 256.853, "cuda_time_us": 197.24599999999998, "pct_cuda_time": 0.22060520824284563, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.51, "pct_cuda_time": 0.21978204613427696, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.088, "cuda_time_us": 43.872, "pct_cuda_time": 0.049067619602071136, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.872, "pct_cuda_time": 0.049067619602071136, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 486.333, "cuda_time_us": 2108.356, "pct_cuda_time": 2.3580418078442813, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 179.101, "cuda_time_us": 1333.038, "pct_cuda_time": 1.4909053952203162, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0014315862757715869, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1331.758, "pct_cuda_time": 1.4894738089445447, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.865, "cuda_time_us": 178.014, "pct_cuda_time": 0.1990956244493776, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.014, "pct_cuda_time": 0.1990956244493776, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.344, "cuda_time_us": 597.304, "pct_cuda_time": 0.6680407881745875, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 596.536, "pct_cuda_time": 0.6671818364091244, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2547.129, "cuda_time_us": 2831.835, "pct_cuda_time": 3.1672000947262746, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.132, "cuda_time_us": 45.439, "pct_cuda_time": 0.050820194363113386, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.439, "pct_cuda_time": 0.050820194363113386, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1858.352, "cuda_time_us": 638.232, "pct_cuda_time": 0.713815759342384, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.978, "cuda_time_us": 282.49199999999996, "pct_cuda_time": 0.3159466173556774, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0014315862757715869, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 281.212, "pct_cuda_time": 0.31451503107990586, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 532.51, "cuda_time_us": 56.0, "pct_cuda_time": 0.06263189956500693, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 56.0, "pct_cuda_time": 0.06263189956500693, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 830.606, "cuda_time_us": 101.08500000000001, "pct_cuda_time": 0.11305617084872725, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.711, "pct_cuda_time": 0.026519017331890698, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.998, "pct_cuda_time": 0.08499819827038209, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.376, "pct_cuda_time": 0.0015389552464544559, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.275, "cuda_time_us": 198.655, "pct_cuda_time": 0.22218107157297237, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.761, "pct_cuda_time": 0.001969549555963879, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.894, "pct_cuda_time": 0.22021152201700844, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.239, "cuda_time_us": 43.583, "pct_cuda_time": 0.04874439426324458, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.583, "pct_cuda_time": 0.04874439426324458, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.108, "cuda_time_us": 2104.581, "pct_cuda_time": 2.353819746757533, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.816, "cuda_time_us": 1327.7910000000002, "pct_cuda_time": 1.485037009916431, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1327.055, "pct_cuda_time": 1.484213847807862, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.787, "cuda_time_us": 178.526, "pct_cuda_time": 0.19966825895968618, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.526, "pct_cuda_time": 0.19966825895968618, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.254, "cuda_time_us": 598.264, "pct_cuda_time": 0.6691144778814162, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.0017894828447144837, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 596.664, "pct_cuda_time": 0.6673249950367017, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2409.814, "cuda_time_us": 2822.107, "pct_cuda_time": 3.156320039030411, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.438, "cuda_time_us": 45.024, "pct_cuda_time": 0.05035604725026557, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.024, "pct_cuda_time": 0.05035604725026557, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1742.913, "cuda_time_us": 634.423, "pct_cuda_time": 0.7095556717451855, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.1, "cuda_time_us": 280.02799999999996, "pct_cuda_time": 0.3131908137748171, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.26, "pct_cuda_time": 0.3123318620093542, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 549.69, "cuda_time_us": 56.159, "pct_cuda_time": 0.06280972942270042, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 56.159, "pct_cuda_time": 0.06280972942270042, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 713.468, "cuda_time_us": 100.51100000000001, "pct_cuda_time": 0.11241419387818592, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.976, "pct_cuda_time": 0.025696973650099986, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 76.031, "pct_cuda_time": 0.08503510635405433, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016821138740316145, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.955, "cuda_time_us": 197.725, "pct_cuda_time": 0.22114093466948206, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.989, "pct_cuda_time": 0.22031777256091337, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.933, "cuda_time_us": 43.552, "pct_cuda_time": 0.04870972303312825, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.552, "pct_cuda_time": 0.04870972303312825, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.332, "cuda_time_us": 2099.108, "pct_cuda_time": 2.347698597001832, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.86, "cuda_time_us": 1324.782, "pct_cuda_time": 1.4816716637415894, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.216, "pct_cuda_time": 0.0013600069619830076, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1323.566, "pct_cuda_time": 1.4803116567796066, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.669, "cuda_time_us": 178.334, "pct_cuda_time": 0.19945352101832045, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.334, "pct_cuda_time": 0.19945352101832045, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.944, "cuda_time_us": 595.992, "pct_cuda_time": 0.6665734122419216, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.257, "pct_cuda_time": 0.6657513685601307, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2466.427, "cuda_time_us": 2830.843, "pct_cuda_time": 3.1660906153625517, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.818, "cuda_time_us": 44.383, "pct_cuda_time": 0.04963913568560184, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.383, "pct_cuda_time": 0.04963913568560184, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1741.403, "cuda_time_us": 633.752, "pct_cuda_time": 0.7088052073771833, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.222, "cuda_time_us": 280.02799999999996, "pct_cuda_time": 0.3131908137748171, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0014315862757715869, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 278.748, "pct_cuda_time": 0.31175922749904555, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 497.792, "cuda_time_us": 55.712, "pct_cuda_time": 0.06230979265295833, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.712, "pct_cuda_time": 0.06230979265295833, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.079, "cuda_time_us": 100.478, "pct_cuda_time": 0.11237728579451367, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.295, "pct_cuda_time": 0.026053751792264934, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.615, "pct_cuda_time": 0.08456984081442855, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.001753693187820194, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 208.629, "cuda_time_us": 197.534, "pct_cuda_time": 0.22092731515489425, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.798, "pct_cuda_time": 0.2201041530463256, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.637, "cuda_time_us": 44.383, "pct_cuda_time": 0.04963913568560184, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.383, "pct_cuda_time": 0.04963913568560184, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 496.489, "cuda_time_us": 2108.325, "pct_cuda_time": 2.3580071366141646, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.491, "cuda_time_us": 1333.743, "pct_cuda_time": 1.4916938860987683, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001968431129185932, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1331.983, "pct_cuda_time": 1.4897254549695826, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.057, "cuda_time_us": 177.918, "pct_cuda_time": 0.1989882554786947, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 177.918, "pct_cuda_time": 0.1989882554786947, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.287, "cuda_time_us": 596.664, "pct_cuda_time": 0.6673249950367017, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.376, "pct_cuda_time": 0.0015389552464544559, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.288, "pct_cuda_time": 0.6657860397902472, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2410.653, "cuda_time_us": 2825.947, "pct_cuda_time": 3.1606147978577255, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.374, "cuda_time_us": 44.288, "pct_cuda_time": 0.0495328851416969, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.288, "pct_cuda_time": 0.0495328851416969, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1699.167, "cuda_time_us": 636.279, "pct_cuda_time": 0.7116314718450544, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.883, "cuda_time_us": 281.564, "pct_cuda_time": 0.3149087173057431, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0018968518153973527, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.868, "pct_cuda_time": 0.3130118654903457, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.811, "cuda_time_us": 55.871, "pct_cuda_time": 0.062487622510651826, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.871, "pct_cuda_time": 0.062487622510651826, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 724.301, "cuda_time_us": 100.76700000000001, "pct_cuda_time": 0.11270051113334024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.2, "pct_cuda_time": 0.02594750124836001, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.903, "pct_cuda_time": 0.08489194772647717, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0018610621585030628, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.235, "cuda_time_us": 198.07700000000003, "pct_cuda_time": 0.22153462089531928, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 197.342, "pct_cuda_time": 0.22071257721352852, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.162, "cuda_time_us": 43.456, "pct_cuda_time": 0.04860235406244538, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.456, "pct_cuda_time": 0.04860235406244538, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.977, "cuda_time_us": 2101.924, "pct_cuda_time": 2.350848086808529, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.735, "cuda_time_us": 1326.6699999999998, "pct_cuda_time": 1.4837832534983524, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0014673759326658768, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1325.358, "pct_cuda_time": 1.4823158775656866, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.764, "cuda_time_us": 179.037, "pct_cuda_time": 0.2002397750432169, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 179.037, "pct_cuda_time": 0.2002397750432169, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.007, "cuda_time_us": 596.217, "pct_cuda_time": 0.6668250582669596, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.481, "pct_cuda_time": 0.6660018961583909, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2482.021, "cuda_time_us": 2825.532, "pct_cuda_time": 3.1601506507448778, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.096, "cuda_time_us": 44.159, "pct_cuda_time": 0.0493886080873418, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.159, "pct_cuda_time": 0.0493886080873418, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1785.037, "cuda_time_us": 637.113, "pct_cuda_time": 0.7125642397778618, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.292, "cuda_time_us": 281.82, "pct_cuda_time": 0.31519503456089737, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 280.572, "pct_cuda_time": 0.3137992379420201, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.627, "cuda_time_us": 56.384, "pct_cuda_time": 0.0630613754477384, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 56.384, "pct_cuda_time": 0.0630613754477384, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 769.097, "cuda_time_us": 100.60700000000001, "pct_cuda_time": 0.11252156284886879, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.328, "pct_cuda_time": 0.026090659875937172, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.775, "pct_cuda_time": 0.0847487890989, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0016821138740316145, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.844, "cuda_time_us": 198.30200000000002, "pct_cuda_time": 0.22178626692035724, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.0020400104429745118, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.478, "pct_cuda_time": 0.2197462564773827, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.174, "cuda_time_us": 43.808, "pct_cuda_time": 0.04899604028828256, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.808, "pct_cuda_time": 0.04899604028828256, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.931, "cuda_time_us": 2100.452, "pct_cuda_time": 2.3492017625913917, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.848, "cuda_time_us": 1323.919, "pct_cuda_time": 1.4807064614322216, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1323.183, "pct_cuda_time": 1.4798832993236528, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.649, "cuda_time_us": 178.845, "pct_cuda_time": 0.20002503710185113, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.845, "pct_cuda_time": 0.20002503710185113, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.309, "cuda_time_us": 597.688, "pct_cuda_time": 0.6684702640573189, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0018968518153973527, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.992, "pct_cuda_time": 0.6665734122419216, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2324.44, "cuda_time_us": 2825.24, "pct_cuda_time": 3.159824070125717, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.237, "cuda_time_us": 44.479, "pct_cuda_time": 0.0497465046562847, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.479, "pct_cuda_time": 0.0497465046562847, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1658.176, "cuda_time_us": 635.127, "pct_cuda_time": 0.7103430441968599, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.52, "cuda_time_us": 280.18899999999996, "pct_cuda_time": 0.3133708804860665, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.453, "pct_cuda_time": 0.31254771837749784, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.28, "cuda_time_us": 56.031, "pct_cuda_time": 0.06266657079512326, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 56.031, "pct_cuda_time": 0.06266657079512326, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.94, "cuda_time_us": 99.743, "pct_cuda_time": 0.11155524211272295, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.879, "pct_cuda_time": 0.025588486252639174, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.52, "pct_cuda_time": 0.08446359027052362, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0015031655895601662, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 180.537, "cuda_time_us": 199.164, "pct_cuda_time": 0.22275035080294714, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.279, "pct_cuda_time": 0.0014304678489936403, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 197.885, "pct_cuda_time": 0.22131988295395347, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.893, "cuda_time_us": 43.295, "pct_cuda_time": 0.04842228735119598, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.295, "pct_cuda_time": 0.04842228735119598, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.056, "cuda_time_us": 2102.339, "pct_cuda_time": 2.3513122339213766, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.967, "cuda_time_us": 1326.5430000000001, "pct_cuda_time": 1.4836412132975534, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1325.807, "pct_cuda_time": 1.4828180511889846, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.885, "cuda_time_us": 178.493, "pct_cuda_time": 0.19963135087601394, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.493, "pct_cuda_time": 0.19963135087601394, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.8, "cuda_time_us": 597.303, "pct_cuda_time": 0.6680396697478095, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0008220436817907158, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 596.568, "pct_cuda_time": 0.6672176260660188, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2442.352, "cuda_time_us": 2824.8599999999997, "pct_cuda_time": 3.159399067950097, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.766, "cuda_time_us": 44.288, "pct_cuda_time": 0.0495328851416969, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.288, "pct_cuda_time": 0.0495328851416969, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1728.921, "cuda_time_us": 633.7839999999999, "pct_cuda_time": 0.7088409970340775, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.84, "cuda_time_us": 281.24399999999997, "pct_cuda_time": 0.31455082073680013, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.996, "pct_cuda_time": 0.31315502411792284, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 476.427, "cuda_time_us": 55.455, "pct_cuda_time": 0.06202235697102606, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.455, "pct_cuda_time": 0.06202235697102606, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 742.747, "cuda_time_us": 100.67099999999999, "pct_cuda_time": 0.11259314216265737, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.392, "pct_cuda_time": 0.026162239189725753, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.551, "pct_cuda_time": 0.08449826150063998, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.463, "cuda_time_us": 196.414, "pct_cuda_time": 0.21967467716359412, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 195.678, "pct_cuda_time": 0.21885151505502543, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.118, "cuda_time_us": 43.936, "pct_cuda_time": 0.04913919891585972, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.936, "pct_cuda_time": 0.04913919891585972, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.515, "cuda_time_us": 2102.852, "pct_cuda_time": 2.351885986858463, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.292, "cuda_time_us": 1327.1180000000002, "pct_cuda_time": 1.4842843086948727, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.0019326414722916423, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1325.39, "pct_cuda_time": 1.482351667222581, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.191, "cuda_time_us": 178.59, "pct_cuda_time": 0.19973983827347475, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.59, "pct_cuda_time": 0.19973983827347475, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.401, "cuda_time_us": 597.144, "pct_cuda_time": 0.667861839890116, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.896, "pct_cuda_time": 0.6664660432712387, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2495.07, "cuda_time_us": 2829.8189999999995, "pct_cuda_time": 3.1649453463419337, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.378, "cuda_time_us": 44.607, "pct_cuda_time": 0.04988966328386186, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.607, "pct_cuda_time": 0.04988966328386186, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1814.501, "cuda_time_us": 636.281, "pct_cuda_time": 0.7116337086986102, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.296, "cuda_time_us": 281.053, "pct_cuda_time": 0.31433720122221237, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0017179035309259042, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.517, "pct_cuda_time": 0.3126192976912865, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.954, "cuda_time_us": 56.064, "pct_cuda_time": 0.0627034788787955, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 56.064, "pct_cuda_time": 0.0627034788787955, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 788.948, "cuda_time_us": 101.023, "pct_cuda_time": 0.11298682838849455, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.815, "pct_cuda_time": 0.025516906938850593, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 76.32, "pct_cuda_time": 0.08535833169288086, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.888, "pct_cuda_time": 0.0021115897567630906, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 224.263, "cuda_time_us": 198.141, "pct_cuda_time": 0.22160620020910782, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.0015031655895601662, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.797, "pct_cuda_time": 0.22010303461954764, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.053, "cuda_time_us": 43.231, "pct_cuda_time": 0.0483507080374074, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.231, "pct_cuda_time": 0.0483507080374074, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.626, "cuda_time_us": 2105.7, "pct_cuda_time": 2.3550712663220548, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.091, "cuda_time_us": 1330.6709999999998, "pct_cuda_time": 1.4882580790369166, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0014673759326658768, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1329.359, "pct_cuda_time": 1.4867907031042507, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.812, "cuda_time_us": 178.717, "pct_cuda_time": 0.199881878474274, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.717, "pct_cuda_time": 0.199881878474274, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.596, "cuda_time_us": 596.312, "pct_cuda_time": 0.6669313088108645, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.576, "pct_cuda_time": 0.6661081467022959, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2449.707, "cuda_time_us": 2828.923, "pct_cuda_time": 3.163943235948894, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.169, "cuda_time_us": 45.184, "pct_cuda_time": 0.05053499553473701, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.184, "pct_cuda_time": 0.05053499553473701, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1746.646, "cuda_time_us": 635.6709999999999, "pct_cuda_time": 0.7109514683640628, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.559, "cuda_time_us": 281.948, "pct_cuda_time": 0.3153381931884745, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 280.7, "pct_cuda_time": 0.3139423965695972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.666, "cuda_time_us": 55.359, "pct_cuda_time": 0.061914988000343195, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.359, "pct_cuda_time": 0.061914988000343195, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.29, "cuda_time_us": 100.223, "pct_cuda_time": 0.11209208696613732, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.944, "pct_cuda_time": 0.025661183993205695, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.967, "pct_cuda_time": 0.08496352704026573, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0014673759326658768, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.607, "cuda_time_us": 198.141, "pct_cuda_time": 0.22160620020910782, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.76, "pct_cuda_time": 0.001968431129185932, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.381, "pct_cuda_time": 0.21963776907992189, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.294, "cuda_time_us": 43.519, "pct_cuda_time": 0.04867281494945601, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.519, "pct_cuda_time": 0.04867281494945601, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 474.099, "cuda_time_us": 2104.549, "pct_cuda_time": 2.3537839571006387, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.585, "cuda_time_us": 1329.423, "pct_cuda_time": 1.4868622824180393, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1328.687, "pct_cuda_time": 1.4860391203094707, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.576, "cuda_time_us": 178.43, "pct_cuda_time": 0.19956088998900334, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.43, "pct_cuda_time": 0.19956088998900334, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.828, "cuda_time_us": 596.6959999999999, "pct_cuda_time": 0.6673607846935958, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.0018252725016087731, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.064, "pct_cuda_time": 0.6655355121919871, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2376.064, "cuda_time_us": 2822.17, "pct_cuda_time": 3.1563904999174213, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.47, "cuda_time_us": 45.567, "pct_cuda_time": 0.05096335299069055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 45.567, "pct_cuda_time": 0.05096335299069055, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1703.922, "cuda_time_us": 634.135, "pct_cuda_time": 0.7092335648331369, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.108, "cuda_time_us": 279.74, "pct_cuda_time": 0.3128687068627685, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.004, "pct_cuda_time": 0.3120455447541999, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 506.777, "cuda_time_us": 55.615, "pct_cuda_time": 0.062201305255497503, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.615, "pct_cuda_time": 0.062201305255497503, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.698, "cuda_time_us": 100.286, "pct_cuda_time": 0.11216254785314794, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.847, "pct_cuda_time": 0.02555269659574488, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.903, "pct_cuda_time": 0.08489194772647717, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0017179035309259042, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.129, "cuda_time_us": 198.494, "pct_cuda_time": 0.22200100486172292, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.28, "pct_cuda_time": 0.0014315862757715869, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 197.214, "pct_cuda_time": 0.2205694185859514, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.973, "cuda_time_us": 43.584, "pct_cuda_time": 0.04874551269002254, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.584, "pct_cuda_time": 0.04874551269002254, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.079, "cuda_time_us": 2098.884, "pct_cuda_time": 2.3474480694035713, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.345, "cuda_time_us": 1324.367, "pct_cuda_time": 1.4812075166287415, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0008589517654629521, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1323.599, "pct_cuda_time": 1.4803485648632786, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.498, "cuda_time_us": 178.525, "pct_cuda_time": 0.19966714053290827, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.525, "pct_cuda_time": 0.19966714053290827, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.721, "cuda_time_us": 595.992, "pct_cuda_time": 0.6665734122419216, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.256, "pct_cuda_time": 0.6657502501333529, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2316.617, "cuda_time_us": 2823.6730000000002, "pct_cuda_time": 3.1580714953646756, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.298, "cuda_time_us": 44.512, "pct_cuda_time": 0.049783412739956935, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.512, "pct_cuda_time": 0.049783412739956935, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1637.944, "cuda_time_us": 633.622, "pct_cuda_time": 0.7086598118960503, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.919, "cuda_time_us": 280.827, "pct_cuda_time": 0.31408443677039644, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.215, "pct_cuda_time": 0.0013588885352050612, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 279.612, "pct_cuda_time": 0.3127255482351914, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 451.367, "cuda_time_us": 56.287, "pct_cuda_time": 0.06295288805027759, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 56.287, "pct_cuda_time": 0.06295288805027759, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.844, "cuda_time_us": 100.351, "pct_cuda_time": 0.11223524559371446, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.552, "pct_cuda_time": 0.0263411874741972, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.263, "pct_cuda_time": 0.08417615458859137, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0017179035309259042, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 183.129, "cuda_time_us": 196.15699999999998, "pct_cuda_time": 0.21938724148166186, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 195.421, "pct_cuda_time": 0.21856407937309316, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.802, "cuda_time_us": 43.647, "pct_cuda_time": 0.04881597357703316, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.647, "pct_cuda_time": 0.04881597357703316, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 448.312, "cuda_time_us": 2101.8920000000003, "pct_cuda_time": 2.3508122971516348, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.648, "cuda_time_us": 1326.9270000000001, "pct_cuda_time": 1.4840706891802848, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.92, "pct_cuda_time": 0.00214737941365738, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1325.007, "pct_cuda_time": 1.4819233097666276, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.697, "cuda_time_us": 178.941, "pct_cuda_time": 0.20013240607253402, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.941, "pct_cuda_time": 0.20013240607253402, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.204, "cuda_time_us": 596.024, "pct_cuda_time": 0.6666092018988159, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 595.288, "pct_cuda_time": 0.6657860397902472, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2373.753, "cuda_time_us": 2822.009, "pct_cuda_time": 3.156210433206172, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.618, "cuda_time_us": 44.319, "pct_cuda_time": 0.04956755637181325, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.319, "pct_cuda_time": 0.04956755637181325, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1686.816, "cuda_time_us": 635.927, "pct_cuda_time": 0.7112377856192171, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.396, "cuda_time_us": 281.884, "pct_cuda_time": 0.31526661387468596, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.0016821138740316145, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 280.38, "pct_cuda_time": 0.31358450000065435, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.525, "cuda_time_us": 55.904, "pct_cuda_time": 0.06252453059432407, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.904, "pct_cuda_time": 0.06252453059432407, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.302, "cuda_time_us": 100.286, "pct_cuda_time": 0.11216254785314794, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 22.879, "pct_cuda_time": 0.025588486252639174, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.871, "pct_cuda_time": 0.08485615806958287, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.536, "pct_cuda_time": 0.0017179035309259042, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.675, "cuda_time_us": 197.853, "pct_cuda_time": 0.2212840932970592, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0014673759326658768, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 196.541, "pct_cuda_time": 0.21981671736439334, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.34, "cuda_time_us": 43.552, "pct_cuda_time": 0.04870972303312825, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 43.552, "pct_cuda_time": 0.04870972303312825, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 464.846, "cuda_time_us": 2098.2110000000002, "pct_cuda_time": 2.3466953681820137, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.461, "cuda_time_us": 1323.9180000000001, "pct_cuda_time": 1.4807053430054438, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1323.182, "pct_cuda_time": 1.479882180896875, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.881, "cuda_time_us": 178.397, "pct_cuda_time": 0.19952398190533105, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.397, "pct_cuda_time": 0.19952398190533105, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.015, "cuda_time_us": 595.896, "pct_cuda_time": 0.6664660432712387, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.376, "pct_cuda_time": 0.0015389552464544559, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 594.52, "pct_cuda_time": 0.6649270880247842, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2325.598, "cuda_time_us": 2825.2110000000002, "pct_cuda_time": 3.1597916357491576, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.606, "cuda_time_us": 44.48, "pct_cuda_time": 0.04974762308306264, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.48, "pct_cuda_time": 0.04974762308306264, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1646.941, "cuda_time_us": 636.248, "pct_cuda_time": 0.7115968006149381, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.647, "cuda_time_us": 281.276, "pct_cuda_time": 0.31458661039369445, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 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": 280.028, "pct_cuda_time": 0.31319081377481717, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[4096, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 452.459, "cuda_time_us": 55.359, "pct_cuda_time": 0.061914988000343195, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 55.359, "pct_cuda_time": 0.061914988000343195, "trace": "_C::rotary_embedding(int64[4096], bfloat16[4096, 4096], bfloat16[4096, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 709.744, "cuda_time_us": 100.575, "pct_cuda_time": 0.1124857731919745, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 23.456, "pct_cuda_time": 0.026233818503514326, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[4096], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 75.423, "pct_cuda_time": 0.08435510287306282, "trace": "_vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.0018968518153973527, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], None, None, bfloat16[4096, 32, 128], int32[9], int32[9], None, None, None, 512, 512, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[4096, 32, 128], bfloat16[4096, 8, 128], bfloat16[4096, 8, 128], bfloat16[4096, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.465, "cuda_time_us": 199.038, "pct_cuda_time": 0.2226094290289259, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.857, "pct_cuda_time": 0.0020769185266467476, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 197.181, "pct_cuda_time": 0.22053251050227915, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[4096, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.845, "cuda_time_us": 44.223, "pct_cuda_time": 0.049460187401130375, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.223, "pct_cuda_time": 0.049460187401130375, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.363, "cuda_time_us": 2100.26, "pct_cuda_time": 2.348987024650026, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.271, "cuda_time_us": 1326.19, "pct_cuda_time": 1.4832464086449382, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.248, "pct_cuda_time": 0.0013957966188772972, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 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": 1324.942, "pct_cuda_time": 1.4818506120260608, "trace": "mm(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[4096, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[4096, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.638, "cuda_time_us": 178.59, "pct_cuda_time": 0.19973983827347475, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 178.59, "pct_cuda_time": 0.19973983827347475, "trace": "_C::silu_and_mul(bfloat16[4096, 14336], bfloat16[4096, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.546, "cuda_time_us": 595.48, "pct_cuda_time": 0.6660007777316129, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0008231621085686625, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x256x64_warpgroupsize2x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 594.744, "pct_cuda_time": 0.6651776156230443, "trace": "mm(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[4096, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[4096, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.191, "cuda_time_us": 44.447, "pct_cuda_time": 0.049710714999390414, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 44.447, "pct_cuda_time": 0.049710714999390414, "trace": "_C::fused_add_rms_norm(bfloat16[4096, 4096], bfloat16[4096, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 440.811, "cuda_time_us": 365.59499999999997, "pct_cuda_time": 0.40889123788336973, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 7.232, "pct_cuda_time": 0.008088462458109466, "trace": "index_select(bfloat16[4096, 4096], 0, int64[8])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.312, "pct_cuda_time": 0.0014673759326658768, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[8, 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": 357.051, "pct_cuda_time": 0.39933539949259445, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[8, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 12097.628, "cuda_time_us": 141.94899999999998, "pct_cuda_time": 0.15875956270273514, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 3.072, "pct_cuda_time": 0.0034358070618518083, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.0026842242670717252, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.002720013923966015, "trace": "copy_(int32[8], int32[8], True) <- _to_copy(int32[8], 3, 0, None, None, True, None) <- to(int32[8], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.002720013923966015, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.0027915932377545944, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.431, "pct_cuda_time": 0.0027188954971880683, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 2.527, "pct_cuda_time": 0.0028262644678709375, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 5.215, "pct_cuda_time": 0.00583259564699127, "trace": "copy_(float32[8, 128256], bfloat16[8, 128256], False) <- _to_copy(bfloat16[8, 128256], 6, None, None, None, False, None) <- to(bfloat16[8, 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": 6.848, "pct_cuda_time": 0.00765898657537799, "trace": "div_(float32[8, 128256], bfloat16[8, 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": 37.888, "pct_cuda_time": 0.042374953762838966, "trace": "_softmax(float32[8, 128256], -1, False) <- softmax(float32[8, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 30.783, "pct_cuda_time": 0.03442853150552872, "trace": "_log_softmax(float32[8, 128256], -1, False) <- log_softmax(float32[8, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.0023632357818010653, "trace": "copy_(int64[8], int32[8], False) <- _to_copy(int32[8], 4, None, None, None, False, None) <- to(int32[8], 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": 7.904, "pct_cuda_time": 0.008840045252889549, "trace": "index(float32[8, 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": 30.08, "pct_cuda_time": 0.03364227748063229, "trace": "argmax(float32[8, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 3.328, "pct_cuda_time": 0.0037221243170061256, "trace": "copy_(int64[8], int64[8], False) <- _to_copy(int64[8], 4, 0, None, None, False, None) <- to(int64[8], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 8 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6574.026, "pct_cuda_time": 93.36317421039266, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 7.296, "pct_cuda_time": 0.10361652342704836, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 7.296, "pct_cuda_time": 0.10361652342704836, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6563.69, "pct_cuda_time": 93.21638413553768, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 200.28800000000012, "pct_cuda_time": 2.844455351446912, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.224, "pct_cuda_time": 0.059988513563027986, "invocations": 1 }, "children": [] }, { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 196.06400000000014, "pct_cuda_time": 2.7844668378838846, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 2059.557, "pct_cuda_time": 29.2494704139037, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 680.5010000000001, "pct_cuda_time": 9.664356881665274, "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": 680.5010000000001, "pct_cuda_time": 9.664356881665274, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 122.78499999999997, "pct_cuda_time": 1.7437712210786906, "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": 122.78499999999997, "pct_cuda_time": 1.7437712210786906, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 681.427, "pct_cuda_time": 9.677507772659439, "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": 86.59100000000001, "pct_cuda_time": 1.2297503262159462, "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": 552.758, "pct_cuda_time": 7.850173006645887, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 42.07800000000002, "pct_cuda_time": 0.5975844397976071, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 574.844, "pct_cuda_time": 8.163834538500298, "invocations": 32 }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cuda_time_us": 506.077, "pct_cuda_time": 7.187217561182887, "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": 68.767, "pct_cuda_time": 0.9766169773174114, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4303.845, "pct_cuda_time": 61.12245837018707, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2607.6149999999993, "pct_cuda_time": 37.03289483774981, "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": 2607.6149999999993, "pct_cuda_time": 37.03289483774981, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 291.76899999999995, "pct_cuda_time": 4.143652607426873, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 291.76899999999995, "pct_cuda_time": 4.143652607426873, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1404.4609999999998, "pct_cuda_time": 19.94591092501038, "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": 1404.4609999999998, "pct_cuda_time": 19.94591092501038, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.04, "pct_cuda_time": 0.04317355142793681, "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": 0.04317355142793681, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 347.899, "pct_cuda_time": 4.940801107969667, "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": 5.568, "pct_cuda_time": 0.07907576787853689, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.704, "pct_cuda_time": 0.009998085593837998, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 341.627, "pct_cuda_time": 4.851727254497292, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 119.42300000000002, "pct_cuda_time": 1.6960246816376638, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.4079999999999995, "pct_cuda_time": 0.07680347569811916, "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": 4.736, "pct_cuda_time": 0.06725984854036471, "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": 6.368, "pct_cuda_time": 0.09043722878062553, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 34.88, "pct_cuda_time": 0.4953596953310645, "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": 28.735, "pct_cuda_time": 0.40808947377689614, "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": 0.02590413085676209, "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": 7.36, "pct_cuda_time": 0.10452544029921544, "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": 27.328, "pct_cuda_time": 0.3881075044153477, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.784, "pct_cuda_time": 0.039537883939268444, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 91535.929, "cuda_time_us": 6574.026, "pct_cuda_time": 93.36317421039266, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 351.56, "cuda_time_us": 7.296, "pct_cuda_time": 0.10361652342704836, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 7.296, "pct_cuda_time": 0.10361652342704836, "trace": "index_select(bfloat16[128256, 4096], 0, int64[8]) <- embedding(bfloat16[128256, 4096], int64[8], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 5195.334, "cuda_time_us": 211.45300000000003, "pct_cuda_time": 3.0030187401616852, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 328.964, "cuda_time_us": 4.224, "pct_cuda_time": 0.059988513563027986, "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": 4.224, "pct_cuda_time": 0.059988513563027986, "trace": "_C::rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3777.478, "cuda_time_us": 69.6, "pct_cuda_time": 0.9884470984817111, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 572.152, "cuda_time_us": 26.24, "pct_cuda_time": 0.3726559175885072, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 26.24, "pct_cuda_time": 0.3726559175885072, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1033.479, "cuda_time_us": 3.616, "pct_cuda_time": 0.05135380327744062, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.05135380327744062, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1470.946, "cuda_time_us": 21.6, "pct_cuda_time": 0.30675944435639313, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.03635667488668363, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.76, "pct_cuda_time": 0.2522244320263677, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 379.478, "cuda_time_us": 18.144, "pct_cuda_time": 0.2576779332593702, "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": 16.0, "pct_cuda_time": 0.22722921804177268, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 139.171, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 779.51, "cuda_time_us": 134.55700000000002, "pct_cuda_time": 1.9109551182529259, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 313.312, "cuda_time_us": 81.151, "pct_cuda_time": 1.1524923920817434, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.151, "pct_cuda_time": 1.1524923920817434, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 197.759, "cuda_time_us": 9.087, "pct_cuda_time": 0.12905199402159928, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.087, "pct_cuda_time": 0.12905199402159928, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 182.177, "cuda_time_us": 44.319, "pct_cuda_time": 0.6294107321495828, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.319, "pct_cuda_time": 0.6294107321495828, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3167.551, "cuda_time_us": 205.791, "pct_cuda_time": 2.922608000627153, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 130.549, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2258.057, "cuda_time_us": 63.52, "pct_cuda_time": 0.9020999956258376, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 208.868, "cuda_time_us": 20.672, "pct_cuda_time": 0.2935801497099703, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.672, "pct_cuda_time": 0.2935801497099703, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 609.963, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1043.78, "cuda_time_us": 21.12, "pct_cuda_time": 0.29994256781513995, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.152, "pct_cuda_time": 0.24358972174078034, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 203.576, "cuda_time_us": 17.92, "pct_cuda_time": 0.25449672420678543, "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": 15.776, "pct_cuda_time": 0.22404800898918786, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.973, "cuda_time_us": 3.2, "pct_cuda_time": 0.04544584360835454, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04544584360835454, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 588.841, "cuda_time_us": 136.063, "pct_cuda_time": 1.9323430684011074, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 209.747, "cuda_time_us": 83.039, "pct_cuda_time": 1.1793054398106728, "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": 1.1793054398106728, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 110.824, "cuda_time_us": 9.184, "pct_cuda_time": 0.1304295711559775, "trace": "" }, "children": [ { "entry": { "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": 0.1304295711559775, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 194.392, "cuda_time_us": 43.84, "pct_cuda_time": 0.6226080574344572, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.84, "pct_cuda_time": 0.6226080574344572, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2874.333, "cuda_time_us": 204.862, "pct_cuda_time": 2.909414504154602, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.858, "cuda_time_us": 2.976, "pct_cuda_time": 0.04226463455576972, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::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": 0.04226463455576972, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2044.665, "cuda_time_us": 64.287, "pct_cuda_time": 0.9129927962657152, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 178.684, "cuda_time_us": 21.344, "pct_cuda_time": 0.3031237768677248, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.344, "pct_cuda_time": 0.3031237768677248, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.452, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 967.274, "cuda_time_us": 21.215, "pct_cuda_time": 0.30129174129726294, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.215, "pct_cuda_time": 0.2444844367868198, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.492, "cuda_time_us": 17.951999999999998, "pct_cuda_time": 0.2549511826428689, "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": 0.22450246742527138, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.634, "cuda_time_us": 3.233, "pct_cuda_time": 0.0459145038705657, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.233, "pct_cuda_time": 0.0459145038705657, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 530.579, "cuda_time_us": 134.36599999999999, "pct_cuda_time": 1.9082425694625518, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 187.753, "cuda_time_us": 81.887, "pct_cuda_time": 1.162944936111665, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.887, "pct_cuda_time": 1.162944936111665, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 106.083, "cuda_time_us": 9.056, "pct_cuda_time": 0.1286117374116433, "trace": "" }, "children": [ { "entry": { "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": 0.1286117374116433, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 172.375, "cuda_time_us": 43.423, "pct_cuda_time": 0.6166858959392435, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6166858959392435, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2974.362, "cuda_time_us": 203.967, "pct_cuda_time": 2.896703869770391, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.622, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2125.204, "cuda_time_us": 63.52, "pct_cuda_time": 0.9020999956258376, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.418, "cuda_time_us": 20.512, "pct_cuda_time": 0.29130785752955257, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29130785752955257, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 545.264, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1007.904, "cuda_time_us": 21.312, "pct_cuda_time": 0.30266931843164124, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.312, "pct_cuda_time": 0.24586201392119808, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.198, "cuda_time_us": 17.92, "pct_cuda_time": 0.25449672420678543, "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": 15.776, "pct_cuda_time": 0.22404800898918786, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.779, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 601.493, "cuda_time_us": 134.239, "pct_cuda_time": 1.9064389375443451, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 237.248, "cuda_time_us": 81.822, "pct_cuda_time": 1.1620218174133703, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.822, "pct_cuda_time": 1.1620218174133703, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 130.898, "cuda_time_us": 9.153, "pct_cuda_time": 0.1299893145460216, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.153, "pct_cuda_time": 0.1299893145460216, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.723, "cuda_time_us": 43.264, "pct_cuda_time": 0.6144278055849534, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.264, "pct_cuda_time": 0.6144278055849534, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2886.367, "cuda_time_us": 204.798, "pct_cuda_time": 2.9085055872824355, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.194, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2068.252, "cuda_time_us": 64.80000000000001, "pct_cuda_time": 0.9202783330691796, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 200.815, "cuda_time_us": 21.568, "pct_cuda_time": 0.3063049859203096, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.568, "pct_cuda_time": 0.3063049859203096, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 500.218, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 930.338, "cuda_time_us": 21.344, "pct_cuda_time": 0.3031237768677248, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.816, "pct_cuda_time": 0.03999234237535199, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.248, "pct_cuda_time": 0.24495309704903098, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 194.753, "cuda_time_us": 18.112000000000002, "pct_cuda_time": 0.2572234748232867, "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": 15.936, "pct_cuda_time": 0.2263203011696056, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 2.176, "pct_cuda_time": 0.030903173653681084, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.602, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 551.003, "cuda_time_us": 133.822, "pct_cuda_time": 1.9005167760491315, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 198.094, "cuda_time_us": 81.695, "pct_cuda_time": 1.1602181854951636, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.695, "pct_cuda_time": 1.1602181854951636, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 121.914, "cuda_time_us": 9.024, "pct_cuda_time": 0.12815727897555979, "trace": "" }, "children": [ { "entry": { "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": 0.12815727897555979, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.334, "cuda_time_us": 43.103, "pct_cuda_time": 0.612141311578408, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.103, "pct_cuda_time": 0.612141311578408, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2964.449, "cuda_time_us": 202.429, "pct_cuda_time": 2.874861461186125, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.014, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2145.743, "cuda_time_us": 63.136, "pct_cuda_time": 0.8966464943928351, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.289, "cuda_time_us": 20.575, "pct_cuda_time": 0.29220257257559207, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.575, "pct_cuda_time": 0.29220257257559207, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.259, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1077.136, "cuda_time_us": 20.96, "pct_cuda_time": 0.2976702756347222, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 16.992, "pct_cuda_time": 0.2413174295603626, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 216.045, "cuda_time_us": 17.793, "pct_cuda_time": 0.25269309228857884, "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": 15.649, "pct_cuda_time": 0.2222443770709813, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 112.939, "cuda_time_us": 3.168, "pct_cuda_time": 0.04499138517227099, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04499138517227099, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 524.889, "cuda_time_us": 133.117, "pct_cuda_time": 1.8905044886291658, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.517, "cuda_time_us": 80.511, "pct_cuda_time": 1.1434032233600724, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.511, "pct_cuda_time": 1.1434032233600724, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 124.508, "cuda_time_us": 8.991, "pct_cuda_time": 0.12768861871334863, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.991, "pct_cuda_time": 0.12768861871334863, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 171.474, "cuda_time_us": 43.615, "pct_cuda_time": 0.6194126465557448, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6194126465557448, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2786.166, "cuda_time_us": 206.557, "pct_cuda_time": 2.933486599440902, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.333, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1963.038, "cuda_time_us": 64.67099999999999, "pct_cuda_time": 0.9184462974987175, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.542, "cuda_time_us": 21.471, "pct_cuda_time": 0.30492740878593133, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.471, "pct_cuda_time": 0.30492740878593133, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 478.627, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 931.802, "cuda_time_us": 21.536, "pct_cuda_time": 0.30585052748422603, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.568, "pct_cuda_time": 0.24949768140986642, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 204.482, "cuda_time_us": 17.856, "pct_cuda_time": 0.25358780733461833, "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": 0.22313909211702077, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 99.321, "cuda_time_us": 3.264, "pct_cuda_time": 0.046354760480521624, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046354760480521624, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 550.042, "cuda_time_us": 135.518, "pct_cuda_time": 1.9246030731615595, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 204.189, "cuda_time_us": 82.271, "pct_cuda_time": 1.1683984373446676, "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": 82.271, "pct_cuda_time": 1.1683984373446676, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.437, "cuda_time_us": 9.056, "pct_cuda_time": 0.1286117374116433, "trace": "" }, "children": [ { "entry": { "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": 0.1286117374116433, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 176.826, "cuda_time_us": 44.191, "pct_cuda_time": 0.6275928984052486, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.191, "pct_cuda_time": 0.6275928984052486, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2797.779, "cuda_time_us": 203.325, "pct_cuda_time": 2.8875862973964646, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.662, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1959.696, "cuda_time_us": 63.199000000000005, "pct_cuda_time": 0.8975412094388745, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.395, "cuda_time_us": 20.512, "pct_cuda_time": 0.29130785752955257, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29130785752955257, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.518, "cuda_time_us": 3.936, "pct_cuda_time": 0.055898387638276086, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.936, "pct_cuda_time": 0.055898387638276086, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 909.522, "cuda_time_us": 21.023, "pct_cuda_time": 0.2985649906807617, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 16.991, "pct_cuda_time": 0.241303227734235, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.891, "cuda_time_us": 17.728, "pct_cuda_time": 0.2517699735902842, "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": 0.2213212583726866, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.084, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 548.945, "cuda_time_us": 133.95, "pct_cuda_time": 1.9023346097934657, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 177.53, "cuda_time_us": 81.151, "pct_cuda_time": 1.1524923920817434, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.151, "pct_cuda_time": 1.1524923920817434, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.619, "cuda_time_us": 8.96, "pct_cuda_time": 0.12724836210339271, "trace": "" }, "children": [ { "entry": { "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": 0.12724836210339271, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 203.089, "cuda_time_us": 43.839, "pct_cuda_time": 0.6225938556083295, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.6225938556083295, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2673.201, "cuda_time_us": 203.452, "pct_cuda_time": 2.889389929314671, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.763, "cuda_time_us": 3.071, "pct_cuda_time": 0.04361380803789275, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.071, "pct_cuda_time": 0.04361380803789275, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1909.394, "cuda_time_us": 64.447, "pct_cuda_time": 0.9152650884461329, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 182.143, "cuda_time_us": 21.312, "pct_cuda_time": 0.30266931843164124, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.312, "pct_cuda_time": 0.30266931843164124, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.074, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 864.261, "cuda_time_us": 21.344, "pct_cuda_time": 0.3031237768677248, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.376, "pct_cuda_time": 0.24677093079336515, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.18, "cuda_time_us": 17.983, "pct_cuda_time": 0.2553914392528249, "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": 15.839, "pct_cuda_time": 0.22494272403522736, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.656, "cuda_time_us": 3.136, "pct_cuda_time": 0.04453692673618745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04453692673618745, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 523.085, "cuda_time_us": 132.798, "pct_cuda_time": 1.8859741060944581, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 182.34, "cuda_time_us": 80.191, "pct_cuda_time": 1.138858638999237, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.191, "pct_cuda_time": 1.138858638999237, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 124.987, "cuda_time_us": 9.024, "pct_cuda_time": 0.12815727897555979, "trace": "" }, "children": [ { "entry": { "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": 0.12815727897555979, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.092, "cuda_time_us": 43.583, "pct_cuda_time": 0.6189581881196612, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.583, "pct_cuda_time": 0.6189581881196612, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2838.984, "cuda_time_us": 205.115, "pct_cuda_time": 2.9130075661648878, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.262, "cuda_time_us": 3.135, "pct_cuda_time": 0.04452272491005983, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04452272491005983, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2024.9, "cuda_time_us": 63.454, "pct_cuda_time": 0.9011626751014152, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 267.045, "cuda_time_us": 20.735, "pct_cuda_time": 0.29447486475600976, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.735, "pct_cuda_time": 0.29447486475600976, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.933, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 914.876, "cuda_time_us": 21.183, "pct_cuda_time": 0.30083728286117944, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.087, "pct_cuda_time": 0.24266660304248563, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019996171187675996, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.989, "cuda_time_us": 17.759999999999998, "pct_cuda_time": 0.2522244320263677, "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": 0.22177571680877012, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 98.032, "cuda_time_us": 3.264, "pct_cuda_time": 0.046354760480521624, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046354760480521624, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 550.17, "cuda_time_us": 135.262, "pct_cuda_time": 1.920967405672891, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 201.279, "cuda_time_us": 81.567, "pct_cuda_time": 1.1584003517508294, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.567, "pct_cuda_time": 1.1584003517508294, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.625, "cuda_time_us": 9.344, "pct_cuda_time": 0.13270186333639522, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.344, "pct_cuda_time": 0.13270186333639522, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 174.708, "cuda_time_us": 44.351, "pct_cuda_time": 0.6298651905856663, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.351, "pct_cuda_time": 0.6298651905856663, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2818.33, "cuda_time_us": 205.438, "pct_cuda_time": 2.917594756004106, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.153, "cuda_time_us": 3.136, "pct_cuda_time": 0.04453692673618745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04453692673618745, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1956.285, "cuda_time_us": 64.896, "pct_cuda_time": 0.9216417083774301, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 208.023, "cuda_time_us": 21.472, "pct_cuda_time": 0.30494161061205893, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.472, "pct_cuda_time": 0.30494161061205893, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 500.091, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 875.036, "cuda_time_us": 21.28, "pct_cuda_time": 0.3022148599955577, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.12, "pct_cuda_time": 0.2431352633046968, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020905088059843084, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.947, "cuda_time_us": 18.336, "pct_cuda_time": 0.2604046838758715, "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": 16.192, "pct_cuda_time": 0.22995596865827395, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 98.057, "cuda_time_us": 3.136, "pct_cuda_time": 0.04453692673618745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04453692673618745, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 603.343, "cuda_time_us": 134.27, "pct_cuda_time": 1.906879194154301, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 201.361, "cuda_time_us": 81.311, "pct_cuda_time": 1.1547646842621613, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.311, "pct_cuda_time": 1.1547646842621613, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 134.241, "cuda_time_us": 9.088, "pct_cuda_time": 0.12906619584772688, "trace": "" }, "children": [ { "entry": { "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": 0.12906619584772688, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.539, "cuda_time_us": 43.871, "pct_cuda_time": 0.6230483140444132, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.871, "pct_cuda_time": 0.6230483140444132, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2669.996, "cuda_time_us": 205.18100000000004, "pct_cuda_time": 2.9139448866893107, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.576, "cuda_time_us": 3.328, "pct_cuda_time": 0.047263677352688716, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.328, "pct_cuda_time": 0.047263677352688716, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1957.885, "cuda_time_us": 63.647000000000006, "pct_cuda_time": 0.9039036275440442, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.93, "cuda_time_us": 20.479, "pct_cuda_time": 0.2908391972673414, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.479, "pct_cuda_time": 0.2908391972673414, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 514.543, "cuda_time_us": 3.936, "pct_cuda_time": 0.055898387638276086, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.936, "pct_cuda_time": 0.055898387638276086, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 840.653, "cuda_time_us": 21.44, "pct_cuda_time": 0.3044871521759754, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.44, "pct_cuda_time": 0.24767984766553225, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 162.699, "cuda_time_us": 17.792, "pct_cuda_time": 0.25267889046245123, "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": 0.2222301752448537, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.598, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 474.474, "cuda_time_us": 135.13400000000001, "pct_cuda_time": 1.9191495719285572, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.247, "cuda_time_us": 81.471, "pct_cuda_time": 1.157036976442579, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.471, "pct_cuda_time": 1.157036976442579, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.841, "cuda_time_us": 8.991, "pct_cuda_time": 0.12768861871334863, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.991, "pct_cuda_time": 0.12768861871334863, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.3, "cuda_time_us": 44.672, "pct_cuda_time": 0.6344239767726293, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.672, "pct_cuda_time": 0.6344239767726293, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2852.414, "cuda_time_us": 204.094, "pct_cuda_time": 2.8985075016885973, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.186, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2082.294, "cuda_time_us": 64.44800000000001, "pct_cuda_time": 0.9152792902722605, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.24, "cuda_time_us": 21.408, "pct_cuda_time": 0.3040326937398919, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3040326937398919, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.908, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1091.954, "cuda_time_us": 21.344, "pct_cuda_time": 0.3031237768677248, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.216, "pct_cuda_time": 0.24449863861294743, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.019996171187675996, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.264, "cuda_time_us": 17.92, "pct_cuda_time": 0.25449672420678543, "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": 15.776, "pct_cuda_time": 0.22404800898918786, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.506, "cuda_time_us": 3.168, "pct_cuda_time": 0.04499138517227099, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04499138517227099, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 511.433, "cuda_time_us": 133.47, "pct_cuda_time": 1.8955177332522124, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 172.433, "cuda_time_us": 81.311, "pct_cuda_time": 1.1547646842621613, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.311, "pct_cuda_time": 1.1547646842621613, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 111.277, "cuda_time_us": 8.895, "pct_cuda_time": 0.126325243405098, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.895, "pct_cuda_time": 0.126325243405098, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.326, "cuda_time_us": 43.264, "pct_cuda_time": 0.6144278055849534, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.264, "pct_cuda_time": 0.6144278055849534, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2591.94, "cuda_time_us": 205.08700000000002, "pct_cuda_time": 2.9126099150333147, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.101, "cuda_time_us": 3.2, "pct_cuda_time": 0.04544584360835454, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04544584360835454, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1854.024, "cuda_time_us": 63.486999999999995, "pct_cuda_time": 0.9016313353636264, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.938, "cuda_time_us": 20.415, "pct_cuda_time": 0.2899302803951743, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2899302803951743, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.877, "cuda_time_us": 4.032, "pct_cuda_time": 0.05726176294652672, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.032, "pct_cuda_time": 0.05726176294652672, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 821.217, "cuda_time_us": 21.088, "pct_cuda_time": 0.29948810937905646, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.12, "pct_cuda_time": 0.2431352633046968, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 159.315, "cuda_time_us": 17.951999999999998, "pct_cuda_time": 0.2549511826428689, "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": 15.776, "pct_cuda_time": 0.22404800898918786, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 2.176, "pct_cuda_time": 0.030903173653681084, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.298, "cuda_time_us": 3.04, "pct_cuda_time": 0.04317355142793681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04317355142793681, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 480.727, "cuda_time_us": 135.36, "pct_cuda_time": 1.922359184633397, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.83, "cuda_time_us": 81.28, "pct_cuda_time": 1.1543244276522053, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.28, "pct_cuda_time": 1.1543244276522053, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.309, "cuda_time_us": 9.312, "pct_cuda_time": 0.1322474049003117, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.312, "pct_cuda_time": 0.1322474049003117, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.196, "cuda_time_us": 44.768, "pct_cuda_time": 0.63578735208088, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.768, "pct_cuda_time": 0.63578735208088, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2622.082, "cuda_time_us": 204.318, "pct_cuda_time": 2.901688710741182, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.432, "cuda_time_us": 3.136, "pct_cuda_time": 0.04453692673618745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04453692673618745, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1854.468, "cuda_time_us": 64.28900000000002, "pct_cuda_time": 0.9130211999179705, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 172.712, "cuda_time_us": 21.439, "pct_cuda_time": 0.3044729503498478, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.439, "pct_cuda_time": 0.3044729503498478, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.188, "cuda_time_us": 3.841, "pct_cuda_time": 0.05454921415615306, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.841, "pct_cuda_time": 0.05454921415615306, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 825.964, "cuda_time_us": 21.153000000000002, "pct_cuda_time": 0.30041122807735116, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.721, "pct_cuda_time": 0.03864316889322897, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.152, "pct_cuda_time": 0.24358972174078034, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 166.584, "cuda_time_us": 17.856, "pct_cuda_time": 0.25358780733461833, "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": 0.2222301752448537, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 2.208, "pct_cuda_time": 0.03135763208976463, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.037, "cuda_time_us": 3.136, "pct_cuda_time": 0.04453692673618745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04453692673618745, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 513.038, "cuda_time_us": 133.757, "pct_cuda_time": 1.8995936573508367, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 177.646, "cuda_time_us": 80.83, "pct_cuda_time": 1.1479336058947802, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.83, "pct_cuda_time": 1.1479336058947802, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.633, "cuda_time_us": 9.312, "pct_cuda_time": 0.1322474049003117, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.312, "pct_cuda_time": 0.1322474049003117, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.728, "cuda_time_us": 43.615, "pct_cuda_time": 0.6194126465557448, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6194126465557448, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2682.979, "cuda_time_us": 205.21099999999998, "pct_cuda_time": 2.914370941473138, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.698, "cuda_time_us": 3.136, "pct_cuda_time": 0.04453692673618745, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04453692673618745, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1889.695, "cuda_time_us": 64.957, "pct_cuda_time": 0.9225080197712142, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.608, "cuda_time_us": 21.44, "pct_cuda_time": 0.3044871521759754, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.44, "pct_cuda_time": 0.3044871521759754, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 496.596, "cuda_time_us": 3.904, "pct_cuda_time": 0.05544392920219253, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.904, "pct_cuda_time": 0.05544392920219253, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 870.979, "cuda_time_us": 21.503, "pct_cuda_time": 0.3053818672220149, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.848, "pct_cuda_time": 0.040446800811435535, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.184, "pct_cuda_time": 0.24404418017686386, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.02089088623371548, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 178.277, "cuda_time_us": 18.11, "pct_cuda_time": 0.2571950711710314, "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": 15.999, "pct_cuda_time": 0.22721501621564508, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.029980054955386387, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 109.852, "cuda_time_us": 3.168, "pct_cuda_time": 0.04499138517227099, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04499138517227099, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 521.326, "cuda_time_us": 133.95, "pct_cuda_time": 1.9023346097934657, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 181.071, "cuda_time_us": 81.823, "pct_cuda_time": 1.1620360192394978, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.823, "pct_cuda_time": 1.1620360192394978, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.317, "cuda_time_us": 9.152, "pct_cuda_time": 0.12997511271989395, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.152, "pct_cuda_time": 0.12997511271989395, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 163.457, "cuda_time_us": 42.975, "pct_cuda_time": 0.6103234778340738, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.975, "pct_cuda_time": 0.6103234778340738, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2756.019, "cuda_time_us": 205.372, "pct_cuda_time": 2.9166574354796837, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.35, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1947.113, "cuda_time_us": 64.89500000000001, "pct_cuda_time": 0.9216275065513025, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.455, "cuda_time_us": 21.631, "pct_cuda_time": 0.307199700966349, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.631, "pct_cuda_time": 0.307199700966349, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 545.861, "cuda_time_us": 4.0, "pct_cuda_time": 0.05680730451044317, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.0, "pct_cuda_time": 0.05680730451044317, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 891.973, "cuda_time_us": 21.248, "pct_cuda_time": 0.30176040155947415, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.248, "pct_cuda_time": 0.24495309704903098, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 175.275, "cuda_time_us": 18.016, "pct_cuda_time": 0.255860099515036, "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": 15.872, "pct_cuda_time": 0.2254113842974385, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 94.351, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 533.27, "cuda_time_us": 134.365, "pct_cuda_time": 1.9082283676364244, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 180.428, "cuda_time_us": 82.175, "pct_cuda_time": 1.167035062036417, "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": 82.175, "pct_cuda_time": 1.167035062036417, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 112.758, "cuda_time_us": 9.087, "pct_cuda_time": 0.12905199402159928, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.087, "pct_cuda_time": 0.12905199402159928, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 166.788, "cuda_time_us": 43.103, "pct_cuda_time": 0.612141311578408, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.103, "pct_cuda_time": 0.612141311578408, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2589.426, "cuda_time_us": 207.67600000000002, "pct_cuda_time": 2.9493784428776992, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.197, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1838.156, "cuda_time_us": 65.18100000000001, "pct_cuda_time": 0.9256892288237992, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.008, "cuda_time_us": 21.536, "pct_cuda_time": 0.30585052748422603, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.536, "pct_cuda_time": 0.30585052748422603, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.518, "cuda_time_us": 3.968, "pct_cuda_time": 0.05635284607435963, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.968, "pct_cuda_time": 0.05635284607435963, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 845.688, "cuda_time_us": 21.726, "pct_cuda_time": 0.30854887444847207, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.695, "pct_cuda_time": 0.251301313328073, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.01861859405329775, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.848, "cuda_time_us": 17.951, "pct_cuda_time": 0.2549369808167413, "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": 0.22450246742527138, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030434513391469922, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.477, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 501.35, "cuda_time_us": 136.351, "pct_cuda_time": 1.936433194325859, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.795, "cuda_time_us": 82.495, "pct_cuda_time": 1.1715796463972523, "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": 82.495, "pct_cuda_time": 1.1715796463972523, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 106.868, "cuda_time_us": 9.344, "pct_cuda_time": 0.13270186333639522, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.344, "pct_cuda_time": 0.13270186333639522, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.777, "cuda_time_us": 44.512, "pct_cuda_time": 0.6321516845922116, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.512, "pct_cuda_time": 0.6321516845922116, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2589.671, "cuda_time_us": 205.532, "pct_cuda_time": 2.9189297276601014, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.47, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1859.438, "cuda_time_us": 64.73400000000001, "pct_cuda_time": 0.9193410125447572, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.975, "cuda_time_us": 21.567, "pct_cuda_time": 0.306290784094182, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.567, "pct_cuda_time": 0.306290784094182, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 508.724, "cuda_time_us": 3.871, "pct_cuda_time": 0.05497526893998137, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.871, "pct_cuda_time": 0.05497526893998137, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 833.185, "cuda_time_us": 21.024, "pct_cuda_time": 0.2985791925068893, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.056, "pct_cuda_time": 0.2422263464325297, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.196, "cuda_time_us": 18.272, "pct_cuda_time": 0.25949576700370436, "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": 0.22904705178610688, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.443, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 486.259, "cuda_time_us": 134.686, "pct_cuda_time": 1.9127871538233872, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.547, "cuda_time_us": 81.631, "pct_cuda_time": 1.1593092686229967, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.631, "pct_cuda_time": 1.1593092686229967, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.273, "cuda_time_us": 8.928, "pct_cuda_time": 0.12679390366730917, "trace": "" }, "children": [ { "entry": { "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": 0.12679390366730917, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.281, "cuda_time_us": 44.127, "pct_cuda_time": 0.6266839815330815, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6266839815330815, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2898.417, "cuda_time_us": 205.598, "pct_cuda_time": 2.919867048184524, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.13, "cuda_time_us": 3.168, "pct_cuda_time": 0.04499138517227099, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04499138517227099, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1884.935, "cuda_time_us": 64.382, "pct_cuda_time": 0.9143419697478382, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.216, "cuda_time_us": 21.151, "pct_cuda_time": 0.3003828244250959, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.151, "pct_cuda_time": 0.3003828244250959, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.647, "cuda_time_us": 3.968, "pct_cuda_time": 0.05635284607435963, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.968, "pct_cuda_time": 0.05635284607435963, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 944.022, "cuda_time_us": 21.247, "pct_cuda_time": 0.30174619973334654, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.247, "pct_cuda_time": 0.24493889522290332, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.606, "cuda_time_us": 18.016, "pct_cuda_time": 0.255860099515036, "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": 15.872, "pct_cuda_time": 0.2254113842974385, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 94.896, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 772.387, "cuda_time_us": 134.976, "pct_cuda_time": 1.9169056834003944, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 174.963, "cuda_time_us": 81.376, "pct_cuda_time": 1.155687802960456, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.376, "pct_cuda_time": 1.155687802960456, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.1, "cuda_time_us": 9.376, "pct_cuda_time": 0.1331563217724788, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.376, "pct_cuda_time": 0.1331563217724788, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 429.991, "cuda_time_us": 44.224, "pct_cuda_time": 0.6280615586674597, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.224, "pct_cuda_time": 0.6280615586674597, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2584.436, "cuda_time_us": 205.82199999999997, "pct_cuda_time": 2.923048257237108, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.098, "cuda_time_us": 3.04, "pct_cuda_time": 0.04317355142793681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04317355142793681, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1863.314, "cuda_time_us": 64.672, "pct_cuda_time": 0.9184604993248451, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.534, "cuda_time_us": 21.472, "pct_cuda_time": 0.30494161061205893, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.472, "pct_cuda_time": 0.30494161061205893, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 539.243, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 847.412, "cuda_time_us": 21.44, "pct_cuda_time": 0.3044871521759754, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.28, "pct_cuda_time": 0.2454075554851145, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.020450629623759542, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.024, "cuda_time_us": 17.951999999999998, "pct_cuda_time": 0.2549511826428689, "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": 0.22450246742527138, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.088, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 492.533, "cuda_time_us": 135.00599999999997, "pct_cuda_time": 1.9173317381842223, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.701, "cuda_time_us": 81.951, "pct_cuda_time": 1.163853852983832, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.951, "pct_cuda_time": 1.163853852983832, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 106.331, "cuda_time_us": 9.088, "pct_cuda_time": 0.12906619584772688, "trace": "" }, "children": [ { "entry": { "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": 0.12906619584772688, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.41, "cuda_time_us": 43.967, "pct_cuda_time": 0.6244116893526638, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.967, "pct_cuda_time": 0.6244116893526638, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2488.759, "cuda_time_us": 204.63800000000003, "pct_cuda_time": 2.9062332951020178, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.756, "cuda_time_us": 3.2, "pct_cuda_time": 0.04544584360835454, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04544584360835454, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1783.698, "cuda_time_us": 63.168000000000006, "pct_cuda_time": 0.8971009528289187, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.568, "cuda_time_us": 20.48, "pct_cuda_time": 0.290853399093469, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.48, "pct_cuda_time": 0.290853399093469, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 471.243, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 797.39, "cuda_time_us": 21.248, "pct_cuda_time": 0.30176040155947415, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.28, "pct_cuda_time": 0.2454075554851145, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 159.415, "cuda_time_us": 17.664, "pct_cuda_time": 0.25086105671811704, "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": 0.22041234150051947, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.845, "cuda_time_us": 3.328, "pct_cuda_time": 0.047263677352688716, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.328, "pct_cuda_time": 0.047263677352688716, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 466.274, "cuda_time_us": 134.942, "pct_cuda_time": 1.9164228213120558, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.234, "cuda_time_us": 81.855, "pct_cuda_time": 1.1624904776755813, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.855, "pct_cuda_time": 1.1624904776755813, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.623, "cuda_time_us": 9.119, "pct_cuda_time": 0.12950645245768283, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.119, "pct_cuda_time": 0.12950645245768283, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.928, "cuda_time_us": 43.968, "pct_cuda_time": 0.6244258911787914, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.968, "pct_cuda_time": 0.6244258911787914, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2741.938, "cuda_time_us": 205.43600000000004, "pct_cuda_time": 2.9175663523518516, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.549, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1935.131, "cuda_time_us": 65.311, "pct_cuda_time": 0.9275354662203885, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.3, "cuda_time_us": 21.407, "pct_cuda_time": 0.30401849191376423, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.407, "pct_cuda_time": 0.30401849191376423, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 488.086, "cuda_time_us": 3.872, "pct_cuda_time": 0.05498947076610899, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.05498947076610899, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 916.307, "cuda_time_us": 21.888, "pct_cuda_time": 0.31084957028114507, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.888, "pct_cuda_time": 0.2540422657707019, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.871, "cuda_time_us": 18.144, "pct_cuda_time": 0.2576779332593702, "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": 16.0, "pct_cuda_time": 0.22722921804177268, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 99.786, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 528.421, "cuda_time_us": 134.04500000000002, "pct_cuda_time": 1.9036837832755888, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 192.027, "cuda_time_us": 81.31, "pct_cuda_time": 1.1547504824360335, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.31, "pct_cuda_time": 1.1547504824360335, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 109.338, "cuda_time_us": 9.024, "pct_cuda_time": 0.12815727897555979, "trace": "" }, "children": [ { "entry": { "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": 0.12815727897555979, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.74, "cuda_time_us": 43.711, "pct_cuda_time": 0.6207760218639953, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.711, "pct_cuda_time": 0.6207760218639953, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2699.09, "cuda_time_us": 203.42000000000002, "pct_cuda_time": 2.8889354708785877, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.531, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1915.901, "cuda_time_us": 63.326, "pct_cuda_time": 0.899344841357081, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.795, "cuda_time_us": 20.64, "pct_cuda_time": 0.29312569127388677, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.64, "pct_cuda_time": 0.29312569127388677, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 512.074, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 879.181, "cuda_time_us": 21.150000000000002, "pct_cuda_time": 0.3003686225989683, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.655, "pct_cuda_time": 0.03770584836880665, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.183, "pct_cuda_time": 0.24402997835073625, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.319, "cuda_time_us": 17.728, "pct_cuda_time": 0.2517699735902842, "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": 0.22177571680877012, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.029994256781513993, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 95.265, "cuda_time_us": 3.168, "pct_cuda_time": 0.04499138517227099, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04499138517227099, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 525.396, "cuda_time_us": 133.854, "pct_cuda_time": 1.9009712344852152, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 191.083, "cuda_time_us": 81.119, "pct_cuda_time": 1.1520379336456599, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.119, "pct_cuda_time": 1.1520379336456599, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.645, "cuda_time_us": 8.96, "pct_cuda_time": 0.12724836210339271, "trace": "" }, "children": [ { "entry": { "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": 0.12724836210339271, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 176.207, "cuda_time_us": 43.775, "pct_cuda_time": 0.6216849387361625, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.775, "pct_cuda_time": 0.6216849387361625, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2510.456, "cuda_time_us": 205.34300000000002, "pct_cuda_time": 2.9162455825219835, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.622, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1745.627, "cuda_time_us": 64.70400000000001, "pct_cuda_time": 0.9189149577609289, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.795, "cuda_time_us": 21.504, "pct_cuda_time": 0.30539606904814254, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.504, "pct_cuda_time": 0.30539606904814254, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.757, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 785.492, "cuda_time_us": 21.28, "pct_cuda_time": 0.3022148599955577, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.248, "pct_cuda_time": 0.24495309704903098, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018632795879425363, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 167.604, "cuda_time_us": 18.144, "pct_cuda_time": 0.2576779332593702, "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": 15.968, "pct_cuda_time": 0.22677475960568913, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 2.176, "pct_cuda_time": 0.030903173653681084, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 94.584, "cuda_time_us": 3.04, "pct_cuda_time": 0.04317355142793681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04317355142793681, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 516.689, "cuda_time_us": 134.495, "pct_cuda_time": 1.9100746050330137, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 182.595, "cuda_time_us": 81.567, "pct_cuda_time": 1.1584003517508294, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.567, "pct_cuda_time": 1.1584003517508294, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.043, "cuda_time_us": 8.992, "pct_cuda_time": 0.12770282053947626, "trace": "" }, "children": [ { "entry": { "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": 0.12770282053947626, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.987, "cuda_time_us": 43.936, "pct_cuda_time": 0.6239714327427078, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6239714327427078, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2521.06, "cuda_time_us": 204.511, "pct_cuda_time": 2.9044296631838105, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.268, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1805.02, "cuda_time_us": 64.225, "pct_cuda_time": 0.912112283045803, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.016, "cuda_time_us": 21.44, "pct_cuda_time": 0.3044871521759754, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.44, "pct_cuda_time": 0.3044871521759754, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.981, "cuda_time_us": 3.84, "pct_cuda_time": 0.054535012330025445, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.054535012330025445, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 795.317, "cuda_time_us": 21.152, "pct_cuda_time": 0.3003970262512235, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.12, "pct_cuda_time": 0.2431352633046968, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.344, "pct_cuda_time": 0.01908725431550891, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 165.146, "cuda_time_us": 17.793, "pct_cuda_time": 0.25269309228857884, "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": 0.2222301752448537, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.03046291704372515, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.369, "cuda_time_us": 3.36, "pct_cuda_time": 0.047718135788772265, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.36, "pct_cuda_time": 0.047718135788772265, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 492.634, "cuda_time_us": 133.918, "pct_cuda_time": 1.901880151357382, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.24, "cuda_time_us": 80.639, "pct_cuda_time": 1.1452210571044068, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.639, "pct_cuda_time": 1.1452210571044068, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.317, "cuda_time_us": 9.088, "pct_cuda_time": 0.12906619584772688, "trace": "" }, "children": [ { "entry": { "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": 0.12906619584772688, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 162.901, "cuda_time_us": 44.191, "pct_cuda_time": 0.6275928984052486, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.191, "pct_cuda_time": 0.6275928984052486, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2469.057, "cuda_time_us": 205.469, "pct_cuda_time": 2.918035012614062, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.347, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1748.415, "cuda_time_us": 64.63900000000001, "pct_cuda_time": 0.9179918390626343, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.515, "cuda_time_us": 21.504, "pct_cuda_time": 0.30539606904814254, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.504, "pct_cuda_time": 0.30539606904814254, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 474.169, "cuda_time_us": 3.776, "pct_cuda_time": 0.05362609545785835, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05362609545785835, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 795.005, "cuda_time_us": 21.376, "pct_cuda_time": 0.3035782353038083, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03862896706710136, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.376, "pct_cuda_time": 0.24677093079336515, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 157.03, "cuda_time_us": 17.983, "pct_cuda_time": 0.2553914392528249, "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": 15.84, "pct_cuda_time": 0.22495692586135496, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030434513391469922, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.169, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 481.222, "cuda_time_us": 134.718, "pct_cuda_time": 1.9132416122594709, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.023, "cuda_time_us": 81.087, "pct_cuda_time": 1.1515834752095764, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.087, "pct_cuda_time": 1.1515834752095764, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.217, "cuda_time_us": 9.311, "pct_cuda_time": 0.1322332030741841, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.311, "pct_cuda_time": 0.1322332030741841, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.89, "cuda_time_us": 44.32, "pct_cuda_time": 0.6294249339757103, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6294249339757103, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2733.058, "cuda_time_us": 204.762, "pct_cuda_time": 2.907994321541841, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.947, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1984.303, "cuda_time_us": 63.452, "pct_cuda_time": 0.90113427144916, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.835, "cuda_time_us": 20.608, "pct_cuda_time": 0.2926712328378032, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2926712328378032, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 468.075, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1005.477, "cuda_time_us": 21.182000000000002, "pct_cuda_time": 0.30082308103505184, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.719, "pct_cuda_time": 0.03861476524097374, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.183, "pct_cuda_time": 0.24402997835073625, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.914, "cuda_time_us": 17.854, "pct_cuda_time": 0.25355940368236307, "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": 15.711, "pct_cuda_time": 0.22312489029089316, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030434513391469922, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.21, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 508.135, "cuda_time_us": 135.13400000000001, "pct_cuda_time": 1.9191495719285572, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 183.651, "cuda_time_us": 81.151, "pct_cuda_time": 1.1524923920817434, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.151, "pct_cuda_time": 1.1524923920817434, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.949, "cuda_time_us": 9.28, "pct_cuda_time": 0.13179294646422815, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.28, "pct_cuda_time": 0.13179294646422815, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.307, "cuda_time_us": 44.703, "pct_cuda_time": 0.6348642333825854, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.703, "pct_cuda_time": 0.6348642333825854, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2603.627, "cuda_time_us": 204.802, "pct_cuda_time": 2.908562394586945, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.238, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1867.942, "cuda_time_us": 65.122, "pct_cuda_time": 0.9248513210822701, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.48, "cuda_time_us": 21.6, "pct_cuda_time": 0.30675944435639313, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.6, "pct_cuda_time": 0.30675944435639313, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.793, "cuda_time_us": 3.777, "pct_cuda_time": 0.05364029728398596, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.777, "pct_cuda_time": 0.05364029728398596, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 864.187, "cuda_time_us": 21.504, "pct_cuda_time": 0.30539606904814254, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.816, "pct_cuda_time": 0.03999234237535199, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.408, "pct_cuda_time": 0.24722538922944867, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.879, "cuda_time_us": 18.241, "pct_cuda_time": 0.25905551039374847, "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": 16.096, "pct_cuda_time": 0.2285925933500233, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.03046291704372515, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.989, "cuda_time_us": 3.105, "pct_cuda_time": 0.044096670126231516, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.044096670126231516, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 494.332, "cuda_time_us": 133.503, "pct_cuda_time": 1.8959863935144234, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 177.211, "cuda_time_us": 80.799, "pct_cuda_time": 1.1474933492848247, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.799, "pct_cuda_time": 1.1474933492848247, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.629, "cuda_time_us": 9.088, "pct_cuda_time": 0.12906619584772688, "trace": "" }, "children": [ { "entry": { "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": 0.12906619584772688, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.095, "cuda_time_us": 43.616, "pct_cuda_time": 0.6194268483818723, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.616, "pct_cuda_time": 0.6194268483818723, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2614.027, "cuda_time_us": 204.22, "pct_cuda_time": 2.900296931780676, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.46, "cuda_time_us": 3.008, "pct_cuda_time": 0.042719092991853266, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.042719092991853266, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1886.596, "cuda_time_us": 63.518, "pct_cuda_time": 0.9020715919735824, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.129, "cuda_time_us": 20.544, "pct_cuda_time": 0.2917623159656361, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.544, "pct_cuda_time": 0.2917623159656361, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 505.675, "cuda_time_us": 3.808, "pct_cuda_time": 0.05408055389394189, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05408055389394189, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 809.924, "cuda_time_us": 21.183, "pct_cuda_time": 0.30083728286117944, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.215, "pct_cuda_time": 0.2444844367868198, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 252.565, "cuda_time_us": 17.983, "pct_cuda_time": 0.2553914392528249, "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": 15.775, "pct_cuda_time": 0.22403380716306026, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 2.208, "pct_cuda_time": 0.03135763208976463, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.85, "cuda_time_us": 3.296, "pct_cuda_time": 0.04680921891660517, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.296, "pct_cuda_time": 0.04680921891660517, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 499.233, "cuda_time_us": 134.398, "pct_cuda_time": 1.9086970278986353, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 175.814, "cuda_time_us": 81.695, "pct_cuda_time": 1.1602181854951636, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.695, "pct_cuda_time": 1.1602181854951636, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.061, "cuda_time_us": 9.311, "pct_cuda_time": 0.1322332030741841, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.311, "pct_cuda_time": 0.1322332030741841, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.854, "cuda_time_us": 43.392, "pct_cuda_time": 0.6162456393292876, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.392, "pct_cuda_time": 0.6162456393292876, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2698.858, "cuda_time_us": 206.01299999999998, "pct_cuda_time": 2.925760806027482, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.583, "cuda_time_us": 3.04, "pct_cuda_time": 0.04317355142793681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04317355142793681, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1959.805, "cuda_time_us": 64.414, "pct_cuda_time": 0.9147964281839217, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 204.882, "cuda_time_us": 21.439, "pct_cuda_time": 0.3044729503498478, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.439, "pct_cuda_time": 0.3044729503498478, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 511.166, "cuda_time_us": 3.744, "pct_cuda_time": 0.053171637021774805, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.053171637021774805, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 897.863, "cuda_time_us": 21.151, "pct_cuda_time": 0.3003828244250959, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.183, "pct_cuda_time": 0.24402997835073625, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 167.583, "cuda_time_us": 18.08, "pct_cuda_time": 0.2567690163872031, "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": 15.936, "pct_cuda_time": 0.2263203011696056, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.030448715217597542, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.753, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.882, "cuda_time_us": 135.45499999999998, "pct_cuda_time": 1.9237083581155194, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.554, "cuda_time_us": 81.535, "pct_cuda_time": 1.157945893314746, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.535, "pct_cuda_time": 1.157945893314746, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 108.062, "cuda_time_us": 9.024, "pct_cuda_time": 0.12815727897555979, "trace": "" }, "children": [ { "entry": { "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": 0.12815727897555979, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 155.296, "cuda_time_us": 44.896, "pct_cuda_time": 0.6376051858252142, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.896, "pct_cuda_time": 0.6376051858252142, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2513.342, "cuda_time_us": 203.998, "pct_cuda_time": 2.897144126380346, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.906, "cuda_time_us": 3.072, "pct_cuda_time": 0.04362800986402036, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04362800986402036, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1807.183, "cuda_time_us": 63.456, "pct_cuda_time": 0.9011910787536705, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.206, "cuda_time_us": 20.384, "pct_cuda_time": 0.2894900237852184, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2894900237852184, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[8, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 489.926, "cuda_time_us": 4.0, "pct_cuda_time": 0.05680730451044317, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 4.0, "pct_cuda_time": 0.05680730451044317, "trace": "_C::rotary_embedding(int64[8], bfloat16[8, 4096], bfloat16[8, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 836.163, "cuda_time_us": 21.183, "pct_cuda_time": 0.30083728286117944, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.688, "pct_cuda_time": 0.03817450863101782, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[8], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 17.215, "pct_cuda_time": 0.2444844367868198, "trace": "_vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.018178337443341814, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[8, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[8, 1, 32, 128], None, None, None, None, int32[8], None, None, int32[8, 33], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[8, 32, 128], bfloat16[8, 8, 128], bfloat16[8, 8, 128], bfloat16[8, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.835, "cuda_time_us": 17.889, "pct_cuda_time": 0.2540564675968295, "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": 0.22359355055310434, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 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": 0.03046291704372515, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[8, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.038, "cuda_time_us": 3.104, "pct_cuda_time": 0.0440824683001039, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.0440824683001039, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 483.278, "cuda_time_us": 134.36599999999999, "pct_cuda_time": 1.9082425694625518, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.275, "cuda_time_us": 81.919, "pct_cuda_time": 1.1633993945477485, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.919, "pct_cuda_time": 1.1633993945477485, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[8, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.63, "cuda_time_us": 9.12, "pct_cuda_time": 0.12952065428381043, "trace": "" }, "children": [ { "entry": { "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": 0.12952065428381043, "trace": "_C::silu_and_mul(bfloat16[8, 14336], bfloat16[8, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.825, "cuda_time_us": 43.327, "pct_cuda_time": 0.6153225206309928, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.327, "pct_cuda_time": 0.6153225206309928, "trace": "mm(bfloat16[8, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[8, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[8, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.899, "cuda_time_us": 3.04, "pct_cuda_time": 0.04317355142793681, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04317355142793681, "trace": "_C::fused_add_rms_norm(bfloat16[8, 4096], bfloat16[8, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 516.199, "cuda_time_us": 347.899, "pct_cuda_time": 4.940801107969667, "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": 5.568, "pct_cuda_time": 0.07907576787853689, "trace": "index_select(bfloat16[8, 4096], 0, int64[8])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.009998085593837998, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[8, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 341.627, "pct_cuda_time": 4.851727254497292, "trace": "mm(bfloat16[8, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[8, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[8, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3556.953, "cuda_time_us": 119.42300000000002, "pct_cuda_time": 1.6960246816376638, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010452544029921542, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 0.010452544029921542, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 0.011361460902088635, "trace": "copy_(int32[8], int32[8], True) <- _to_copy(int32[8], 3, 0, None, None, True, None) <- to(int32[8], 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": 0.01090700246600509, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 0.011361460902088635, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 0.011361460902088635, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 0.01090700246600509, "trace": "copy_(bfloat16[8], bfloat16[8], True) <- _to_copy(bfloat16[8], 15, 0, None, None, True, None) <- to(bfloat16[8], 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": 4.736, "pct_cuda_time": 0.06725984854036471, "trace": "copy_(float32[8, 128256], bfloat16[8, 128256], False) <- _to_copy(bfloat16[8, 128256], 6, None, None, None, False, None) <- to(bfloat16[8, 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": 6.368, "pct_cuda_time": 0.09043722878062553, "trace": "div_(float32[8, 128256], bfloat16[8, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.88, "pct_cuda_time": 0.4953596953310645, "trace": "_softmax(float32[8, 128256], -1, False) <- softmax(float32[8, 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": 28.735, "pct_cuda_time": 0.40808947377689614, "trace": "_log_softmax(float32[8, 128256], -1, False) <- log_softmax(float32[8, 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": 0.02590413085676209, "trace": "copy_(int64[8], int32[8], False) <- _to_copy(int32[8], 4, None, None, None, False, None) <- to(int32[8], 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": 7.36, "pct_cuda_time": 0.10452544029921544, "trace": "index(float32[8, 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": 27.328, "pct_cuda_time": 0.3881075044153477, "trace": "argmax(float32[8, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.784, "pct_cuda_time": 0.039537883939268444, "trace": "copy_(int64[8], int64[8], False) <- _to_copy(int64[8], 4, 0, None, None, False, None) <- to(int64[8], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }