{ "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": 2, "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": 12525.517, "pct_cuda_time": 96.32804200821334, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 16.191, "pct_cuda_time": 0.1245176010024163, "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": 16.191, "pct_cuda_time": 0.1245176010024163, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 12502.926, "pct_cuda_time": 96.15430492438618, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 412.63200000000006, "pct_cuda_time": 3.173364630771975, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 7.712, "pct_cuda_time": 0.05930947680381906, "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": 404.92, "pct_cuda_time": 3.1140551539681556, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 3154.2830000000004, "pct_cuda_time": 24.25815280357635, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 1410.218, "pct_cuda_time": 10.845343848460596, "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": 1410.218, "pct_cuda_time": 10.845343848460596, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 219.22900000000004, "pct_cuda_time": 1.6859903125291045, "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": 219.22900000000004, "pct_cuda_time": 1.6859903125291045, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 531.319, "pct_cuda_time": 4.0861322492126995, "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": 118.52900000000001, "pct_cuda_time": 0.9115525124584893, "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": 368.09100000000007, "pct_cuda_time": 2.830820101944316, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 44.699000000000005, "pct_cuda_time": 0.3437596348098948, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 993.5169999999998, "pct_cuda_time": 7.640686393373949, "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": 993.5169999999998, "pct_cuda_time": 7.640686393373949, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 8936.011000000002, "pct_cuda_time": 68.72278749003789, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 5443.577000000001, "pct_cuda_time": 41.86406947760672, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 24.864, "pct_cuda_time": 0.19121769077413864, "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": 5418.713000000001, "pct_cuda_time": 41.672851786832574, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 727.8920000000002, "pct_cuda_time": 5.5978855925422035, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 727.8920000000002, "pct_cuda_time": 5.5978855925422035, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 2764.5420000000004, "pct_cuda_time": 21.26083241988895, "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": 2764.5420000000004, "pct_cuda_time": 21.26083241988895, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "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": 6.4, "pct_cuda_time": 0.049219482824746114, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 363.099, "pct_cuda_time": 2.7924289053410134, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 2.976, "pct_cuda_time": 0.022887059513506943, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "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": 359.387, "pct_cuda_time": 2.763881605302661, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 114.36500000000001, "pct_cuda_time": 0.8795290864456389, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.438000000000001, "pct_cuda_time": 0.04182117931265147, "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": 4.064, "pct_cuda_time": 0.031254371593713776, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 4.64, "pct_cuda_time": 0.035684125047940926, "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": 34.303, "pct_cuda_time": 0.2638087373964478, "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": 28.031, "pct_cuda_time": 0.21557364422819658, "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": 0.014027552605052642, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 4.608, "pct_cuda_time": 0.0354380276338172, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cuda_time_us": 28.832, "pct_cuda_time": 0.22173377012548123, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.625, "pct_cuda_time": 0.020187678502337273, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 91371.419, "cuda_time_us": 12525.517, "pct_cuda_time": 96.32804200821334, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 321.891, "cuda_time_us": 16.191, "pct_cuda_time": 0.1245176010024163, "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": 16.191, "pct_cuda_time": 0.1245176010024163, "trace": "index_select(bfloat16[128256, 4096], 0, int64[512]) <- embedding(bfloat16[128256, 4096], int64[512], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 8040.812, "cuda_time_us": 393.597, "pct_cuda_time": 3.0269751220893117, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 3653.952, "cuda_time_us": 7.712, "pct_cuda_time": 0.05930947680381906, "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": 7.712, "pct_cuda_time": 0.05930947680381906, "trace": "_C::rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3503.411, "cuda_time_us": 99.616, "pct_cuda_time": 0.7661012501671732, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 739.921, "cuda_time_us": 44.831, "pct_cuda_time": 0.34477478664315514, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.831, "pct_cuda_time": 0.34477478664315514, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 1018.475, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1155.891, "cuda_time_us": 16.929000000000002, "pct_cuda_time": 0.13019322261564484, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.841, "pct_cuda_time": 0.029539380239039036, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.6, "pct_cuda_time": 0.012304870706186528, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 296.219, "cuda_time_us": 31.264, "pct_cuda_time": 0.24043717359888472, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.264, "pct_cuda_time": 0.24043717359888472, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 124.26, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 606.507, "cuda_time_us": 279.933, "pct_cuda_time": 2.1528371071218206, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 211.609, "cuda_time_us": 169.726, "pct_cuda_time": 1.305285303423884, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.99, "pct_cuda_time": 1.2996250628990382, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 150.007, "cuda_time_us": 22.528, "pct_cuda_time": 0.1732525795431063, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.528, "pct_cuda_time": 0.1732525795431063, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.136, "cuda_time_us": 87.679, "pct_cuda_time": 0.6742992241548303, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 87.679, "pct_cuda_time": 0.6742992241548303, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2703.973, "cuda_time_us": 396.474, "pct_cuda_time": 3.049100817727873, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.128, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1924.784, "cuda_time_us": 97.726, "pct_cuda_time": 0.7515661216454903, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.552, "cuda_time_us": 43.903, "pct_cuda_time": 0.33763796163356696, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.903, "pct_cuda_time": 0.33763796163356696, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 551.216, "cuda_time_us": 6.752, "pct_cuda_time": 0.05192655438010714, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.752, "pct_cuda_time": 0.05192655438010714, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 844.581, "cuda_time_us": 16.512, "pct_cuda_time": 0.12698626568784496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.776, "pct_cuda_time": 0.029039494866600207, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.424, "pct_cuda_time": 0.08785677684217182, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.010089993979072953, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.49, "cuda_time_us": 30.559, "pct_cuda_time": 0.23501533994397134, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.559, "pct_cuda_time": 0.23501533994397134, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 103.889, "cuda_time_us": 6.528, "pct_cuda_time": 0.05020387248124103, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.528, "pct_cuda_time": 0.05020387248124103, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 492.532, "cuda_time_us": 285.852, "pct_cuda_time": 2.1983574381905195, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.171, "cuda_time_us": 171.006, "pct_cuda_time": 1.3151291999888335, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.27, "pct_cuda_time": 1.3094689594639877, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.596, "cuda_time_us": 22.688, "pct_cuda_time": 0.17448306661372495, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.688, "pct_cuda_time": 0.17448306661372495, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.412, "cuda_time_us": 92.158, "pct_cuda_time": 0.7087451715879612, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 92.158, "pct_cuda_time": 0.7087451715879612, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2652.736, "cuda_time_us": 387.034, "pct_cuda_time": 2.9765020805613727, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.602, "cuda_time_us": 6.303, "pct_cuda_time": 0.048473500038183556, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.303, "pct_cuda_time": 0.048473500038183556, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1893.795, "cuda_time_us": 99.007, "pct_cuda_time": 0.761417708754631, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.472, "cuda_time_us": 44.128, "pct_cuda_time": 0.3393683340766244, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.128, "pct_cuda_time": 0.3393683340766244, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 534.349, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 843.1, "cuda_time_us": 16.352, "pct_cuda_time": 0.1257557786172263, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.456, "pct_cuda_time": 0.08810287425629552, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.897, "cuda_time_us": 31.423, "pct_cuda_time": 0.24165997012531204, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.423, "pct_cuda_time": 0.24165997012531204, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.071, "cuda_time_us": 6.56, "pct_cuda_time": 0.05044996989536476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.56, "pct_cuda_time": 0.05044996989536476, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 495.514, "cuda_time_us": 275.164, "pct_cuda_time": 2.1161609018731933, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.331, "cuda_time_us": 169.59799999999998, "pct_cuda_time": 1.304300913767389, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.862, "pct_cuda_time": 1.2986406732425433, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.907, "cuda_time_us": 22.911, "pct_cuda_time": 0.1761980579683997, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.911, "pct_cuda_time": 0.1761980579683997, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.471, "cuda_time_us": 82.655, "pct_cuda_time": 0.6356619301374047, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.655, "pct_cuda_time": 0.6356619301374047, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2512.688, "cuda_time_us": 385.27699999999993, "pct_cuda_time": 2.9629897944171413, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.878, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1809.349, "cuda_time_us": 98.078, "pct_cuda_time": 0.7542731932008514, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.64, "cuda_time_us": 43.711, "pct_cuda_time": 0.33616137714882455, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.711, "pct_cuda_time": 0.33616137714882455, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.393, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 788.985, "cuda_time_us": 16.704, "pct_cuda_time": 0.12846285017258735, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.028793397452476475, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011320481049691605, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.505, "cuda_time_us": 31.039, "pct_cuda_time": 0.2387068011558273, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.039, "pct_cuda_time": 0.2387068011558273, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.468, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.489, "cuda_time_us": 274.27099999999996, "pct_cuda_time": 2.109293245910303, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.78, "cuda_time_us": 170.271, "pct_cuda_time": 1.3094766500081787, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.005667931069037169, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.534, "pct_cuda_time": 1.3038087189391416, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 106.125, "cuda_time_us": 22.56, "pct_cuda_time": 0.17349867695723004, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.56, "pct_cuda_time": 0.17349867695723004, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.027, "cuda_time_us": 81.44, "pct_cuda_time": 0.6263179189448942, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.44, "pct_cuda_time": 0.6263179189448942, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2415.705, "cuda_time_us": 385.20899999999995, "pct_cuda_time": 2.9624668374121286, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.494, "cuda_time_us": 6.303, "pct_cuda_time": 0.048473500038183556, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.303, "pct_cuda_time": 0.048473500038183556, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1680.797, "cuda_time_us": 98.205, "pct_cuda_time": 0.755249892313155, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.05, "cuda_time_us": 44.095, "pct_cuda_time": 0.3391145461183093, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.095, "pct_cuda_time": 0.3391145461183093, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 461.481, "cuda_time_us": 6.623, "pct_cuda_time": 0.05093447417942086, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.623, "pct_cuda_time": 0.05093447417942086, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.526, "cuda_time_us": 16.735, "pct_cuda_time": 0.1287012570425197, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.028793397452476475, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.01155888791962397, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 172.233, "cuda_time_us": 30.752, "pct_cuda_time": 0.23649961497290506, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.752, "pct_cuda_time": 0.23649961497290506, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.517, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 500.279, "cuda_time_us": 274.205, "pct_cuda_time": 2.1087856699936727, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.49, "cuda_time_us": 169.85399999999998, "pct_cuda_time": 1.3062696930803788, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.118, "pct_cuda_time": 1.3006094525555334, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 138.128, "cuda_time_us": 22.656, "pct_cuda_time": 0.17423696919960122, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.656, "pct_cuda_time": 0.17423696919960122, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.014, "cuda_time_us": 81.695, "pct_cuda_time": 0.6282790077136927, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.695, "pct_cuda_time": 0.6282790077136927, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2816.741, "cuda_time_us": 387.738, "pct_cuda_time": 2.9819162236720946, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.767, "cuda_time_us": 6.304, "pct_cuda_time": 0.04848119058237492, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.304, "pct_cuda_time": 0.04848119058237492, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2051.623, "cuda_time_us": 98.11, "pct_cuda_time": 0.7545192906149751, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.32, "cuda_time_us": 43.647, "pct_cuda_time": 0.33566918232057713, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.647, "pct_cuda_time": 0.33566918232057713, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 716.616, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 827.042, "cuda_time_us": 16.48, "pct_cuda_time": 0.12674016827372125, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.02854730003835275, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.762, "cuda_time_us": 30.879, "pct_cuda_time": 0.23747631408520864, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.879, "pct_cuda_time": 0.23747631408520864, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.405, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 488.192, "cuda_time_us": 276.668, "pct_cuda_time": 2.127727480337009, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.948, "cuda_time_us": 170.398, "pct_cuda_time": 1.3104533491204824, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.005906337938969533, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.63, "pct_cuda_time": 1.3045470111815127, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.101, "cuda_time_us": 22.559, "pct_cuda_time": 0.17349098641303867, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.559, "pct_cuda_time": 0.17349098641303867, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.381, "cuda_time_us": 83.711, "pct_cuda_time": 0.6437831448034878, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.711, "pct_cuda_time": 0.6437831448034878, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2518.311, "cuda_time_us": 386.524, "pct_cuda_time": 2.9725799030237763, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.333, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1816.193, "cuda_time_us": 97.343, "pct_cuda_time": 0.748620643220197, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.956, "cuda_time_us": 43.775, "pct_cuda_time": 0.336653571977072, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.775, "pct_cuda_time": 0.336653571977072, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 545.541, "cuda_time_us": 7.04, "pct_cuda_time": 0.05414143110722072, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.04, "pct_cuda_time": 0.05414143110722072, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 785.754, "cuda_time_us": 16.417, "pct_cuda_time": 0.12625566398966515, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.585, "pct_cuda_time": 0.027570600926049186, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.552, "pct_cuda_time": 0.08884116649866673, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 184.501, "cuda_time_us": 30.111, "pct_cuda_time": 0.23156997614623911, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.111, "pct_cuda_time": 0.23156997614623911, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.753, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.507, "cuda_time_us": 276.189, "pct_cuda_time": 2.1240437096693445, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.371, "cuda_time_us": 170.847, "pct_cuda_time": 1.3139064034624062, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.005667931069037169, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.11, "pct_cuda_time": 1.3082384723933689, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.777, "cuda_time_us": 22.911, "pct_cuda_time": 0.1761980579683997, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.911, "pct_cuda_time": 0.1761980579683997, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.469, "cuda_time_us": 82.431, "pct_cuda_time": 0.6339392482385385, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.431, "pct_cuda_time": 0.6339392482385385, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2460.471, "cuda_time_us": 392.88800000000003, "pct_cuda_time": 3.021522526257633, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.598, "cuda_time_us": 6.239, "pct_cuda_time": 0.04798130520993609, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.239, "pct_cuda_time": 0.04798130520993609, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1724.806, "cuda_time_us": 99.422, "pct_cuda_time": 0.7646092845940481, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.576, "cuda_time_us": 44.063, "pct_cuda_time": 0.33886844870418564, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.063, "pct_cuda_time": 0.33886844870418564, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.219, "cuda_time_us": 6.784, "pct_cuda_time": 0.052172651794230876, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.784, "pct_cuda_time": 0.052172651794230876, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 762.723, "cuda_time_us": 16.671, "pct_cuda_time": 0.12820906221427225, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.68, "pct_cuda_time": 0.028301202624229017, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.503, "pct_cuda_time": 0.01155888791962397, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.219, "cuda_time_us": 31.904, "pct_cuda_time": 0.2453591218813594, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.904, "pct_cuda_time": 0.2453591218813594, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.372, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 501.114, "cuda_time_us": 280.827, "pct_cuda_time": 2.1597124536289027, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.96, "cuda_time_us": 169.533, "pct_cuda_time": 1.3038010283949502, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.797, "pct_cuda_time": 1.2981407878701046, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 126.817, "cuda_time_us": 22.751, "pct_cuda_time": 0.17496757089778106, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.751, "pct_cuda_time": 0.17496757089778106, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.64, "cuda_time_us": 88.543, "pct_cuda_time": 0.6809438543361712, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 88.543, "pct_cuda_time": 0.6809438543361712, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2572.064, "cuda_time_us": 391.73699999999997, "pct_cuda_time": 3.01267070989337, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.856, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1777.19, "cuda_time_us": 98.36600000000001, "pct_cuda_time": 0.7564880699279651, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.212, "cuda_time_us": 44.383, "pct_cuda_time": 0.34132942284542295, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.383, "pct_cuda_time": 0.34132942284542295, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 522.538, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 754.941, "cuda_time_us": 16.479, "pct_cuda_time": 0.12673247772952986, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.027562910381857826, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.423, "pct_cuda_time": 0.08784908629798044, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011320481049691605, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 175.315, "cuda_time_us": 30.4, "pct_cuda_time": 0.23379254341754402, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.4, "pct_cuda_time": 0.23379254341754402, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 99.725, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 535.492, "cuda_time_us": 280.50699999999995, "pct_cuda_time": 2.1572514794876647, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.351, "cuda_time_us": 169.75699999999998, "pct_cuda_time": 1.3055237102938164, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.021, "pct_cuda_time": 1.2998634697689706, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.655, "cuda_time_us": 22.815, "pct_cuda_time": 0.17545976572602853, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.815, "pct_cuda_time": 0.17545976572602853, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 207.02, "cuda_time_us": 87.935, "pct_cuda_time": 0.6762680034678202, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 87.935, "pct_cuda_time": 0.6762680034678202, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2426.159, "cuda_time_us": 391.609, "pct_cuda_time": 3.011686320236875, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.298, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1733.731, "cuda_time_us": 97.886, "pct_cuda_time": 0.752796608716109, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.067, "cuda_time_us": 44.031, "pct_cuda_time": 0.33862235129006185, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.031, "pct_cuda_time": 0.33862235129006185, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.371, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 782.642, "cuda_time_us": 16.416, "pct_cuda_time": 0.12624797344547378, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.028055105210105284, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.415, "cuda_time_us": 30.847, "pct_cuda_time": 0.23723021667108488, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.847, "pct_cuda_time": 0.23723021667108488, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.79, "cuda_time_us": 6.24, "pct_cuda_time": 0.04798899575412746, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.24, "pct_cuda_time": 0.04798899575412746, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.72, "cuda_time_us": 281.115, "pct_cuda_time": 2.1619273303560163, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.637, "cuda_time_us": 169.11700000000002, "pct_cuda_time": 1.3006017620113421, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.005652549980654436, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.382, "pct_cuda_time": 1.2949492120306876, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.993, "cuda_time_us": 23.456, "pct_cuda_time": 0.1803894045526945, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 23.456, "pct_cuda_time": 0.1803894045526945, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.995, "cuda_time_us": 88.542, "pct_cuda_time": 0.6809361637919797, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 88.542, "pct_cuda_time": 0.6809361637919797, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2506.048, "cuda_time_us": 392.057, "pct_cuda_time": 3.0151316840346074, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.987, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1808.96, "cuda_time_us": 99.29400000000001, "pct_cuda_time": 0.7636248949375533, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.625, "cuda_time_us": 44.127, "pct_cuda_time": 0.33936064353243306, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.127, "pct_cuda_time": 0.33936064353243306, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 512.27, "cuda_time_us": 6.976, "pct_cuda_time": 0.05364923627897326, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.976, "pct_cuda_time": 0.05364923627897326, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 814.084, "cuda_time_us": 16.767, "pct_cuda_time": 0.12894735445664343, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.029777787108971398, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.615, "pct_cuda_time": 0.08932567078272283, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.713, "cuda_time_us": 31.424, "pct_cuda_time": 0.2416676606695034, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.424, "pct_cuda_time": 0.2416676606695034, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.541, "cuda_time_us": 6.367, "pct_cuda_time": 0.04896569486643102, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.367, "pct_cuda_time": 0.04896569486643102, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.763, "cuda_time_us": 280.06, "pct_cuda_time": 2.1538138062341243, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.315, "cuda_time_us": 170.845, "pct_cuda_time": 1.3138910223740232, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.109, "pct_cuda_time": 1.3082307818491776, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.339, "cuda_time_us": 22.784, "pct_cuda_time": 0.17522135885609616, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.784, "pct_cuda_time": 0.17522135885609616, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.784, "cuda_time_us": 86.431, "pct_cuda_time": 0.6647014250040049, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 86.431, "pct_cuda_time": 0.6647014250040049, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2569.748, "cuda_time_us": 396.408, "pct_cuda_time": 3.0485932418112434, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 94.716, "cuda_time_us": 6.24, "pct_cuda_time": 0.04798899575412746, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.24, "pct_cuda_time": 0.04798899575412746, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1846.149, "cuda_time_us": 100.76499999999999, "pct_cuda_time": 0.7749376854430533, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 148.271, "cuda_time_us": 44.415, "pct_cuda_time": 0.34157552025954663, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.415, "pct_cuda_time": 0.34157552025954663, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 473.454, "cuda_time_us": 7.135, "pct_cuda_time": 0.054872032805400546, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.135, "pct_cuda_time": 0.054872032805400546, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 773.166, "cuda_time_us": 16.928, "pct_cuda_time": 0.13018553207145347, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.936, "pct_cuda_time": 0.030269981937218856, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.52, "pct_cuda_time": 0.088595069084543, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011320481049691605, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.861, "cuda_time_us": 32.287, "pct_cuda_time": 0.24830460030665275, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 32.287, "pct_cuda_time": 0.24830460030665275, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.978, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.308, "cuda_time_us": 282.90700000000004, "pct_cuda_time": 2.1757087855469455, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.711, "cuda_time_us": 170.55700000000002, "pct_cuda_time": 1.3116761456469097, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.703, "pct_cuda_time": 0.005406452566530705, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.854, "pct_cuda_time": 1.3062696930803792, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.457, "cuda_time_us": 22.368, "pct_cuda_time": 0.17202209247248765, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.368, "pct_cuda_time": 0.17202209247248765, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.552, "cuda_time_us": 89.982, "pct_cuda_time": 0.6920105474275475, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 89.982, "pct_cuda_time": 0.6920105474275475, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2763.176, "cuda_time_us": 386.713, "pct_cuda_time": 2.9740334158759443, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.736, "cuda_time_us": 6.399, "pct_cuda_time": 0.049211792280554736, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.399, "pct_cuda_time": 0.049211792280554736, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2031.805, "cuda_time_us": 97.406, "pct_cuda_time": 0.749105147504253, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.496, "cuda_time_us": 43.552, "pct_cuda_time": 0.3349385806223973, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.552, "pct_cuda_time": 0.3349385806223973, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.828, "cuda_time_us": 6.848, "pct_cuda_time": 0.05266484662247834, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.848, "pct_cuda_time": 0.05266484662247834, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1041.715, "cuda_time_us": 16.607, "pct_cuda_time": 0.1277168673860248, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.424, "pct_cuda_time": 0.08785677684217182, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.567, "pct_cuda_time": 0.012051082747871431, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 211.585, "cuda_time_us": 30.399, "pct_cuda_time": 0.23378485287335266, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.399, "pct_cuda_time": 0.23378485287335266, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 96.598, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 490.278, "cuda_time_us": 276.31600000000003, "pct_cuda_time": 2.125020408781648, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 174.083, "cuda_time_us": 172.061, "pct_cuda_time": 1.3232427241107252, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 1.919, "pct_cuda_time": 0.014758154303232468, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.142, "pct_cuda_time": 1.3084845698074925, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 110.002, "cuda_time_us": 22.912, "pct_cuda_time": 0.17620574851259105, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.912, "pct_cuda_time": 0.17620574851259105, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.01, "cuda_time_us": 81.343, "pct_cuda_time": 0.6255719361583317, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.343, "pct_cuda_time": 0.6255719361583317, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2478.618, "cuda_time_us": 385.625, "pct_cuda_time": 2.9656661037957375, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.054, "cuda_time_us": 6.56, "pct_cuda_time": 0.05044996989536476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.56, "pct_cuda_time": 0.05044996989536476, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1777.866, "cuda_time_us": 98.36600000000001, "pct_cuda_time": 0.7564880699279651, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.096, "cuda_time_us": 44.063, "pct_cuda_time": 0.33886844870418564, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.063, "pct_cuda_time": 0.33886844870418564, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 522.627, "cuda_time_us": 6.591, "pct_cuda_time": 0.05068837676529713, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.591, "pct_cuda_time": 0.05068837676529713, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 767.55, "cuda_time_us": 16.544, "pct_cuda_time": 0.1272323631019687, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.028793397452476475, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.52, "pct_cuda_time": 0.088595069084543, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.559, "cuda_time_us": 31.168, "pct_cuda_time": 0.23969888135651354, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.168, "pct_cuda_time": 0.23969888135651354, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.434, "cuda_time_us": 6.591, "pct_cuda_time": 0.05068837676529713, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.591, "pct_cuda_time": 0.05068837676529713, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 461.707, "cuda_time_us": 274.108, "pct_cuda_time": 2.1080396872071105, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.32, "cuda_time_us": 169.278, "pct_cuda_time": 1.3018399396261517, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.005906337938969533, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.51, "pct_cuda_time": 1.2959336016871823, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.742, "cuda_time_us": 22.591, "pct_cuda_time": 0.1737370838271624, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.591, "pct_cuda_time": 0.1737370838271624, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.946, "cuda_time_us": 82.239, "pct_cuda_time": 0.6324626637537961, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.239, "pct_cuda_time": 0.6324626637537961, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2579.568, "cuda_time_us": 387.385, "pct_cuda_time": 2.9792014615725426, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.738, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1866.02, "cuda_time_us": 99.038, "pct_cuda_time": 0.7616561156245633, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.507, "cuda_time_us": 44.159, "pct_cuda_time": 0.3396067409465568, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.159, "pct_cuda_time": 0.3396067409465568, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 534.833, "cuda_time_us": 7.232, "pct_cuda_time": 0.0556180155919631, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.232, "pct_cuda_time": 0.0556180155919631, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 811.648, "cuda_time_us": 16.543, "pct_cuda_time": 0.12722467255777734, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.647, "pct_cuda_time": 0.028047414665913917, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.08933336132691419, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 183.006, "cuda_time_us": 31.104, "pct_cuda_time": 0.2392066865282661, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.104, "pct_cuda_time": 0.2392066865282661, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.472, "cuda_time_us": 6.495, "pct_cuda_time": 0.04995008452292594, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.495, "pct_cuda_time": 0.04995008452292594, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.282, "cuda_time_us": 275.484, "pct_cuda_time": 2.118621876014431, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.498, "cuda_time_us": 169.59799999999998, "pct_cuda_time": 1.304300913767389, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.862, "pct_cuda_time": 1.2986406732425433, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.547, "cuda_time_us": 22.463, "pct_cuda_time": 0.1727526941706675, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.463, "pct_cuda_time": 0.1727526941706675, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.144, "cuda_time_us": 83.423, "pct_cuda_time": 0.6415682680763742, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.423, "pct_cuda_time": 0.6415682680763742, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2460.145, "cuda_time_us": 400.50800000000004, "pct_cuda_time": 3.0801244729958466, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.39, "cuda_time_us": 6.304, "pct_cuda_time": 0.04848119058237492, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.304, "pct_cuda_time": 0.04848119058237492, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1762.563, "cuda_time_us": 98.97500000000001, "pct_cuda_time": 0.7611716113405073, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.854, "cuda_time_us": 43.871, "pct_cuda_time": 0.3373918642194432, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.871, "pct_cuda_time": 0.3373918642194432, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 482.283, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 783.084, "cuda_time_us": 16.608, "pct_cuda_time": 0.12772455793021614, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011566578463815336, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 180.144, "cuda_time_us": 31.84, "pct_cuda_time": 0.2448669270531119, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.84, "pct_cuda_time": 0.2448669270531119, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.392, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.592, "pct_cuda_time": 0.050696067309488495, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.925, "cuda_time_us": 288.637, "pct_cuda_time": 2.2197756037634755, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.997, "cuda_time_us": 171.07, "pct_cuda_time": 1.3156213948170807, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.334, "pct_cuda_time": 1.309961154292235, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.434, "cuda_time_us": 23.008, "pct_cuda_time": 0.17694404075496226, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 23.008, "pct_cuda_time": 0.17694404075496226, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.34, "cuda_time_us": 94.559, "pct_cuda_time": 0.7272101681914324, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 94.559, "pct_cuda_time": 0.7272101681914324, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2380.733, "cuda_time_us": 389.367, "pct_cuda_time": 2.9944441201598315, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.434, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1701.62, "cuda_time_us": 99.197, "pct_cuda_time": 0.7628789121509907, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.086, "cuda_time_us": 44.255, "pct_cuda_time": 0.340345033188928, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.255, "pct_cuda_time": 0.340345033188928, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.848, "cuda_time_us": 6.784, "pct_cuda_time": 0.052172651794230876, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.784, "pct_cuda_time": 0.052172651794230876, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 766.659, "cuda_time_us": 16.735, "pct_cuda_time": 0.1287012570425197, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.68, "pct_cuda_time": 0.028301202624229017, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.615, "pct_cuda_time": 0.08932567078272283, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011074383635567875, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.854, "cuda_time_us": 31.423, "pct_cuda_time": 0.24165997012531204, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.423, "pct_cuda_time": 0.24165997012531204, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.915, "cuda_time_us": 6.559, "pct_cuda_time": 0.0504422793511734, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.559, "pct_cuda_time": 0.0504422793511734, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.919, "cuda_time_us": 277.115, "pct_cuda_time": 2.13116515359055, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.439, "cuda_time_us": 170.429, "pct_cuda_time": 1.3106917559904148, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.693, "pct_cuda_time": 1.3050315154655692, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.343, "cuda_time_us": 22.911, "pct_cuda_time": 0.1761980579683997, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.911, "pct_cuda_time": 0.1761980579683997, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.876, "cuda_time_us": 83.775, "pct_cuda_time": 0.6442753396317353, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.775, "pct_cuda_time": 0.6442753396317353, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2460.181, "cuda_time_us": 388.41100000000006, "pct_cuda_time": 2.987091959912885, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.014, "cuda_time_us": 6.144, "pct_cuda_time": 0.04725070351175627, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.144, "pct_cuda_time": 0.04725070351175627, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1745.679, "cuda_time_us": 99.102, "pct_cuda_time": 0.7621483104528108, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.005, "cuda_time_us": 44.287, "pct_cuda_time": 0.3405911306030517, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.287, "pct_cuda_time": 0.3405911306030517, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.575, "cuda_time_us": 7.488, "pct_cuda_time": 0.05758679490495295, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.488, "pct_cuda_time": 0.05758679490495295, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 782.965, "cuda_time_us": 16.448, "pct_cuda_time": 0.12649407085959752, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.028055105210105284, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.36, "pct_cuda_time": 0.08736458201392434, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011074383635567875, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 184.946, "cuda_time_us": 30.879, "pct_cuda_time": 0.23747631408520864, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.879, "pct_cuda_time": 0.23747631408520864, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.792, "cuda_time_us": 6.464, "pct_cuda_time": 0.04971167765299358, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.464, "pct_cuda_time": 0.04971167765299358, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 486.583, "cuda_time_us": 276.701, "pct_cuda_time": 2.1279812682953243, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.455, "cuda_time_us": 170.559, "pct_cuda_time": 1.3116915267352924, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.005667931069037169, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.822, "pct_cuda_time": 1.3060235956662554, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 111.746, "cuda_time_us": 22.815, "pct_cuda_time": 0.17545976572602853, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.815, "pct_cuda_time": 0.17545976572602853, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.196, "cuda_time_us": 83.327, "pct_cuda_time": 0.640829975834003, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.327, "pct_cuda_time": 0.640829975834003, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2517.81, "cuda_time_us": 394.647, "pct_cuda_time": 3.0350501934902465, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.986, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1761.574, "cuda_time_us": 97.98, "pct_cuda_time": 0.7535195198700976, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.232, "cuda_time_us": 43.935, "pct_cuda_time": 0.3378840590476907, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.935, "pct_cuda_time": 0.3378840590476907, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 525.782, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 753.159, "cuda_time_us": 16.445999999999998, "pct_cuda_time": 0.12647868977121476, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.02854730003835275, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.455, "pct_cuda_time": 0.08809518371210417, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.279, "pct_cuda_time": 0.009836206020757856, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.472, "cuda_time_us": 30.943, "pct_cuda_time": 0.23796850891345608, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.943, "pct_cuda_time": 0.23796850891345608, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 108.879, "cuda_time_us": 6.431, "pct_cuda_time": 0.04945788969467847, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.431, "pct_cuda_time": 0.04945788969467847, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 501.671, "cuda_time_us": 283.804, "pct_cuda_time": 2.1826072036866004, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 199.491, "cuda_time_us": 170.81300000000002, "pct_cuda_time": 1.3136449249598998, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.005652549980654436, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.078, "pct_cuda_time": 1.3079923749792453, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.745, "cuda_time_us": 22.848, "pct_cuda_time": 0.17571355368434363, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.848, "pct_cuda_time": 0.17571355368434363, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.061, "cuda_time_us": 90.143, "pct_cuda_time": 0.6932487250423576, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 90.143, "pct_cuda_time": 0.6932487250423576, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2787.222, "cuda_time_us": 394.172, "pct_cuda_time": 3.0313971849993475, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.117, "cuda_time_us": 6.561, "pct_cuda_time": 0.050457660439556136, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.561, "pct_cuda_time": 0.050457660439556136, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2058.669, "cuda_time_us": 99.135, "pct_cuda_time": 0.7624020984111259, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.785, "cuda_time_us": 44.128, "pct_cuda_time": 0.3393683340766244, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.128, "pct_cuda_time": 0.3393683340766244, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 485.186, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.656, "pct_cuda_time": 0.05118826213773596, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1066.225, "cuda_time_us": 16.672, "pct_cuda_time": 0.12821675275846361, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.028055105210105284, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.552, "pct_cuda_time": 0.08884116649866673, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011320481049691605, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 204.79, "cuda_time_us": 31.679, "pct_cuda_time": 0.24362874943830184, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.679, "pct_cuda_time": 0.24362874943830184, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.607, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 483.441, "cuda_time_us": 281.98, "pct_cuda_time": 2.1685796510815485, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.192, "cuda_time_us": 169.502, "pct_cuda_time": 1.303562621525018, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.005906337938969533, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.734, "pct_cuda_time": 1.2976562835860486, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 107.597, "cuda_time_us": 22.751, "pct_cuda_time": 0.17496757089778106, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.751, "pct_cuda_time": 0.17496757089778106, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.644, "cuda_time_us": 89.727, "pct_cuda_time": 0.6900494586587491, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 89.727, "pct_cuda_time": 0.6900494586587491, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2548.484, "cuda_time_us": 386.365, "pct_cuda_time": 2.9713571064973485, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.39, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1822.389, "cuda_time_us": 98.369, "pct_cuda_time": 0.7565111415605391, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.318, "cuda_time_us": 43.744, "pct_cuda_time": 0.33641516510713965, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.744, "pct_cuda_time": 0.33641516510713965, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 554.404, "cuda_time_us": 7.393, "pct_cuda_time": 0.056856193206773126, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.393, "pct_cuda_time": 0.056856193206773126, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 766.513, "cuda_time_us": 16.608, "pct_cuda_time": 0.12772455793021614, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011566578463815336, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 190.807, "cuda_time_us": 30.624, "pct_cuda_time": 0.23551522531641014, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.624, "pct_cuda_time": 0.23551522531641014, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.693, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 473.115, "cuda_time_us": 275.164, "pct_cuda_time": 2.1161609018731933, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.636, "cuda_time_us": 170.36499999999998, "pct_cuda_time": 1.310199561162167, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.629, "pct_cuda_time": 1.3045393206373215, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.518, "cuda_time_us": 22.656, "pct_cuda_time": 0.17423696919960122, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.656, "pct_cuda_time": 0.17423696919960122, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.24, "cuda_time_us": 82.143, "pct_cuda_time": 0.631724371511425, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.143, "pct_cuda_time": 0.631724371511425, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2476.937, "cuda_time_us": 396.02700000000004, "pct_cuda_time": 3.0456631444743327, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.83, "cuda_time_us": 6.528, "pct_cuda_time": 0.05020387248124103, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.528, "pct_cuda_time": 0.05020387248124103, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1711.036, "cuda_time_us": 98.143, "pct_cuda_time": 0.7547730785732902, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.163, "cuda_time_us": 43.807, "pct_cuda_time": 0.33689966939119576, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.807, "pct_cuda_time": 0.33689966939119576, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.288, "cuda_time_us": 6.593, "pct_cuda_time": 0.050703757853679865, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.593, "pct_cuda_time": 0.050703757853679865, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 748.919, "cuda_time_us": 16.352, "pct_cuda_time": 0.1257557786172263, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009597799150825492, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.428, "cuda_time_us": 31.391, "pct_cuda_time": 0.2414138727111883, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.391, "pct_cuda_time": 0.2414138727111883, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.296, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 525.611, "cuda_time_us": 284.956, "pct_cuda_time": 2.1914667105950554, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.908, "cuda_time_us": 169.661, "pct_cuda_time": 1.3047854180514453, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.005652549980654436, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.926, "pct_cuda_time": 1.2991328680707908, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 167.979, "cuda_time_us": 22.88, "pct_cuda_time": 0.17595965109846734, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.88, "pct_cuda_time": 0.17595965109846734, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.44, "cuda_time_us": 92.415, "pct_cuda_time": 0.7107216414451425, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 92.415, "pct_cuda_time": 0.7107216414451425, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2497.372, "cuda_time_us": 389.81800000000004, "pct_cuda_time": 2.997912555590138, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.612, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1782.066, "cuda_time_us": 98.04599999999999, "pct_cuda_time": 0.7540270957867276, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.46, "cuda_time_us": 44.288, "pct_cuda_time": 0.34059882114724305, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.288, "pct_cuda_time": 0.34059882114724305, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 492.19, "cuda_time_us": 6.463, "pct_cuda_time": 0.0497039871088022, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.463, "pct_cuda_time": 0.0497039871088022, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 792.261, "cuda_time_us": 16.608, "pct_cuda_time": 0.12772455793021614, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.02854730003835275, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.584, "pct_cuda_time": 0.08908726391279045, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.010089993979072953, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 193.195, "cuda_time_us": 30.687, "pct_cuda_time": 0.23599972960046625, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.687, "pct_cuda_time": 0.23599972960046625, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 90.342, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.803, "cuda_time_us": 278.74800000000005, "pct_cuda_time": 2.1437238122550517, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.687, "cuda_time_us": 170.11, "pct_cuda_time": 1.3082384723933689, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.005906337938969533, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.342, "pct_cuda_time": 1.3023321344543994, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.324, "cuda_time_us": 22.847, "pct_cuda_time": 0.17570586314015227, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.847, "pct_cuda_time": 0.17570586314015227, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.18, "cuda_time_us": 85.791, "pct_cuda_time": 0.6597794767215303, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 85.791, "pct_cuda_time": 0.6597794767215303, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2507.626, "cuda_time_us": 387.42, "pct_cuda_time": 2.9794706306192404, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.721, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1792.713, "cuda_time_us": 99.71000000000001, "pct_cuda_time": 0.7668241613211617, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.532, "cuda_time_us": 44.32, "pct_cuda_time": 0.34084491856136684, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.32, "pct_cuda_time": 0.34084491856136684, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.048, "cuda_time_us": 7.36, "pct_cuda_time": 0.05660240524845803, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.05660240524845803, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 790.578, "cuda_time_us": 16.671, "pct_cuda_time": 0.12820906221427225, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.584, "pct_cuda_time": 0.08908726391279045, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.01131279050550024, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.61, "cuda_time_us": 31.359, "pct_cuda_time": 0.2411677752970646, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.359, "pct_cuda_time": 0.2411677752970646, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.873, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.432, "pct_cuda_time": 0.04946558023886984, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.493, "cuda_time_us": 274.878, "pct_cuda_time": 2.1139614062344627, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.058, "cuda_time_us": 169.631, "pct_cuda_time": 1.3045547017257042, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.005667931069037169, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.894, "pct_cuda_time": 1.2988867706566674, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.588, "cuda_time_us": 23.2, "pct_cuda_time": 0.17842062523970464, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 23.2, "pct_cuda_time": 0.17842062523970464, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.041, "cuda_time_us": 82.047, "pct_cuda_time": 0.6309860792690538, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.047, "pct_cuda_time": 0.6309860792690538, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2421.56, "cuda_time_us": 394.522, "pct_cuda_time": 3.0340888754663258, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.19, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1731.9, "cuda_time_us": 98.11, "pct_cuda_time": 0.7545192906149751, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.052, "cuda_time_us": 43.68, "pct_cuda_time": 0.33592297027889223, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.68, "pct_cuda_time": 0.33592297027889223, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.197, "cuda_time_us": 6.72, "pct_cuda_time": 0.05168045696598341, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.72, "pct_cuda_time": 0.05168045696598341, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.517, "cuda_time_us": 16.671, "pct_cuda_time": 0.12820906221427225, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.028055105210105284, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.583, "pct_cuda_time": 0.08907957336859909, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011074383635567875, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.942, "cuda_time_us": 31.039, "pct_cuda_time": 0.2387068011558273, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.039, "pct_cuda_time": 0.2387068011558273, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.346, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.375, "cuda_time_us": 283.676, "pct_cuda_time": 2.181622814030106, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.407, "cuda_time_us": 169.117, "pct_cuda_time": 1.300601762011342, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.381, "pct_cuda_time": 1.294941521486496, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.313, "cuda_time_us": 22.688, "pct_cuda_time": 0.17448306661372495, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.688, "pct_cuda_time": 0.17448306661372495, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.518, "cuda_time_us": 91.871, "pct_cuda_time": 0.7065379854050391, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 91.871, "pct_cuda_time": 0.7065379854050391, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2432.28, "cuda_time_us": 394.62, "pct_cuda_time": 3.0348425487970796, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.914, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1748.922, "cuda_time_us": 98.143, "pct_cuda_time": 0.7547730785732902, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.428, "cuda_time_us": 43.839, "pct_cuda_time": 0.3371457668053195, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.3371457668053195, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 496.646, "cuda_time_us": 6.848, "pct_cuda_time": 0.05266484662247834, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.848, "pct_cuda_time": 0.05266484662247834, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 768.567, "cuda_time_us": 16.352, "pct_cuda_time": 0.1257557786172263, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.02854730003835275, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.328, "pct_cuda_time": 0.08711848459980061, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.010089993979072953, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 176.397, "cuda_time_us": 31.104, "pct_cuda_time": 0.2392066865282661, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.104, "pct_cuda_time": 0.2392066865282661, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.967, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.287, "cuda_time_us": 283.453, "pct_cuda_time": 2.179907822675431, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.036, "cuda_time_us": 171.327, "pct_cuda_time": 1.317597864674262, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.005667931069037169, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 170.59, "pct_cuda_time": 1.311929933605225, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.58, "cuda_time_us": 22.623, "pct_cuda_time": 0.17398318124128614, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.623, "pct_cuda_time": 0.17398318124128614, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.698, "cuda_time_us": 89.503, "pct_cuda_time": 0.6883267767598831, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 89.503, "pct_cuda_time": 0.6883267767598831, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2612.881, "cuda_time_us": 393.116, "pct_cuda_time": 3.0232759703332643, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.316, "cuda_time_us": 6.304, "pct_cuda_time": 0.04848119058237492, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.304, "pct_cuda_time": 0.04848119058237492, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1934.093, "cuda_time_us": 98.464, "pct_cuda_time": 0.7572417432587188, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.631, "cuda_time_us": 44.384, "pct_cuda_time": 0.3413371133896143, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.384, "pct_cuda_time": 0.3413371133896143, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 466.599, "cuda_time_us": 6.848, "pct_cuda_time": 0.05266484662247834, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.848, "pct_cuda_time": 0.05266484662247834, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 980.537, "cuda_time_us": 16.448, "pct_cuda_time": 0.12649407085959752, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.02780900779598155, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.552, "pct_cuda_time": 0.08884116649866673, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.511, "cuda_time_us": 30.784, "pct_cuda_time": 0.2367457123870288, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.784, "pct_cuda_time": 0.2367457123870288, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.25, "cuda_time_us": 6.433, "pct_cuda_time": 0.049473270783061206, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.433, "pct_cuda_time": 0.049473270783061206, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.652, "cuda_time_us": 281.915, "pct_cuda_time": 2.1680797657091095, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.413, "cuda_time_us": 168.989, "pct_cuda_time": 1.299617372354847, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.005652549980654436, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.254, "pct_cuda_time": 1.2939648223741924, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.821, "cuda_time_us": 22.592, "pct_cuda_time": 0.17374477437135377, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.592, "pct_cuda_time": 0.17374477437135377, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.403, "cuda_time_us": 90.334, "pct_cuda_time": 0.6947176189829086, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 90.334, "pct_cuda_time": 0.6947176189829086, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2557.763, "cuda_time_us": 391.16299999999995, "pct_cuda_time": 3.0082563375275253, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.334, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1739.918, "cuda_time_us": 98.527, "pct_cuda_time": 0.7577262475427751, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.94, "cuda_time_us": 43.776, "pct_cuda_time": 0.33666126252126344, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.776, "pct_cuda_time": 0.33666126252126344, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.445, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 752.855, "cuda_time_us": 17.184, "pct_cuda_time": 0.1321543113844433, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 4.0, "pct_cuda_time": 0.03076217676546632, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.552, "pct_cuda_time": 0.08884116649866673, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.012550968120310257, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.415, "cuda_time_us": 30.943, "pct_cuda_time": 0.23796850891345608, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.943, "pct_cuda_time": 0.23796850891345608, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 97.087, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.496, "pct_cuda_time": 0.04995777506711731, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 576.987, "cuda_time_us": 279.804, "pct_cuda_time": 2.1518450269211344, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 248.04, "cuda_time_us": 169.149, "pct_cuda_time": 1.3008478594254655, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.005906337938969533, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.381, "pct_cuda_time": 1.294941521486496, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 123.024, "cuda_time_us": 22.272, "pct_cuda_time": 0.17128380023011647, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.272, "pct_cuda_time": 0.17128380023011647, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.388, "cuda_time_us": 88.383, "pct_cuda_time": 0.6797133672655523, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 88.383, "pct_cuda_time": 0.6797133672655523, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3760.842, "cuda_time_us": 385.82099999999997, "pct_cuda_time": 2.9671734504572447, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.362, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2576.896, "cuda_time_us": 98.015, "pct_cuda_time": 0.7537886889167953, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 163.428, "cuda_time_us": 43.616, "pct_cuda_time": 0.3354307754506447, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.616, "pct_cuda_time": 0.3354307754506447, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 538.767, "cuda_time_us": 6.72, "pct_cuda_time": 0.05168045696598341, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.72, "pct_cuda_time": 0.05168045696598341, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1330.426, "cuda_time_us": 16.864, "pct_cuda_time": 0.129693337243206, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.872, "pct_cuda_time": 0.029777787108971398, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.52, "pct_cuda_time": 0.088595069084543, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011320481049691605, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 353.332, "cuda_time_us": 30.815, "pct_cuda_time": 0.2369841192569612, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.815, "pct_cuda_time": 0.2369841192569612, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 156.719, "cuda_time_us": 6.528, "pct_cuda_time": 0.05020387248124103, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.528, "pct_cuda_time": 0.05020387248124103, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 862.468, "cuda_time_us": 274.90999999999997, "pct_cuda_time": 2.1142075036485863, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 311.384, "cuda_time_us": 170.17399999999998, "pct_cuda_time": 1.3087306672216161, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.438, "pct_cuda_time": 1.3030704266967705, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 175.547, "cuda_time_us": 22.08, "pct_cuda_time": 0.16980721574537408, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.08, "pct_cuda_time": 0.16980721574537408, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 276.749, "cuda_time_us": 82.656, "pct_cuda_time": 0.635669620681596, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.656, "pct_cuda_time": 0.635669620681596, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4014.025, "cuda_time_us": 392.827, "pct_cuda_time": 3.0210534030619596, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.926, "cuda_time_us": 6.271, "pct_cuda_time": 0.04822740262405982, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.271, "pct_cuda_time": 0.04822740262405982, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2787.758, "cuda_time_us": 99.039, "pct_cuda_time": 0.7616638061687547, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 249.793, "cuda_time_us": 44.799, "pct_cuda_time": 0.3445286892290314, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.799, "pct_cuda_time": 0.3445286892290314, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 526.04, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 7.104, "pct_cuda_time": 0.05463362593546819, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1460.259, "cuda_time_us": 16.576, "pct_cuda_time": 0.12747846051609243, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.68, "pct_cuda_time": 0.028301202624229017, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.616, "pct_cuda_time": 0.08933336132691419, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 356.177, "cuda_time_us": 30.56, "pct_cuda_time": 0.2350230304881627, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.56, "pct_cuda_time": 0.2350230304881627, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 153.325, "cuda_time_us": 6.464, "pct_cuda_time": 0.04971167765299358, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.464, "pct_cuda_time": 0.04971167765299358, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 908.05, "cuda_time_us": 281.053, "pct_cuda_time": 2.1614505166161515, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 324.361, "cuda_time_us": 169.694, "pct_cuda_time": 1.3050392060097604, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 168.958, "pct_cuda_time": 1.2993789654849146, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 194.271, "cuda_time_us": 22.72, "pct_cuda_time": 0.1747291640278487, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.72, "pct_cuda_time": 0.1747291640278487, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 285.444, "cuda_time_us": 88.639, "pct_cuda_time": 0.6816821465785422, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 88.639, "pct_cuda_time": 0.6816821465785422, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2426.768, "cuda_time_us": 388.538, "pct_cuda_time": 2.9880686590251884, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.933, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.368, "pct_cuda_time": 0.04897338541062238, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1750.071, "cuda_time_us": 97.277, "pct_cuda_time": 0.7481130673035667, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.744, "cuda_time_us": 43.999, "pct_cuda_time": 0.33837625387593817, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.999, "pct_cuda_time": 0.33837625387593817, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 525.585, "cuda_time_us": 6.591, "pct_cuda_time": 0.05068837676529713, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.591, "pct_cuda_time": 0.05068837676529713, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.264, "cuda_time_us": 16.384, "pct_cuda_time": 0.12600187603135005, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.744, "pct_cuda_time": 0.028793397452476475, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.36, "pct_cuda_time": 0.08736458201392434, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.009843896564949222, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 171.353, "cuda_time_us": 30.303, "pct_cuda_time": 0.23304656063098148, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 30.303, "pct_cuda_time": 0.23304656063098148, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.27, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.336, "pct_cuda_time": 0.048727287996498655, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 443.661, "cuda_time_us": 278.557, "pct_cuda_time": 2.1422549183145003, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.141, "cuda_time_us": 170.01399999999998, "pct_cuda_time": 1.3075001801509976, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.278, "pct_cuda_time": 1.3018399396261517, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.058, "cuda_time_us": 23.264, "pct_cuda_time": 0.17891282006795212, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 23.264, "pct_cuda_time": 0.17891282006795212, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 134.334, "cuda_time_us": 85.279, "pct_cuda_time": 0.6558419180955505, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 85.279, "pct_cuda_time": 0.6558419180955505, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2410.681, "cuda_time_us": 389.30899999999997, "pct_cuda_time": 2.9939980685967313, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.991, "cuda_time_us": 6.272, "pct_cuda_time": 0.04823509316825119, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.272, "pct_cuda_time": 0.04823509316825119, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1697.865, "cuda_time_us": 99.423, "pct_cuda_time": 0.7646169751382395, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 134.371, "cuda_time_us": 44.607, "pct_cuda_time": 0.343052104744289, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.607, "pct_cuda_time": 0.343052104744289, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[512, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.012, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.624, "pct_cuda_time": 0.050942164723612224, "trace": "_C::rotary_embedding(int64[512], bfloat16[512, 4096], bfloat16[512, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 736.799, "cuda_time_us": 16.608, "pct_cuda_time": 0.12772455793021614, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.028055105210105284, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[512], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 11.488, "pct_cuda_time": 0.08834897167041926, "trace": "_vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.011320481049691605, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], None, None, bfloat16[512, 32, 128], int32[3], int32[3], None, None, None, 256, 256, None, None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[512, 32, 128], bfloat16[512, 8, 128], bfloat16[512, 8, 128], bfloat16[512, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.355, "cuda_time_us": 31.584, "pct_cuda_time": 0.24289814774012208, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 31.584, "pct_cuda_time": 0.24289814774012208, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[512, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 111.664, "cuda_time_us": 6.56, "pct_cuda_time": 0.05044996989536476, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.56, "pct_cuda_time": 0.05044996989536476, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.19, "cuda_time_us": 277.054, "pct_cuda_time": 2.1306960303948763, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.356, "cuda_time_us": 170.527, "pct_cuda_time": 1.3114454293211686, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.005667931069037169, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 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": 169.79, "pct_cuda_time": 1.3057774982521315, "trace": "mm(bfloat16[512, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[512, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[512, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.454, "cuda_time_us": 22.784, "pct_cuda_time": 0.17522135885609616, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 22.784, "pct_cuda_time": 0.17522135885609616, "trace": "_C::silu_and_mul(bfloat16[512, 14336], bfloat16[512, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.308, "cuda_time_us": 83.743, "pct_cuda_time": 0.6440292422176115, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.743, "pct_cuda_time": 0.6440292422176115, "trace": "mm(bfloat16[512, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[512, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[512, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.758, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 6.4, "pct_cuda_time": 0.049219482824746114, "trace": "_C::fused_add_rms_norm(bfloat16[512, 4096], bfloat16[512, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 426.736, "cuda_time_us": 363.099, "pct_cuda_time": 2.7924289053410134, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 2.976, "pct_cuda_time": 0.022887059513506943, "trace": "index_select(bfloat16[512, 4096], 0, int64[2])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.005660240524845803, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 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": 359.387, "pct_cuda_time": 2.763881605302661, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3178.254, "cuda_time_us": 114.36500000000001, "pct_cuda_time": 0.8795290864456389, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.005906337938969533, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.005906337938969533, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.006152435353093264, "trace": "copy_(int32[2], int32[2], True) <- _to_copy(int32[2], 3, 0, None, None, True, None) <- to(int32[2], 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": 0.005906337938969533, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.006152435353093264, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.005898647394778167, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.005898647394778167, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 4.064, "pct_cuda_time": 0.031254371593713776, "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.64, "pct_cuda_time": 0.035684125047940926, "trace": "div_(float32[2, 128256], bfloat16[2, 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": 34.303, "pct_cuda_time": 0.2638087373964478, "trace": "_softmax(float32[2, 128256], -1, False) <- softmax(float32[2, 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": 28.031, "pct_cuda_time": 0.21557364422819658, "trace": "_log_softmax(float32[2, 128256], -1, False) <- log_softmax(float32[2, 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": 0.014027552605052642, "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.608, "pct_cuda_time": 0.0354380276338172, "trace": "index(float32[2, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 28.832, "pct_cuda_time": 0.22173377012548123, "trace": "argmax(float32[2, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.625, "pct_cuda_time": 0.020187678502337273, "trace": "copy_(int64[2], int64[2], False) <- _to_copy(int64[2], 4, 0, None, None, False, None) <- to(int64[2], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 2 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6368.478000000001, "pct_cuda_time": 93.30209237092885, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 3.68, "pct_cuda_time": 0.05391424763106948, "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": 3.68, "pct_cuda_time": 0.05391424763106948, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6361.502, "pct_cuda_time": 93.19988971020211, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 211.38800000000006, "pct_cuda_time": 3.0969633093034017, "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": 0.06188417988957541, "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": 207.16400000000004, "pct_cuda_time": 3.035079129413826, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 1868.4150000000002, "pct_cuda_time": 27.373420920544756, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 666.838, "pct_cuda_time": 9.769583984186713, "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": 666.838, "pct_cuda_time": 9.769583984186713, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 115.55200000000005, "pct_cuda_time": 1.6929073756155826, "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": 115.55200000000005, "pct_cuda_time": 1.6929073756155826, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 523.89, "pct_cuda_time": 7.675308475935052, "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": 73.18299999999999, "pct_cuda_time": 1.0721756479305864, "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": 403.51000000000005, "pct_cuda_time": 5.911667951525231, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 47.196999999999996, "pct_cuda_time": 0.6914648764792354, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 562.1349999999999, "pct_cuda_time": 8.235621084807402, "invocations": 32 }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cuda_time_us": 494.58399999999983, "pct_cuda_time": 7.245957676729581, "invocations": 32 }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cuda_time_us": 67.55100000000002, "pct_cuda_time": 0.9896634080778195, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4281.6990000000005, "pct_cuda_time": 62.729505480353964, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2614.4259999999995, "pct_cuda_time": 38.30293771117022, "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": 2614.4259999999995, "pct_cuda_time": 38.30293771117022, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 276.86299999999994, "pct_cuda_time": 4.0562120494241265, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 276.86299999999994, "pct_cuda_time": 4.0562120494241265, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1390.4099999999999, "pct_cuda_time": 20.370355719759594, "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": 1390.4099999999999, "pct_cuda_time": 20.370355719759594, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.296, "pct_cuda_time": 0.04828841309565353, "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": 3.296, "pct_cuda_time": 0.04828841309565353, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 342.46, "pct_cuda_time": 5.01724816405871, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 2.977, "pct_cuda_time": 0.04361486826024289, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.768, "pct_cuda_time": 0.011251669070831891, "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": 338.715, "pct_cuda_time": 4.962381626727636, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 114.71600000000001, "pct_cuda_time": 1.6806594650124365, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.536, "pct_cuda_time": 0.08110578121891321, "invocations": 7 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 4.096, "pct_cuda_time": 0.06000890171110342, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 4.735, "pct_cuda_time": 0.06937064199269402, "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": 34.655, "pct_cuda_time": 0.5077169162105198, "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": 28.191, "pct_cuda_time": 0.41301536819768464, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cuda_time_us": 1.952, "pct_cuda_time": 0.028597992221697722, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cuda_time_us": 4.672, "pct_cuda_time": 0.06844765351422734, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cuda_time_us": 28.384, "pct_cuda_time": 0.415842936076162, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.495, "pct_cuda_time": 0.03655327386943434, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 81264.749, "cuda_time_us": 6368.478000000001, "pct_cuda_time": 93.30209237092885, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 311.542, "cuda_time_us": 3.68, "pct_cuda_time": 0.05391424763106948, "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": 3.68, "pct_cuda_time": 0.05391424763106948, "trace": "index_select(bfloat16[128256, 4096], 0, int64[2]) <- embedding(bfloat16[128256, 4096], int64[2], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4278.972, "cuda_time_us": 203.998, "pct_cuda_time": 2.9886952957181823, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 278.836, "cuda_time_us": 4.224, "pct_cuda_time": 0.06188417988957541, "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": 0.06188417988957541, "trace": "_C::rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3121.433, "cuda_time_us": 61.792, "pct_cuda_time": 0.9052905406573494, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 541.161, "cuda_time_us": 23.872, "pct_cuda_time": 0.34973938028502466, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 23.872, "pct_cuda_time": 0.34973938028502466, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 943.349, "cuda_time_us": 3.456, "pct_cuda_time": 0.05063251081874351, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.456, "pct_cuda_time": 0.05063251081874351, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1089.964, "cuda_time_us": 16.224, "pct_cuda_time": 0.23769150912132372, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.144, "pct_cuda_time": 0.0314109094894057, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.608, "pct_cuda_time": 0.18471490057949022, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 288.272, "cuda_time_us": 18.240000000000002, "pct_cuda_time": 0.2672271404322574, "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": 0.23628505048746973, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 124.216, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 624.902, "cuda_time_us": 134.718, "pct_cuda_time": 1.9737009816202224, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 236.129, "cuda_time_us": 82.047, "pct_cuda_time": 1.2020386617897711, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.047, "pct_cuda_time": 1.2020386617897711, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 141.692, "cuda_time_us": 8.544, "pct_cuda_time": 0.1251748184130048, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.544, "pct_cuda_time": 0.1251748184130048, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 168.2, "cuda_time_us": 44.127, "pct_cuda_time": 0.6464875014174465, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6464875014174465, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2566.064, "cuda_time_us": 200.22, "pct_cuda_time": 2.9333452882317204, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.342, "cuda_time_us": 3.168, "pct_cuda_time": 0.046413134917181555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046413134917181555, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1846.39, "cuda_time_us": 59.134, "pct_cuda_time": 0.866349217232517, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.825, "cuda_time_us": 21.311, "pct_cuda_time": 0.31221916610481565, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.311, "pct_cuda_time": 0.31221916610481565, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 556.663, "cuda_time_us": 3.776, "pct_cuda_time": 0.05532070626492346, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05532070626492346, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 803.553, "cuda_time_us": 16.447, "pct_cuda_time": 0.2409585953228804, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.4, "pct_cuda_time": 0.03516146584634966, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.607, "pct_cuda_time": 0.18470024996872092, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.762, "cuda_time_us": 17.6, "pct_cuda_time": 0.25785074953989756, "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": 15.488, "pct_cuda_time": 0.2269086595951098, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.736, "cuda_time_us": 3.455, "pct_cuda_time": 0.0506178602079742, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.455, "pct_cuda_time": 0.0506178602079742, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.947, "cuda_time_us": 134.463, "pct_cuda_time": 1.9699650758740477, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.491, "cuda_time_us": 81.695, "pct_cuda_time": 1.1968816467989731, "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": 1.1968816467989731, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.317, "cuda_time_us": 8.576, "pct_cuda_time": 0.1256436379576228, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.1256436379576228, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.14, "cuda_time_us": 44.192, "pct_cuda_time": 0.6474397911174518, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.192, "pct_cuda_time": 0.6474397911174518, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2375.803, "cuda_time_us": 198.91000000000003, "pct_cuda_time": 2.9141529881239214, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.138, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1678.171, "cuda_time_us": 58.71900000000001, "pct_cuda_time": 0.8602692137632526, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.375, "cuda_time_us": 21.311, "pct_cuda_time": 0.31221916610481565, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.311, "pct_cuda_time": 0.31221916610481565, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 501.69, "cuda_time_us": 3.744, "pct_cuda_time": 0.05485188672030547, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05485188672030547, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 720.715, "cuda_time_us": 16.321, "pct_cuda_time": 0.23911261836594705, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.305, "pct_cuda_time": 0.033769657823264985, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 158.53, "cuda_time_us": 17.343, "pct_cuda_time": 0.25408554257218424, "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": 15.263, "pct_cuda_time": 0.22361227217201451, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.03047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.442, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.412, "cuda_time_us": 133.727, "pct_cuda_time": 1.9591822263478338, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.177, "cuda_time_us": 80.735, "pct_cuda_time": 1.1828170604604333, "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": 80.735, "pct_cuda_time": 1.1828170604604333, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.882, "cuda_time_us": 8.832, "pct_cuda_time": 0.12939419431456678, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.832, "pct_cuda_time": 0.12939419431456678, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.041, "cuda_time_us": 44.16, "pct_cuda_time": 0.6469709715728338, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.16, "pct_cuda_time": 0.6469709715728338, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2482.409, "cuda_time_us": 198.142, "pct_cuda_time": 2.9029013190530892, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.533, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1753.978, "cuda_time_us": 58.112, "pct_cuda_time": 0.8513762930262798, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.368, "cuda_time_us": 20.544, "pct_cuda_time": 0.3009821476447531, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3009821476447531, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 527.937, "cuda_time_us": 3.776, "pct_cuda_time": 0.05532070626492346, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05532070626492346, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 753.399, "cuda_time_us": 16.288, "pct_cuda_time": 0.23862914821055972, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.576, "pct_cuda_time": 0.18424608103487222, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.811, "cuda_time_us": 17.503999999999998, "pct_cuda_time": 0.25644429090604354, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.22550220096125584, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 91.914, "cuda_time_us": 3.392, "pct_cuda_time": 0.04969487172950752, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.392, "pct_cuda_time": 0.04969487172950752, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.11, "cuda_time_us": 133.374, "pct_cuda_time": 1.9540105607462666, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 181.782, "cuda_time_us": 81.535, "pct_cuda_time": 1.194537549075883, "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": 1.194537549075883, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.517, "cuda_time_us": 8.704, "pct_cuda_time": 0.12751891613609478, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.704, "pct_cuda_time": 0.12751891613609478, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.178, "cuda_time_us": 43.135, "pct_cuda_time": 0.6319540955342886, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.135, "pct_cuda_time": 0.6319540955342886, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2423.29, "cuda_time_us": 199.325, "pct_cuda_time": 2.9202329915931857, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.641, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1685.062, "cuda_time_us": 58.399, "pct_cuda_time": 0.8555810183170723, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.796, "cuda_time_us": 20.447, "pct_cuda_time": 0.29956103840012976, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.447, "pct_cuda_time": 0.29956103840012976, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 510.388, "cuda_time_us": 3.712, "pct_cuda_time": 0.05438306717568748, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.712, "pct_cuda_time": 0.05438306717568748, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 717.621, "cuda_time_us": 16.64, "pct_cuda_time": 0.24378616320135768, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.464, "pct_cuda_time": 0.03609910493558566, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.672, "pct_cuda_time": 0.18565253966872622, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 158.068, "cuda_time_us": 17.6, "pct_cuda_time": 0.25785074953989756, "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": 15.488, "pct_cuda_time": 0.2269086595951098, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.725, "cuda_time_us": 3.328, "pct_cuda_time": 0.04875723264027153, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04875723264027153, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 493.902, "cuda_time_us": 134.334, "pct_cuda_time": 1.9680751470848066, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 189.386, "cuda_time_us": 81.311, "pct_cuda_time": 1.1912558122635575, "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": 1.1912558122635575, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.901, "cuda_time_us": 8.864, "pct_cuda_time": 0.12986301385918478, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.864, "pct_cuda_time": 0.12986301385918478, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.103, "cuda_time_us": 44.159, "pct_cuda_time": 0.6469563209620645, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.159, "pct_cuda_time": 0.6469563209620645, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2538.877, "cuda_time_us": 197.276, "pct_cuda_time": 2.8902138901268652, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.101, "cuda_time_us": 3.199, "pct_cuda_time": 0.04686730385103023, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.199, "pct_cuda_time": 0.04686730385103023, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1801.884, "cuda_time_us": 58.20700000000001, "pct_cuda_time": 0.8527681010493646, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 141.582, "cuda_time_us": 20.32, "pct_cuda_time": 0.29770041083242715, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.32, "pct_cuda_time": 0.29770041083242715, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 480.34, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 841.897, "cuda_time_us": 16.254, "pct_cuda_time": 0.23813102744440312, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.207, "pct_cuda_time": 0.03233389796787237, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.607, "pct_cuda_time": 0.18470024996872092, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 165.556, "cuda_time_us": 18.081, "pct_cuda_time": 0.26489769331993673, "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": 15.969, "pct_cuda_time": 0.23395560337514903, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.299, "cuda_time_us": 3.424, "pct_cuda_time": 0.05016369127412552, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.424, "pct_cuda_time": 0.05016369127412552, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.005, "cuda_time_us": 132.446, "pct_cuda_time": 1.9404147939523446, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.366, "cuda_time_us": 81.183, "pct_cuda_time": 1.1893805340850854, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.183, "pct_cuda_time": 1.1893805340850854, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.749, "cuda_time_us": 8.416, "pct_cuda_time": 0.12329954023453282, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.12329954023453282, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.196, "cuda_time_us": 42.847, "pct_cuda_time": 0.6277347196327266, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.847, "pct_cuda_time": 0.6277347196327266, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2433.812, "cuda_time_us": 200.99099999999999, "pct_cuda_time": 2.9446409091348604, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.29, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1737.871, "cuda_time_us": 59.040000000000006, "pct_cuda_time": 0.8649720598202018, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.509, "cuda_time_us": 21.248, "pct_cuda_time": 0.31129617762634904, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.248, "pct_cuda_time": 0.31129617762634904, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 509.348, "cuda_time_us": 3.744, "pct_cuda_time": 0.05485188672030547, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05485188672030547, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 760.002, "cuda_time_us": 16.384, "pct_cuda_time": 0.2400356068444137, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.368, "pct_cuda_time": 0.03469264630173167, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 161.283, "cuda_time_us": 17.664, "pct_cuda_time": 0.25878838862913356, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.552, "pct_cuda_time": 0.22784629868434578, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.97, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.771, "cuda_time_us": 135.487, "pct_cuda_time": 1.9849673013018234, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.137, "cuda_time_us": 82.879, "pct_cuda_time": 1.214227969949839, "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": 82.879, "pct_cuda_time": 1.214227969949839, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.353, "cuda_time_us": 8.672, "pct_cuda_time": 0.12705009659147679, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.672, "pct_cuda_time": 0.12705009659147679, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.745, "cuda_time_us": 43.936, "pct_cuda_time": 0.6436892347605078, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6436892347605078, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2281.922, "cuda_time_us": 198.077, "pct_cuda_time": 2.901949029353084, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.61, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1619.797, "cuda_time_us": 57.759, "pct_cuda_time": 0.8462046274247126, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.062, "cuda_time_us": 20.575, "pct_cuda_time": 0.30143631657860176, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30143631657860176, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.307, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 694.395, "cuda_time_us": 16.416, "pct_cuda_time": 0.2405044263890317, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.736, "pct_cuda_time": 0.18659017875796222, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 155.696, "cuda_time_us": 17.216, "pct_cuda_time": 0.2522249150044816, "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": 15.104, "pct_cuda_time": 0.22128282505969385, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.49, "cuda_time_us": 3.36, "pct_cuda_time": 0.04922605218488953, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04922605218488953, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 444.497, "cuda_time_us": 133.758, "pct_cuda_time": 1.9596363952816827, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.507, "cuda_time_us": 82.015, "pct_cuda_time": 1.2015698422451533, "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": 82.015, "pct_cuda_time": 1.2015698422451533, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.772, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.251, "cuda_time_us": 43.103, "pct_cuda_time": 0.6314852759896706, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6314852759896706, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2303.193, "cuda_time_us": 198.078, "pct_cuda_time": 2.9019636799638535, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.147, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1613.754, "cuda_time_us": 58.305, "pct_cuda_time": 0.8542038609047571, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 156.77, "cuda_time_us": 20.832, "pct_cuda_time": 0.3052015235463151, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.832, "pct_cuda_time": 0.3052015235463151, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 460.778, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 696.306, "cuda_time_us": 16.289, "pct_cuda_time": 0.23864379882132902, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.273, "pct_cuda_time": 0.033300838278646994, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 153.947, "cuda_time_us": 17.631999999999998, "pct_cuda_time": 0.2583195690845155, "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": 0.2273774791397278, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.517, "cuda_time_us": 3.295, "pct_cuda_time": 0.04827376248488422, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.295, "pct_cuda_time": 0.04827376248488422, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.486, "cuda_time_us": 133.246, "pct_cuda_time": 1.952135282567795, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.181, "cuda_time_us": 80.799, "pct_cuda_time": 1.1837546995496695, "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": 1.1837546995496695, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.183, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.065, "cuda_time_us": 43.807, "pct_cuda_time": 0.6417993059712666, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.807, "pct_cuda_time": 0.6417993059712666, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2460.95, "cuda_time_us": 198.14100000000002, "pct_cuda_time": 2.9028866684423202, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.64, "cuda_time_us": 3.296, "pct_cuda_time": 0.04828841309565353, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04828841309565353, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1787.537, "cuda_time_us": 58.144, "pct_cuda_time": 0.8518451125708978, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 209.714, "cuda_time_us": 20.608, "pct_cuda_time": 0.30191978673398906, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30191978673398906, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 514.918, "cuda_time_us": 3.776, "pct_cuda_time": 0.05532070626492346, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05532070626492346, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 725.954, "cuda_time_us": 16.256, "pct_cuda_time": 0.2381603286659417, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.94, "cuda_time_us": 17.503999999999998, "pct_cuda_time": 0.25644429090604354, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.22550220096125584, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.002, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.527, "cuda_time_us": 133.437, "pct_cuda_time": 1.9549335492247335, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.563, "cuda_time_us": 81.086, "pct_cuda_time": 1.187959424840462, "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": 81.086, "pct_cuda_time": 1.187959424840462, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.178, "cuda_time_us": 8.544, "pct_cuda_time": 0.1251748184130048, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.544, "pct_cuda_time": 0.1251748184130048, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.035, "cuda_time_us": 43.807, "pct_cuda_time": 0.6417993059712666, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.807, "pct_cuda_time": 0.6417993059712666, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2265.557, "cuda_time_us": 199.484, "pct_cuda_time": 2.9225624387055067, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.282, "cuda_time_us": 3.456, "pct_cuda_time": 0.05063251081874351, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.456, "pct_cuda_time": 0.05063251081874351, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1606.966, "cuda_time_us": 58.271, "pct_cuda_time": 0.8537057401386005, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.868, "cuda_time_us": 20.832, "pct_cuda_time": 0.3052015235463151, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.832, "pct_cuda_time": 0.3052015235463151, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.278, "cuda_time_us": 3.584, "pct_cuda_time": 0.0525077889972155, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.0525077889972155, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 692.104, "cuda_time_us": 16.351000000000003, "pct_cuda_time": 0.23955213668902642, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.303, "pct_cuda_time": 0.03374035660172636, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.576, "pct_cuda_time": 0.18424608103487222, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 146.211, "cuda_time_us": 17.503999999999998, "pct_cuda_time": 0.25644429090604354, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.22550220096125584, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.798, "cuda_time_us": 3.488, "pct_cuda_time": 0.05110133036336151, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.488, "pct_cuda_time": 0.05110133036336151, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 433.491, "cuda_time_us": 134.269, "pct_cuda_time": 1.9671228573848014, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.356, "cuda_time_us": 81.599, "pct_cuda_time": 1.1954751881651193, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.599, "pct_cuda_time": 1.1954751881651193, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.381, "cuda_time_us": 8.543, "pct_cuda_time": 0.1251601678022355, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.543, "pct_cuda_time": 0.1251601678022355, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 134.02, "cuda_time_us": 44.127, "pct_cuda_time": 0.6464875014174465, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6464875014174465, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2435.907, "cuda_time_us": 198.87800000000004, "pct_cuda_time": 2.9136841685793042, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.31, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1725.429, "cuda_time_us": 58.367, "pct_cuda_time": 0.8551121987724545, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 132.185, "cuda_time_us": 20.768, "pct_cuda_time": 0.3042638844570791, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.768, "pct_cuda_time": 0.3042638844570791, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 465.251, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 727.984, "cuda_time_us": 16.543, "pct_cuda_time": 0.24236505395673433, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.304, "pct_cuda_time": 0.03375500721249567, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.799, "pct_cuda_time": 0.18751316723642888, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 166.09, "cuda_time_us": 17.503999999999998, "pct_cuda_time": 0.25644429090604354, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.392, "pct_cuda_time": 0.22550220096125584, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.183, "cuda_time_us": 3.36, "pct_cuda_time": 0.04922605218488953, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04922605218488953, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.527, "cuda_time_us": 133.95100000000002, "pct_cuda_time": 1.96246396316016, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.294, "cuda_time_us": 81.855, "pct_cuda_time": 1.1992257445220633, "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": 1.1992257445220633, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.743, "cuda_time_us": 8.832, "pct_cuda_time": 0.12939419431456678, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.832, "pct_cuda_time": 0.12939419431456678, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.088, "cuda_time_us": 43.264, "pct_cuda_time": 0.6338440243235299, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6338440243235299, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2566.094, "cuda_time_us": 197.724, "pct_cuda_time": 2.8967773637515166, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.366, "cuda_time_us": 3.456, "pct_cuda_time": 0.05063251081874351, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.456, "pct_cuda_time": 0.05063251081874351, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1841.722, "cuda_time_us": 58.141999999999996, "pct_cuda_time": 0.8518158113493591, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.858, "cuda_time_us": 20.448, "pct_cuda_time": 0.29957568901089915, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.448, "pct_cuda_time": 0.29957568901089915, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 498.178, "cuda_time_us": 3.616, "pct_cuda_time": 0.052976608541833496, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.052976608541833496, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 892.29, "cuda_time_us": 16.351, "pct_cuda_time": 0.23955213668902636, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.575, "pct_cuda_time": 0.18423143042410292, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.471, "cuda_time_us": 17.727, "pct_cuda_time": 0.25971137710760017, "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": 15.647, "pct_cuda_time": 0.2292381067074305, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.03047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.91, "cuda_time_us": 3.168, "pct_cuda_time": 0.046413134917181555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046413134917181555, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 498.284, "cuda_time_us": 132.958, "pct_cuda_time": 1.9479159066662324, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.717, "cuda_time_us": 81.759, "pct_cuda_time": 1.197819285888209, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.759, "pct_cuda_time": 1.197819285888209, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.896, "cuda_time_us": 8.512, "pct_cuda_time": 0.1247059988683868, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.512, "pct_cuda_time": 0.1247059988683868, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.159, "cuda_time_us": 42.687, "pct_cuda_time": 0.6253906219096366, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.687, "pct_cuda_time": 0.6253906219096366, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2274.4, "cuda_time_us": 197.887, "pct_cuda_time": 2.899165413306915, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.813, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1627.721, "cuda_time_us": 58.144, "pct_cuda_time": 0.8518451125708978, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.596, "cuda_time_us": 20.512, "pct_cuda_time": 0.30051332810013515, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30051332810013515, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 500.933, "cuda_time_us": 3.488, "pct_cuda_time": 0.05110133036336151, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.488, "pct_cuda_time": 0.05110133036336151, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 687.077, "cuda_time_us": 16.32, "pct_cuda_time": 0.2390979677551777, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.608, "pct_cuda_time": 0.18471490057949022, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 151.39, "cuda_time_us": 17.823999999999998, "pct_cuda_time": 0.26113248635222347, "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": 0.2306592159520538, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.03047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.71, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 434.457, "cuda_time_us": 133.279, "pct_cuda_time": 1.9526187527231817, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.308, "cuda_time_us": 81.087, "pct_cuda_time": 1.1879740754512313, "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": 1.1879740754512313, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.078, "cuda_time_us": 8.576, "pct_cuda_time": 0.1256436379576228, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.576, "pct_cuda_time": 0.1256436379576228, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.001, "cuda_time_us": 43.616, "pct_cuda_time": 0.6390010393143278, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6390010393143278, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2388.128, "cuda_time_us": 199.003, "pct_cuda_time": 2.915515494925467, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.988, "cuda_time_us": 3.231, "pct_cuda_time": 0.04733612339564823, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.231, "pct_cuda_time": 0.04733612339564823, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1690.935, "cuda_time_us": 58.74999999999999, "pct_cuda_time": 0.8607233826971009, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.307, "cuda_time_us": 21.215, "pct_cuda_time": 0.3108127074709617, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.215, "pct_cuda_time": 0.3108127074709617, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 464.362, "cuda_time_us": 3.551, "pct_cuda_time": 0.05202431884182819, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.551, "pct_cuda_time": 0.05202431884182819, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.007, "cuda_time_us": 16.352, "pct_cuda_time": 0.23956678729979572, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.336, "pct_cuda_time": 0.03422382675711367, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.029, "cuda_time_us": 17.631999999999998, "pct_cuda_time": 0.2583195690845155, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.22643984005049184, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.031879729034023696, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.764, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.494, "cuda_time_us": 133.75799999999998, "pct_cuda_time": 1.9596363952816824, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 171.255, "cuda_time_us": 81.567, "pct_cuda_time": 1.1950063686205012, "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": 1.1950063686205012, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.937, "cuda_time_us": 8.639, "pct_cuda_time": 0.12656662643608946, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.639, "pct_cuda_time": 0.12656662643608946, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.507, "cuda_time_us": 43.552, "pct_cuda_time": 0.6380634002250919, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.552, "pct_cuda_time": 0.6380634002250919, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2289.703, "cuda_time_us": 199.07100000000003, "pct_cuda_time": 2.916511736457781, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.336, "cuda_time_us": 3.328, "pct_cuda_time": 0.04875723264027153, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04875723264027153, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1625.988, "cuda_time_us": 58.145, "pct_cuda_time": 0.8518597631816671, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.176, "cuda_time_us": 20.8, "pct_cuda_time": 0.3047327040016971, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.8, "pct_cuda_time": 0.3047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.409, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 678.79, "cuda_time_us": 16.256, "pct_cuda_time": 0.2381603286659417, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.544, "pct_cuda_time": 0.18377726149025425, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 150.272, "cuda_time_us": 17.537, "pct_cuda_time": 0.25692776106143084, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.2259710205058738, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030956740555557014, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.878, "cuda_time_us": 3.424, "pct_cuda_time": 0.05016369127412552, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.424, "pct_cuda_time": 0.05016369127412552, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 441.179, "cuda_time_us": 134.174, "pct_cuda_time": 1.9657310493617166, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 152.476, "cuda_time_us": 81.982, "pct_cuda_time": 1.2010863720897658, "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": 81.982, "pct_cuda_time": 1.2010863720897658, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.8, "cuda_time_us": 8.928, "pct_cuda_time": 0.13080065294842075, "trace": "" }, "children": [ { "entry": { "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": 0.13080065294842075, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.791, "cuda_time_us": 43.264, "pct_cuda_time": 0.6338440243235299, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6338440243235299, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2338.028, "cuda_time_us": 199.839, "pct_cuda_time": 2.927763405528612, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.681, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1675.267, "cuda_time_us": 58.432, "pct_cuda_time": 0.8560644884724599, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.867, "cuda_time_us": 20.704, "pct_cuda_time": 0.3033262453678431, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.704, "pct_cuda_time": 0.3033262453678431, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.063, "cuda_time_us": 3.744, "pct_cuda_time": 0.05485188672030547, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05485188672030547, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 708.535, "cuda_time_us": 16.512, "pct_cuda_time": 0.24191088502288566, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.304, "pct_cuda_time": 0.03375500721249567, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.736, "pct_cuda_time": 0.18659017875796222, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 155.839, "cuda_time_us": 17.472, "pct_cuda_time": 0.2559754713614255, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.36, "pct_cuda_time": 0.2250333814166378, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.63, "cuda_time_us": 3.393, "pct_cuda_time": 0.049709522340276834, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.393, "pct_cuda_time": 0.049709522340276834, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.313, "cuda_time_us": 134.78199999999998, "pct_cuda_time": 1.9746386207094582, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.978, "cuda_time_us": 83.231, "pct_cuda_time": 1.2193849849406369, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.231, "pct_cuda_time": 1.2193849849406369, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.062, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 138.737, "cuda_time_us": 42.911, "pct_cuda_time": 0.6286723587219626, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.911, "pct_cuda_time": 0.6286723587219626, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2342.749, "cuda_time_us": 197.757, "pct_cuda_time": 2.8972608339069046, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.562, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1639.587, "cuda_time_us": 57.887, "pct_cuda_time": 0.8480799056031845, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.52, "cuda_time_us": 20.576, "pct_cuda_time": 0.3014509671893711, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.576, "pct_cuda_time": 0.3014509671893711, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 487.934, "cuda_time_us": 3.648, "pct_cuda_time": 0.05344542808645149, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.05344542808645149, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 702.969, "cuda_time_us": 16.352, "pct_cuda_time": 0.23956678729979572, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.64, "pct_cuda_time": 0.18518372012410822, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 161.996, "cuda_time_us": 17.311, "pct_cuda_time": 0.25361672302756627, "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": 15.199, "pct_cuda_time": 0.22267463308277852, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.459, "cuda_time_us": 3.328, "pct_cuda_time": 0.04875723264027153, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04875723264027153, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.939, "cuda_time_us": 133.278, "pct_cuda_time": 1.9526041021124123, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.948, "cuda_time_us": 80.735, "pct_cuda_time": 1.1828170604604333, "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": 80.735, "pct_cuda_time": 1.1828170604604333, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.612, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.586, "cuda_time_us": 43.903, "pct_cuda_time": 0.6432057646051205, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.903, "pct_cuda_time": 0.6432057646051205, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2331.283, "cuda_time_us": 198.75, "pct_cuda_time": 2.9118088904008315, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.823, "cuda_time_us": 3.168, "pct_cuda_time": 0.046413134917181555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046413134917181555, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1640.806, "cuda_time_us": 58.495999999999995, "pct_cuda_time": 0.8570021275616957, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.463, "cuda_time_us": 20.8, "pct_cuda_time": 0.3047327040016971, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.8, "pct_cuda_time": 0.3047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.623, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 711.036, "cuda_time_us": 16.543, "pct_cuda_time": 0.24236505395673433, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.496, "pct_cuda_time": 0.03656792448020365, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.575, "pct_cuda_time": 0.18423143042410292, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 157.092, "cuda_time_us": 17.601, "pct_cuda_time": 0.25786540015066683, "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": 15.488, "pct_cuda_time": 0.2269086595951098, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030956740555557014, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.992, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 452.119, "cuda_time_us": 133.886, "pct_cuda_time": 1.9615116734601545, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 155.907, "cuda_time_us": 81.886, "pct_cuda_time": 1.1996799134559117, "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": 81.886, "pct_cuda_time": 1.1996799134559117, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.91, "cuda_time_us": 8.768, "pct_cuda_time": 0.12845655522533075, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.768, "pct_cuda_time": 0.12845655522533075, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 147.679, "cuda_time_us": 43.232, "pct_cuda_time": 0.6333752047789118, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.232, "pct_cuda_time": 0.6333752047789118, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2590.614, "cuda_time_us": 199.165, "pct_cuda_time": 2.9178888938700958, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.838, "cuda_time_us": 3.328, "pct_cuda_time": 0.04875723264027153, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04875723264027153, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1675.313, "cuda_time_us": 57.854, "pct_cuda_time": 0.8475964354477972, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 132.9, "cuda_time_us": 20.64, "pct_cuda_time": 0.3023886062786071, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3023886062786071, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 449.716, "cuda_time_us": 3.584, "pct_cuda_time": 0.0525077889972155, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.0525077889972155, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 760.453, "cuda_time_us": 16.159, "pct_cuda_time": 0.2367392194213184, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.447, "pct_cuda_time": 0.18235615224563093, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 180.42, "cuda_time_us": 17.471, "pct_cuda_time": 0.25596082075065624, "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": 15.359, "pct_cuda_time": 0.2250187308058685, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.845, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 678.917, "cuda_time_us": 134.719, "pct_cuda_time": 1.9737156322309917, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.491, "cuda_time_us": 81.695, "pct_cuda_time": 1.1968816467989731, "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": 1.1968816467989731, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.728, "cuda_time_us": 8.704, "pct_cuda_time": 0.12751891613609478, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.704, "pct_cuda_time": 0.12751891613609478, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 355.25, "cuda_time_us": 44.32, "pct_cuda_time": 0.6493150692959238, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6493150692959238, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2350.525, "cuda_time_us": 198.845, "pct_cuda_time": 2.9132006984239163, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.718, "cuda_time_us": 3.168, "pct_cuda_time": 0.046413134917181555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046413134917181555, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1673.795, "cuda_time_us": 58.111000000000004, "pct_cuda_time": 0.8513616424155105, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.2, "cuda_time_us": 20.511, "pct_cuda_time": 0.30049867748936576, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.511, "pct_cuda_time": 0.30049867748936576, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 518.441, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 706.777, "cuda_time_us": 16.64, "pct_cuda_time": 0.24378616320135768, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.896, "pct_cuda_time": 0.1889342764810522, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.905, "cuda_time_us": 17.408, "pct_cuda_time": 0.25503783227218957, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.22456456187201984, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.03047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.836, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.135, "cuda_time_us": 134.302, "pct_cuda_time": 1.9676063275401883, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.098, "cuda_time_us": 81.567, "pct_cuda_time": 1.1950063686205012, "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": 1.1950063686205012, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.65, "cuda_time_us": 8.672, "pct_cuda_time": 0.12705009659147679, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.672, "pct_cuda_time": 0.12705009659147679, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.651, "cuda_time_us": 44.063, "pct_cuda_time": 0.6455498623282105, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.063, "pct_cuda_time": 0.6455498623282105, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2400.948, "cuda_time_us": 197.723, "pct_cuda_time": 2.8967627131407476, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.912, "cuda_time_us": 3.328, "pct_cuda_time": 0.04875723264027153, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04875723264027153, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1689.796, "cuda_time_us": 57.757999999999996, "pct_cuda_time": 0.8461899768139433, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 133.083, "cuda_time_us": 20.607, "pct_cuda_time": 0.3019051361232198, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.607, "pct_cuda_time": 0.3019051361232198, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.029, "cuda_time_us": 3.519, "pct_cuda_time": 0.05155549929721019, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.519, "pct_cuda_time": 0.05155549929721019, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 714.496, "cuda_time_us": 16.224, "pct_cuda_time": 0.23769150912132372, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 168.719, "cuda_time_us": 17.408, "pct_cuda_time": 0.25503783227218957, "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": 15.296, "pct_cuda_time": 0.22409574232740181, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.139, "cuda_time_us": 3.328, "pct_cuda_time": 0.04875723264027153, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04875723264027153, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 479.497, "cuda_time_us": 133.309, "pct_cuda_time": 1.9530582710462614, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 178.939, "cuda_time_us": 82.046, "pct_cuda_time": 1.202024011179002, "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": 82.046, "pct_cuda_time": 1.202024011179002, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.757, "cuda_time_us": 8.416, "pct_cuda_time": 0.12329954023453282, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.416, "pct_cuda_time": 0.12329954023453282, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.53, "cuda_time_us": 42.847, "pct_cuda_time": 0.6277347196327266, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.847, "pct_cuda_time": 0.6277347196327266, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2405.896, "cuda_time_us": 198.558, "pct_cuda_time": 2.9089959731331234, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.155, "cuda_time_us": 3.136, "pct_cuda_time": 0.04594431537256356, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04594431537256356, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1706.15, "cuda_time_us": 57.952, "pct_cuda_time": 0.8490321953031899, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.965, "cuda_time_us": 20.48, "pct_cuda_time": 0.3000445085555171, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3000445085555171, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 497.846, "cuda_time_us": 3.68, "pct_cuda_time": 0.05391424763106948, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.68, "pct_cuda_time": 0.05391424763106948, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 717.556, "cuda_time_us": 16.352, "pct_cuda_time": 0.23956678729979572, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.304, "pct_cuda_time": 0.03375500721249567, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.544, "pct_cuda_time": 0.18377726149025425, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 183.009, "cuda_time_us": 17.439999999999998, "pct_cuda_time": 0.25550665181680754, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.22456456187201984, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.274, "cuda_time_us": 3.52, "pct_cuda_time": 0.0515701499079795, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.52, "pct_cuda_time": 0.0515701499079795, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 478.268, "cuda_time_us": 133.95, "pct_cuda_time": 1.9624493125493903, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.729, "cuda_time_us": 82.175, "pct_cuda_time": 1.203913939968243, "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": 1.203913939968243, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.489, "cuda_time_us": 8.832, "pct_cuda_time": 0.12939419431456678, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.832, "pct_cuda_time": 0.12939419431456678, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.398, "cuda_time_us": 42.943, "pct_cuda_time": 0.6291411782665806, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.943, "pct_cuda_time": 0.6291411782665806, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2351.863, "cuda_time_us": 197.053, "pct_cuda_time": 2.886946803925308, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.438, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1660.607, "cuda_time_us": 57.983000000000004, "pct_cuda_time": 0.8494863642370386, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.24, "cuda_time_us": 20.703, "pct_cuda_time": 0.30331159475707375, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.703, "pct_cuda_time": 0.30331159475707375, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 472.697, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.806, "cuda_time_us": 16.192, "pct_cuda_time": 0.23722268957670573, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.512, "pct_cuda_time": 0.18330844194563625, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 166.318, "cuda_time_us": 17.536, "pct_cuda_time": 0.2569131104506615, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.2259710205058738, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.025, "cuda_time_us": 3.36, "pct_cuda_time": 0.04922605218488953, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04922605218488953, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.307, "cuda_time_us": 132.478, "pct_cuda_time": 1.9408836134969627, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.604, "cuda_time_us": 81.055, "pct_cuda_time": 1.1875052559066135, "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": 81.055, "pct_cuda_time": 1.1875052559066135, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.669, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.318, "cuda_time_us": 42.783, "pct_cuda_time": 0.6267970805434907, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.783, "pct_cuda_time": 0.6267970805434907, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2483.227, "cuda_time_us": 198.108, "pct_cuda_time": 2.902403198286933, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.198, "cuda_time_us": 3.36, "pct_cuda_time": 0.04922605218488953, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04922605218488953, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1772.548, "cuda_time_us": 58.045, "pct_cuda_time": 0.850394702104736, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.408, "cuda_time_us": 20.544, "pct_cuda_time": 0.3009821476447531, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3009821476447531, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 488.698, "cuda_time_us": 3.52, "pct_cuda_time": 0.0515701499079795, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.52, "pct_cuda_time": 0.0515701499079795, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 827.117, "cuda_time_us": 16.543, "pct_cuda_time": 0.24236505395673433, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.639, "pct_cuda_time": 0.1851690695133389, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.02390979677551777, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.048, "cuda_time_us": 17.438, "pct_cuda_time": 0.2554773505952689, "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": 15.327, "pct_cuda_time": 0.22454991126125054, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030927439334018394, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.718, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 476.636, "cuda_time_us": 133.503, "pct_cuda_time": 1.9559004895355077, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.978, "cuda_time_us": 82.271, "pct_cuda_time": 1.205320398602097, "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": 1.205320398602097, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 116.998, "cuda_time_us": 8.768, "pct_cuda_time": 0.12845655522533075, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.768, "pct_cuda_time": 0.12845655522533075, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.979, "cuda_time_us": 42.464, "pct_cuda_time": 0.62212353570808, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.464, "pct_cuda_time": 0.62212353570808, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2964.309, "cuda_time_us": 198.077, "pct_cuda_time": 2.901949029353084, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.912, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2251.72, "cuda_time_us": 58.304, "pct_cuda_time": 0.8541892102939878, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.838, "cuda_time_us": 20.736, "pct_cuda_time": 0.3037950649124611, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.736, "pct_cuda_time": 0.3037950649124611, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 635.23, "cuda_time_us": 3.713, "pct_cuda_time": 0.05439771778645679, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.713, "pct_cuda_time": 0.05439771778645679, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1143.759, "cuda_time_us": 16.319000000000003, "pct_cuda_time": 0.23908331714440845, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.608, "pct_cuda_time": 0.18471490057949022, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.439, "pct_cuda_time": 0.021082228897040485, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 161.906, "cuda_time_us": 17.536, "pct_cuda_time": 0.2569131104506615, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.2259710205058738, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.839, "cuda_time_us": 3.391, "pct_cuda_time": 0.04968022111873821, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.391, "pct_cuda_time": 0.04968022111873821, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.295, "cuda_time_us": 133.182, "pct_cuda_time": 1.9511976434785583, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.812, "cuda_time_us": 81.279, "pct_cuda_time": 1.1907869927189392, "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": 81.279, "pct_cuda_time": 1.1907869927189392, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 117.475, "cuda_time_us": 8.672, "pct_cuda_time": 0.12705009659147679, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.672, "pct_cuda_time": 0.12705009659147679, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.5, "cuda_time_us": 43.231, "pct_cuda_time": 0.6333605541681426, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.231, "pct_cuda_time": 0.6333605541681426, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2309.019, "cuda_time_us": 199.325, "pct_cuda_time": 2.9202329915931857, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.664, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1636.326, "cuda_time_us": 58.815, "pct_cuda_time": 0.8616756723971063, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.498, "cuda_time_us": 20.64, "pct_cuda_time": 0.3023886062786071, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3023886062786071, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.445, "cuda_time_us": 3.584, "pct_cuda_time": 0.0525077889972155, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.584, "pct_cuda_time": 0.0525077889972155, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 687.945, "cuda_time_us": 16.448, "pct_cuda_time": 0.24097324593364966, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.704, "pct_cuda_time": 0.18612135921334422, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 148.377, "cuda_time_us": 18.143, "pct_cuda_time": 0.26580603118763413, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.903, "pct_cuda_time": 0.23298866306437443, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.892, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 450.752, "cuda_time_us": 134.078, "pct_cuda_time": 1.9643245907278624, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.096, "cuda_time_us": 81.567, "pct_cuda_time": 1.1950063686205012, "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": 1.1950063686205012, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.493, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.64, "pct_cuda_time": 0.1265812770468588, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 139.467, "cuda_time_us": 43.871, "pct_cuda_time": 0.6427369450605025, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6427369450605025, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2534.161, "cuda_time_us": 198.61899999999997, "pct_cuda_time": 2.909889660390051, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 65.594, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1866.04, "cuda_time_us": 58.492999999999995, "pct_cuda_time": 0.8569581757293877, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.678, "cuda_time_us": 21.119, "pct_cuda_time": 0.3094062488371077, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.119, "pct_cuda_time": 0.3094062488371077, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.663, "cuda_time_us": 3.648, "pct_cuda_time": 0.05344542808645149, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.648, "pct_cuda_time": 0.05344542808645149, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 918.532, "cuda_time_us": 16.447000000000003, "pct_cuda_time": 0.24095859532288041, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.272, "pct_cuda_time": 0.03328618766787768, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.704, "pct_cuda_time": 0.18612135921334422, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.02155104844165848, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 174.899, "cuda_time_us": 17.279, "pct_cuda_time": 0.25314790348294824, "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": 15.2, "pct_cuda_time": 0.22268928369354787, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 2.079, "pct_cuda_time": 0.030458619789400395, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.746, "cuda_time_us": 3.296, "pct_cuda_time": 0.04828841309565353, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04828841309565353, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.653, "cuda_time_us": 133.59799999999998, "pct_cuda_time": 1.9572922975585925, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.7, "cuda_time_us": 82.207, "pct_cuda_time": 1.204382759512861, "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": 82.207, "pct_cuda_time": 1.204382759512861, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.368, "cuda_time_us": 8.544, "pct_cuda_time": 0.1251748184130048, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.544, "pct_cuda_time": 0.1251748184130048, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.146, "cuda_time_us": 42.847, "pct_cuda_time": 0.6277347196327266, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 42.847, "pct_cuda_time": 0.6277347196327266, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2244.886, "cuda_time_us": 198.938, "pct_cuda_time": 2.9145632052254618, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.324, "cuda_time_us": 3.136, "pct_cuda_time": 0.04594431537256356, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04594431537256356, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1589.827, "cuda_time_us": 58.653000000000006, "pct_cuda_time": 0.8593022734524778, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.956, "cuda_time_us": 21.023, "pct_cuda_time": 0.30799979020325374, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.023, "pct_cuda_time": 0.30799979020325374, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.583, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 667.589, "cuda_time_us": 16.351, "pct_cuda_time": 0.23955213668902636, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.304, "pct_cuda_time": 0.03375500721249567, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.543, "pct_cuda_time": 0.1837626108794849, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.022034518597045787, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 148.392, "cuda_time_us": 17.727, "pct_cuda_time": 0.25971137710760017, "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": 15.647, "pct_cuda_time": 0.2292381067074305, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.03047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.518, "cuda_time_us": 3.264, "pct_cuda_time": 0.04781959355103554, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04781959355103554, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 442.41, "cuda_time_us": 133.885, "pct_cuda_time": 1.961497022849385, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.48, "cuda_time_us": 81.95, "pct_cuda_time": 1.200617552545148, "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": 81.95, "pct_cuda_time": 1.200617552545148, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.837, "cuda_time_us": 8.608, "pct_cuda_time": 0.1261124575022408, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.608, "pct_cuda_time": 0.1261124575022408, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 137.662, "cuda_time_us": 43.327, "pct_cuda_time": 0.6347670128019965, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6347670128019965, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2435.315, "cuda_time_us": 198.303, "pct_cuda_time": 2.9052600673869486, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.755, "cuda_time_us": 3.2, "pct_cuda_time": 0.046881954461799547, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.046881954461799547, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1777.162, "cuda_time_us": 57.855000000000004, "pct_cuda_time": 0.8476110860585665, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 131.242, "cuda_time_us": 20.544, "pct_cuda_time": 0.3009821476447531, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3009821476447531, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 539.129, "cuda_time_us": 3.617, "pct_cuda_time": 0.052991259152602804, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.617, "pct_cuda_time": 0.052991259152602804, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 744.759, "cuda_time_us": 16.319000000000003, "pct_cuda_time": 0.23908331714440845, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.608, "pct_cuda_time": 0.18471490057949022, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.02155104844165848, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 216.019, "cuda_time_us": 17.375, "pct_cuda_time": 0.2545543621168022, "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": 15.295, "pct_cuda_time": 0.22408109171663254, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.08, "pct_cuda_time": 0.03047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.696, "cuda_time_us": 3.425, "pct_cuda_time": 0.050178341884894825, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.425, "pct_cuda_time": 0.050178341884894825, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 443.494, "cuda_time_us": 133.823, "pct_cuda_time": 1.960588684981688, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.628, "cuda_time_us": 82.079, "pct_cuda_time": 1.202507481334389, "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": 82.079, "pct_cuda_time": 1.202507481334389, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.369, "cuda_time_us": 8.513, "pct_cuda_time": 0.1247206494791561, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.513, "pct_cuda_time": 0.1247206494791561, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 136.248, "cuda_time_us": 43.231, "pct_cuda_time": 0.6333605541681426, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.231, "pct_cuda_time": 0.6333605541681426, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2359.61, "cuda_time_us": 198.81, "pct_cuda_time": 2.9126879270469903, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.494, "cuda_time_us": 3.135, "pct_cuda_time": 0.04592966476179425, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04592966476179425, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1649.75, "cuda_time_us": 58.237, "pct_cuda_time": 0.8532076193724438, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.406, "cuda_time_us": 20.768, "pct_cuda_time": 0.3042638844570791, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.768, "pct_cuda_time": 0.3042638844570791, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.403, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 715.139, "cuda_time_us": 16.478, "pct_cuda_time": 0.24141276425672908, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.271, "pct_cuda_time": 0.033271537057108364, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.735, "pct_cuda_time": 0.1865755281471929, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021565699052427792, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 157.156, "cuda_time_us": 17.439, "pct_cuda_time": 0.2554920012060382, "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": 15.327, "pct_cuda_time": 0.22454991126125054, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030942089944787705, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.487, "cuda_time_us": 3.616, "pct_cuda_time": 0.052976608541833496, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.616, "pct_cuda_time": 0.052976608541833496, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.455, "cuda_time_us": 133.822, "pct_cuda_time": 1.9605740343709186, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 188.815, "cuda_time_us": 81.247, "pct_cuda_time": 1.190318173174321, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.247, "pct_cuda_time": 1.190318173174321, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.761, "cuda_time_us": 8.896, "pct_cuda_time": 0.13033183340380278, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.896, "pct_cuda_time": 0.13033183340380278, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.44, "cuda_time_us": 43.679, "pct_cuda_time": 0.6399240277927946, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.679, "pct_cuda_time": 0.6399240277927946, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2391.74, "cuda_time_us": 198.427, "pct_cuda_time": 2.907076743122343, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.962, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.232, "pct_cuda_time": 0.047350774006417545, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1726.536, "cuda_time_us": 58.11, "pct_cuda_time": 0.8513469918047412, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.616, "cuda_time_us": 20.8, "pct_cuda_time": 0.3047327040016971, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.8, "pct_cuda_time": 0.3047327040016971, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[2, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 533.318, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.052038969452597506, "trace": "_C::rotary_embedding(int64[2], bfloat16[2, 4096], bfloat16[2, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 724.156, "cuda_time_us": 16.319, "pct_cuda_time": 0.23908331714440836, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.03281736812325969, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[2], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.639, "pct_cuda_time": 0.1851690695133389, "trace": "_vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.021096879507809797, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[2, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[2, 1, 32, 128], None, None, None, None, int32[2], None, None, int32[2, 17], None, None, None, None, None, None, None, 0.08838834764831845, True, -1, -1, 0, 0.0, True, 0, None, 0) <- vllm::unified_attention_with_output(bfloat16[2, 32, 128], bfloat16[2, 8, 128], bfloat16[2, 8, 128], bfloat16[2, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.148, "cuda_time_us": 17.439, "pct_cuda_time": 0.2554920012060382, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.22456456187201984, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 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": 0.030927439334018394, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[2, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.712, "cuda_time_us": 3.36, "pct_cuda_time": 0.04922605218488953, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04922605218488953, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 447.44, "cuda_time_us": 133.725, "pct_cuda_time": 1.9591529251262951, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.395, "cuda_time_us": 82.302, "pct_cuda_time": 1.205774567535946, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.302, "pct_cuda_time": 1.205774567535946, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[2, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.034, "cuda_time_us": 8.448, "pct_cuda_time": 0.12376835977915082, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.448, "pct_cuda_time": 0.12376835977915082, "trace": "_C::silu_and_mul(bfloat16[2, 14336], bfloat16[2, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 135.077, "cuda_time_us": 42.975, "pct_cuda_time": 0.6296099978111986, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6296099978111986, "trace": "mm(bfloat16[2, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[2, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[2, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.994, "cuda_time_us": 3.296, "pct_cuda_time": 0.04828841309565353, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04828841309565353, "trace": "_C::fused_add_rms_norm(bfloat16[2, 4096], bfloat16[2, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 478.214, "cuda_time_us": 342.46, "pct_cuda_time": 5.01724816405871, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 2.977, "pct_cuda_time": 0.04361486826024289, "trace": "index_select(bfloat16[2, 4096], 0, int64[2])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011251669070831891, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 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": 338.715, "pct_cuda_time": 4.962381626727636, "trace": "mm(bfloat16[2, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[2, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[2, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3217.004, "cuda_time_us": 114.71600000000001, "pct_cuda_time": 1.6806594650124365, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011251669070831891, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.011251669070831891, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.011251669070831891, "trace": "copy_(int32[2], int32[2], True) <- _to_copy(int32[2], 3, 0, None, None, True, None) <- to(int32[2], 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": 0.011251669070831891, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.011720488615449887, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.896, "pct_cuda_time": 0.013126947249303874, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 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": 0.011251669070831891, "trace": "copy_(bfloat16[2], bfloat16[2], True) <- _to_copy(bfloat16[2], 15, 0, None, None, True, None) <- to(bfloat16[2], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.096, "pct_cuda_time": 0.06000890171110342, "trace": "copy_(float32[2, 128256], bfloat16[2, 128256], False) <- _to_copy(bfloat16[2, 128256], 6, None, None, None, False, None) <- to(bfloat16[2, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.735, "pct_cuda_time": 0.06937064199269402, "trace": "div_(float32[2, 128256], bfloat16[2, 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": 34.655, "pct_cuda_time": 0.5077169162105198, "trace": "_softmax(float32[2, 128256], -1, False) <- softmax(float32[2, 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": 28.191, "pct_cuda_time": 0.41301536819768464, "trace": "_log_softmax(float32[2, 128256], -1, False) <- log_softmax(float32[2, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.952, "pct_cuda_time": 0.028597992221697722, "trace": "copy_(int64[2], int32[2], False) <- _to_copy(int32[2], 4, None, None, None, False, None) <- to(int32[2], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 4.672, "pct_cuda_time": 0.06844765351422734, "trace": "index(float32[2, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 28.384, "pct_cuda_time": 0.415842936076162, "trace": "argmax(float32[2, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.495, "pct_cuda_time": 0.03655327386943434, "trace": "copy_(int64[2], int64[2], False) <- _to_copy(int64[2], 4, 0, None, None, False, None) <- to(int64[2], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }