{ "context": { "python_version": "3.12.9 | packaged by Anaconda, Inc. | (main, Feb 6 2025, 18:56:27) [GCC 11.2.0]", "torch_version": "2.5.1+cu124", "engine_args": { "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "served_model_name": null, "tokenizer": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "task": "auto", "skip_tokenizer_init": false, "tokenizer_mode": "auto", "trust_remote_code": false, "allowed_local_media_path": null, "download_dir": null, "load_format": "dummy", "config_format": "auto", "dtype": "auto", "kv_cache_dtype": "auto", "seed": 0, "max_model_len": null, "distributed_executor_backend": null, "pipeline_parallel_size": 1, "tensor_parallel_size": 1, "max_parallel_loading_workers": null, "block_size": null, "enable_prefix_caching": false, "disable_sliding_window": false, "use_v2_block_manager": true, "swap_space": 4, "cpu_offload_gb": 0, "gpu_memory_utilization": 0.9, "max_num_batched_tokens": 8000, "max_num_partial_prefills": 1, "max_long_partial_prefills": 1, "long_prefill_token_threshold": 0, "max_num_seqs": 256, "max_logprobs": 20, "disable_log_stats": false, "revision": null, "code_revision": null, "rope_scaling": null, "rope_theta": null, "hf_overrides": null, "tokenizer_revision": null, "quantization": null, "enforce_eager": true, "max_seq_len_to_capture": 8192, "disable_custom_all_reduce": false, "tokenizer_pool_size": 0, "tokenizer_pool_type": "ray", "tokenizer_pool_extra_config": null, "limit_mm_per_prompt": null, "mm_processor_kwargs": null, "disable_mm_preprocessor_cache": false, "enable_lora": false, "enable_lora_bias": false, "max_loras": 1, "max_lora_rank": 16, "enable_prompt_adapter": false, "max_prompt_adapters": 1, "max_prompt_adapter_token": 0, "fully_sharded_loras": false, "lora_extra_vocab_size": 256, "long_lora_scaling_factors": null, "lora_dtype": "auto", "max_cpu_loras": null, "device": "auto", "num_scheduler_steps": 1, "multi_step_stream_outputs": true, "ray_workers_use_nsight": false, "num_gpu_blocks_override": null, "num_lookahead_slots": 0, "model_loader_extra_config": null, "ignore_patterns": [], "preemption_mode": null, "scheduler_delay_factor": 0.0, "enable_chunked_prefill": null, "guided_decoding_backend": "xgrammar", "logits_processor_pattern": null, "speculative_model": null, "speculative_model_quantization": null, "speculative_draft_tensor_parallel_size": null, "num_speculative_tokens": null, "speculative_disable_mqa_scorer": false, "speculative_max_model_len": null, "speculative_disable_by_batch_size": null, "ngram_prompt_lookup_max": null, "ngram_prompt_lookup_min": null, "spec_decoding_acceptance_method": "rejection_sampler", "typical_acceptance_sampler_posterior_threshold": null, "typical_acceptance_sampler_posterior_alpha": null, "qlora_adapter_name_or_path": null, "disable_logprobs_during_spec_decoding": null, "otlp_traces_endpoint": null, "collect_detailed_traces": null, "disable_async_output_proc": false, "scheduling_policy": "fcfs", "scheduler_cls": "vllm.core.scheduler.Scheduler", "override_neuron_config": null, "override_pooler_config": null, "compilation_config": null, "worker_cls": "auto", "kv_transfer_config": null, "generation_config": null, "override_generation_config": null, "enable_sleep_mode": false, "model_impl": "auto", "calculate_kv_scales": false, "additional_config": null }, "prompt_len": 0, "batch_size": 6, "num_steps": 2, "complete_num_requests_per_step": null, "save_chrome_traces_folder": null }, "prefill": { "metadata": { "num_running_seqs": null }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 33598.87900000001, "pct_cuda_time": 98.6029086422082, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 42.239, "pct_cuda_time": 0.12395914334339046, "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": 42.239, "pct_cuda_time": 0.12395914334339046, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 33541.984000000004, "pct_cuda_time": 98.43593841420748, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 968.4659999999998, "pct_cuda_time": 2.8421651960794514, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 18.56, "pct_cuda_time": 0.05446818581058564, "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": 949.9059999999997, "pct_cuda_time": 2.787697010268866, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 7665.080000000001, "pct_cuda_time": 22.49477379811443, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 3374.614, "pct_cuda_time": 9.903507671929077, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.81000000000001, "pct_cuda_time": 0.06987540431842913, "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": 3350.803999999999, "pct_cuda_time": 9.833632267610644, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 608.147, "pct_cuda_time": 1.784734040740853, "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": 608.147, "pct_cuda_time": 1.784734040740853, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 1059.6989999999998, "pct_cuda_time": 3.1099074372463247, "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": 241.66200000000003, "pct_cuda_time": 0.709207474103327, "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": 772.6319999999998, "pct_cuda_time": 2.267449533362306, "invocations": 32 }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cuda_time_us": 45.405, "pct_cuda_time": 0.13325042978069188, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 2622.6200000000003, "pct_cuda_time": 7.696624648198175, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.71100000000001, "pct_cuda_time": 0.06958486819799552, "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": 2598.909, "pct_cuda_time": 7.627039780000179, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 24908.438000000006, "pct_cuda_time": 73.09899942001361, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 15300.471000000001, "pct_cuda_time": 44.902419041889935, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 23.712000000000007, "pct_cuda_time": 0.06958780290628272, "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": 15276.758999999996, "pct_cuda_time": 44.83283123898364, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 2134.466, "pct_cuda_time": 6.264035058964304, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 2134.466, "pct_cuda_time": 6.264035058964304, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 7473.501, "pct_cuda_time": 21.932545319159356, "invocations": 32 }, "children": [ { "entry": { "name": "Memset (Device)", "cuda_time_us": 25.921000000000014, "pct_cuda_time": 0.07607057351272582, "invocations": 32 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cuda_time_us": 7447.580000000001, "pct_cuda_time": 21.856474745646633, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 14.656, "pct_cuda_time": 0.043011084657324526, "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": 14.656, "pct_cuda_time": 0.043011084657324526, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 361.02, "pct_cuda_time": 1.0594883858479325, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 4.801, "pct_cuda_time": 0.014089534486886945, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "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": 355.483, "pct_cuda_time": 1.0432389060616605, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 115.038, "pct_cuda_time": 0.3376029719438659, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.440999999999999, "pct_cuda_time": 0.01596774779070024, "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": 4.48, "pct_cuda_time": 0.013147493126693086, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 5.503, "pct_cuda_time": 0.016149699704507154, "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": 34.495, "pct_cuda_time": 0.10123276236724955, "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": 27.295, "pct_cuda_time": 0.08010286269934996, "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": 0.005352907915867899, "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": 5.856, "pct_cuda_time": 0.017185651729891678, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cuda_time_us": 27.616, "pct_cuda_time": 0.08104490405954382, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.528, "pct_cuda_time": 0.007418942550062527, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 83843.571, "cuda_time_us": 33598.87900000001, "pct_cuda_time": 98.6029086422082, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 294.152, "cuda_time_us": 42.239, "pct_cuda_time": 0.12395914334339046, "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": 42.239, "pct_cuda_time": 0.12395914334339046, "trace": "index_select(bfloat16[128256, 4096], 0, int64[1536]) <- embedding(bfloat16[128256, 4096], int64[1536], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 3958.713, "cuda_time_us": 1054.32, "pct_cuda_time": 3.0941216413694317, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 233.055, "cuda_time_us": 18.56, "pct_cuda_time": 0.05446818581058564, "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": 18.56, "pct_cuda_time": 0.05446818581058564, "trace": "_C::rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2888.075, "cuda_time_us": 243.99599999999998, "pct_cuda_time": 0.716057083245671, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 370.179, "cuda_time_us": 109.31, "pct_cuda_time": 0.3207929628747369, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 108.542, "pct_cuda_time": 0.31853910691016096, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 944.758, "cuda_time_us": 19.2, "pct_cuda_time": 0.05634639911439894, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.2, "pct_cuda_time": 0.05634639911439894, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 996.903, "cuda_time_us": 33.119, "pct_cuda_time": 0.09719460376405097, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.648, "pct_cuda_time": 0.022444648980568912, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.999, "pct_cuda_time": 0.07043006418471147, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 304.066, "cuda_time_us": 82.367, "pct_cuda_time": 0.24172311749248426, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.631, "pct_cuda_time": 0.23956317219309894, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 122.463, "cuda_time_us": 15.424, "pct_cuda_time": 0.04526494062190048, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.424, "pct_cuda_time": 0.04526494062190048, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 584.055, "cuda_time_us": 776.34, "pct_cuda_time": 2.278331431691275, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 196.195, "cuda_time_us": 476.441, "pct_cuda_time": 1.3982153510657993, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 475.705, "pct_cuda_time": 1.3960554057664138, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 133.428, "cuda_time_us": 66.526, "pct_cuda_time": 0.1952344035148179, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.526, "pct_cuda_time": 0.1952344035148179, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 181.515, "cuda_time_us": 233.37300000000002, "pct_cuda_time": 0.6848816771106576, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.573, "pct_cuda_time": 0.6825339104808908, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2702.978, "cuda_time_us": 1047.187, "pct_cuda_time": 3.073188367156775, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.861, "cuda_time_us": 15.104, "pct_cuda_time": 0.04432583396999383, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.04432583396999383, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1942.747, "cuda_time_us": 239.99699999999999, "pct_cuda_time": 0.7043211848051251, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 163.663, "cuda_time_us": 105.79, "pct_cuda_time": 0.3104627897037638, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 105.054, "pct_cuda_time": 0.30830284440437844, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 571.764, "cuda_time_us": 18.912, "pct_cuda_time": 0.05550120312768295, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.912, "pct_cuda_time": 0.05550120312768295, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 802.531, "cuda_time_us": 33.056, "pct_cuda_time": 0.09700971714195684, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.552, "pct_cuda_time": 0.022162916984996915, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.0, "pct_cuda_time": 0.07043299889299869, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004413801263961251, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 207.122, "cuda_time_us": 82.239, "pct_cuda_time": 0.2413474748317216, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.503, "pct_cuda_time": 0.2391875295323363, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 98.454, "cuda_time_us": 15.232, "pct_cuda_time": 0.04470147663075649, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.232, "pct_cuda_time": 0.04470147663075649, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 490.614, "cuda_time_us": 776.854, "pct_cuda_time": 2.2798398717509, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.459, "cuda_time_us": 476.89, "pct_cuda_time": 1.3995330350867556, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.154, "pct_cuda_time": 1.3973730897873704, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.563, "cuda_time_us": 66.464, "pct_cuda_time": 0.195052451601011, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.464, "pct_cuda_time": 0.195052451601011, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.198, "cuda_time_us": 233.5, "pct_cuda_time": 0.685254385063133, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.668, "pct_cuda_time": 0.6828127077681757, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2492.126, "cuda_time_us": 1048.6889999999999, "pct_cuda_time": 3.0775962990041617, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.919, "cuda_time_us": 14.56, "pct_cuda_time": 0.04272935266175253, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.56, "pct_cuda_time": 0.04272935266175253, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1778.458, "cuda_time_us": 239.804, "pct_cuda_time": 0.7037547861056939, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.047, "cuda_time_us": 105.343, "pct_cuda_time": 0.30915097509938166, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.607, "pct_cuda_time": 0.3069910297999964, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 514.261, "cuda_time_us": 18.847, "pct_cuda_time": 0.055310447089014426, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.847, "pct_cuda_time": 0.055310447089014426, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 764.587, "cuda_time_us": 32.959, "pct_cuda_time": 0.09672505043809765, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.328, "pct_cuda_time": 0.021505542328662263, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.224, "pct_cuda_time": 0.07109037354933333, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.407, "pct_cuda_time": 0.004129134560102048, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 196.513, "cuda_time_us": 82.655, "pct_cuda_time": 0.24256831347920024, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.887, "pct_cuda_time": 0.24031445751462427, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.143, "cuda_time_us": 15.488, "pct_cuda_time": 0.04545276195228181, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.488, "pct_cuda_time": 0.04545276195228181, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 473.13, "cuda_time_us": 778.837, "pct_cuda_time": 2.2856593982844338, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.006, "cuda_time_us": 477.849, "pct_cuda_time": 1.4023474203341884, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.113, "pct_cuda_time": 1.4001874750348033, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.084, "cuda_time_us": 67.199, "pct_cuda_time": 0.19720946219210908, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.199, "pct_cuda_time": 0.19720946219210908, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 159.535, "cuda_time_us": 233.78900000000002, "pct_cuda_time": 0.6861025157581362, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.989, "pct_cuda_time": 0.6837547491283695, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2577.412, "cuda_time_us": 1044.85, "pct_cuda_time": 3.066329953889569, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.071, "cuda_time_us": 15.072, "pct_cuda_time": 0.04423192330480317, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.04423192330480317, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1857.811, "cuda_time_us": 236.572, "pct_cuda_time": 0.6942698089214369, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 180.195, "cuda_time_us": 105.182, "pct_cuda_time": 0.3086784870651411, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.446, "pct_cuda_time": 0.3065185417657558, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 497.284, "cuda_time_us": 18.784, "pct_cuda_time": 0.05512556046692029, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.784, "pct_cuda_time": 0.05512556046692029, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 764.459, "cuda_time_us": 33.119, "pct_cuda_time": 0.09719460376405097, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.583, "pct_cuda_time": 0.022253892941900374, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.064, "pct_cuda_time": 0.07062082022338001, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 235.411, "cuda_time_us": 79.48700000000001, "pct_cuda_time": 0.2332711576253244, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 78.751, "pct_cuda_time": 0.23111121232593912, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.879, "cuda_time_us": 15.488, "pct_cuda_time": 0.04545276195228181, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.488, "pct_cuda_time": 0.04545276195228181, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 487.882, "cuda_time_us": 777.718, "pct_cuda_time": 2.282375459711048, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.702, "cuda_time_us": 477.754, "pct_cuda_time": 1.4020686230469037, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.704, "pct_cuda_time": 0.002066034634194628, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.05, "pct_cuda_time": 1.4000025884127092, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.371, "cuda_time_us": 67.007, "pct_cuda_time": 0.1966459982009651, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.007, "pct_cuda_time": 0.1966459982009651, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.008, "cuda_time_us": 232.95700000000002, "pct_cuda_time": 0.6836608384631789, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.157, "pct_cuda_time": 0.6813130718334123, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2504.93, "cuda_time_us": 1049.493, "pct_cuda_time": 3.079955804467077, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.548, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1809.169, "cuda_time_us": 240.99, "pct_cuda_time": 0.707235350134323, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 159.945, "cuda_time_us": 105.119, "pct_cuda_time": 0.30849360044304697, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.383, "pct_cuda_time": 0.3063336551436617, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 556.584, "cuda_time_us": 18.912, "pct_cuda_time": 0.05550120312768295, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.912, "pct_cuda_time": 0.05550120312768295, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 739.967, "cuda_time_us": 33.152, "pct_cuda_time": 0.09729144913752884, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.456, "pct_cuda_time": 0.021881184989424922, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.256, "pct_cuda_time": 0.071184284214524, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.00422597993357992, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 196.363, "cuda_time_us": 83.807, "pct_cuda_time": 0.24594909742606416, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.071, "pct_cuda_time": 0.24378915212667887, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.188, "cuda_time_us": 15.136, "pct_cuda_time": 0.0444197446351845, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.136, "pct_cuda_time": 0.0444197446351845, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.379, "cuda_time_us": 778.487, "pct_cuda_time": 2.2846322503839107, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.479, "cuda_time_us": 477.914, "pct_cuda_time": 1.402538176372857, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.178, "pct_cuda_time": 1.4003782310734716, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.761, "cuda_time_us": 66.304, "pct_cuda_time": 0.19458289827505768, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.304, "pct_cuda_time": 0.19458289827505768, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.333, "cuda_time_us": 234.269, "pct_cuda_time": 0.6875111757359961, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.469, "pct_cuda_time": 0.6851634091062294, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2754.638, "cuda_time_us": 1048.0520000000001, "pct_cuda_time": 3.075726889825211, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.754, "cuda_time_us": 14.944, "pct_cuda_time": 0.04385628064404051, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.944, "pct_cuda_time": 0.04385628064404051, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1987.506, "cuda_time_us": 239.74099999999999, "pct_cuda_time": 0.7035698994835997, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.592, "cuda_time_us": 105.119, "pct_cuda_time": 0.30849360044304697, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.383, "pct_cuda_time": 0.3063336551436617, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 692.244, "cuda_time_us": 19.424, "pct_cuda_time": 0.05700377377073359, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.424, "pct_cuda_time": 0.05700377377073359, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 740.239, "cuda_time_us": 33.086999999999996, "pct_cuda_time": 0.09710069309886028, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.616, "pct_cuda_time": 0.022350738315378246, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.159, "pct_cuda_time": 0.07089961751066479, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0038503372728172615, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 245.373, "cuda_time_us": 82.111, "pct_cuda_time": 0.24097183217095894, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.375, "pct_cuda_time": 0.23881188687157362, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 90.701, "cuda_time_us": 15.296, "pct_cuda_time": 0.044889297961137824, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.296, "pct_cuda_time": 0.044889297961137824, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 499.253, "cuda_time_us": 778.071, "pct_cuda_time": 2.283411411736432, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 176.69, "cuda_time_us": 477.531, "pct_cuda_time": 1.4014141830988565, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.794, "pct_cuda_time": 1.3992513030911837, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 106.333, "cuda_time_us": 66.655, "pct_cuda_time": 0.1956129808838678, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.655, "pct_cuda_time": 0.1956129808838678, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 160.304, "cuda_time_us": 233.885, "pct_cuda_time": 0.6863842477537081, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.053, "pct_cuda_time": 0.6839425704587508, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2466.14, "cuda_time_us": 1045.906, "pct_cuda_time": 3.0694290058408615, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.724, "cuda_time_us": 14.592, "pct_cuda_time": 0.042823263326943195, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.592, "pct_cuda_time": 0.042823263326943195, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1785.074, "cuda_time_us": 239.80599999999998, "pct_cuda_time": 0.7037606555222683, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.876, "cuda_time_us": 105.215, "pct_cuda_time": 0.308775332438619, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.479, "pct_cuda_time": 0.3066153871392337, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 542.588, "cuda_time_us": 19.168, "pct_cuda_time": 0.05625248844920828, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.168, "pct_cuda_time": 0.05625248844920828, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 742.623, "cuda_time_us": 33.150999999999996, "pct_cuda_time": 0.09728851442924162, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.52, "pct_cuda_time": 0.02206900631980625, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.127, "pct_cuda_time": 0.07080570684547412, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004413801263961251, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.439, "cuda_time_us": 82.27199999999999, "pct_cuda_time": 0.24144432020519946, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.535, "pct_cuda_time": 0.23928144019752695, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.804, "cuda_time_us": 15.295, "pct_cuda_time": 0.044886363252850614, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.295, "pct_cuda_time": 0.044886363252850614, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.368, "cuda_time_us": 776.213, "pct_cuda_time": 2.277958723738799, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.045, "cuda_time_us": 477.17699999999996, "pct_cuda_time": 1.4003752963651843, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.441, "pct_cuda_time": 1.3982153510657993, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.551, "cuda_time_us": 66.175, "pct_cuda_time": 0.19420432090600778, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.175, "pct_cuda_time": 0.19420432090600778, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 150.343, "cuda_time_us": 232.86100000000002, "pct_cuda_time": 0.6833791064676069, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.061, "pct_cuda_time": 0.6810313398378403, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2456.951, "cuda_time_us": 1046.865, "pct_cuda_time": 3.072243391088294, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.338, "cuda_time_us": 14.527, "pct_cuda_time": 0.042632507288274654, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.527, "pct_cuda_time": 0.042632507288274654, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1747.934, "cuda_time_us": 239.613, "pct_cuda_time": 0.7031942568228372, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.462, "cuda_time_us": 105.216, "pct_cuda_time": 0.30877826714690615, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.479, "pct_cuda_time": 0.3066153871392337, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 515.283, "cuda_time_us": 19.359, "pct_cuda_time": 0.05681301773206506, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.359, "pct_cuda_time": 0.05681301773206506, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 719.434, "cuda_time_us": 33.150999999999996, "pct_cuda_time": 0.09728851442924162, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.744, "pct_cuda_time": 0.022726380976140905, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.935, "pct_cuda_time": 0.07024224285433014, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.559, "cuda_time_us": 81.887, "pct_cuda_time": 0.24031445751462427, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.119, "pct_cuda_time": 0.23806060155004832, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.832, "cuda_time_us": 15.456, "pct_cuda_time": 0.04535885128709115, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.04535885128709115, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 480.974, "cuda_time_us": 777.269, "pct_cuda_time": 2.2810577756900914, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.807, "cuda_time_us": 477.657, "pct_cuda_time": 1.4017839563430445, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.921, "pct_cuda_time": 1.3996240110436593, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 113.452, "cuda_time_us": 66.303, "pct_cuda_time": 0.19457996356677046, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.303, "pct_cuda_time": 0.19457996356677046, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 156.543, "cuda_time_us": 233.309, "pct_cuda_time": 0.6846938557802762, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.509, "pct_cuda_time": 0.6823460891505095, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2621.256, "cuda_time_us": 1049.49, "pct_cuda_time": 3.079947000342216, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.878, "cuda_time_us": 14.528, "pct_cuda_time": 0.042635441996561864, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.528, "pct_cuda_time": 0.042635441996561864, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1768.597, "cuda_time_us": 239.71, "pct_cuda_time": 0.7034789235266964, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 164.278, "cuda_time_us": 105.183, "pct_cuda_time": 0.30868142177342833, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.447, "pct_cuda_time": 0.30652147647404304, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 498.393, "cuda_time_us": 19.04, "pct_cuda_time": 0.05587684578844561, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.04, "pct_cuda_time": 0.05587684578844561, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 735.123, "cuda_time_us": 33.184, "pct_cuda_time": 0.09738535980271949, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.68, "pct_cuda_time": 0.022538559645759574, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.064, "pct_cuda_time": 0.07062082022338001, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.00422597993357992, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 214.594, "cuda_time_us": 82.303, "pct_cuda_time": 0.24153529616210292, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.535, "pct_cuda_time": 0.23928144019752695, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.823, "cuda_time_us": 15.616, "pct_cuda_time": 0.04582840461304447, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.04582840461304447, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 623.081, "cuda_time_us": 779.636, "pct_cuda_time": 2.288004230205913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 230.718, "cuda_time_us": 478.873, "pct_cuda_time": 1.4053525616202898, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 478.137, "pct_cuda_time": 1.4031926163209045, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.917, "cuda_time_us": 66.527, "pct_cuda_time": 0.19523733822310513, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.527, "pct_cuda_time": 0.19523733822310513, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 231.179, "cuda_time_us": 234.23600000000002, "pct_cuda_time": 0.6874143303625183, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.436, "pct_cuda_time": 0.6850665637327515, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2489.352, "cuda_time_us": 1045.7469999999998, "pct_cuda_time": 3.068962387223195, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.224, "cuda_time_us": 15.136, "pct_cuda_time": 0.0444197446351845, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.136, "pct_cuda_time": 0.0444197446351845, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1785.957, "cuda_time_us": 238.55700000000002, "pct_cuda_time": 0.7000952048715452, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 159.5, "cuda_time_us": 104.991, "pct_cuda_time": 0.30811795778228435, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.255, "pct_cuda_time": 0.305958012482899, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 532.355, "cuda_time_us": 18.879, "pct_cuda_time": 0.05540435775420509, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.879, "pct_cuda_time": 0.05540435775420509, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 743.721, "cuda_time_us": 33.088, "pct_cuda_time": 0.09710362780714751, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.52, "pct_cuda_time": 0.02206900631980625, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.32, "pct_cuda_time": 0.07137210554490533, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003662515942435931, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 195.416, "cuda_time_us": 81.599, "pct_cuda_time": 0.2394692615279083, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.863, "pct_cuda_time": 0.23730931622852303, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.502, "cuda_time_us": 15.84, "pct_cuda_time": 0.04648577926937913, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.84, "pct_cuda_time": 0.04648577926937913, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 472.791, "cuda_time_us": 776.2139999999999, "pct_cuda_time": 2.2779616584470865, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.908, "cuda_time_us": 475.99399999999997, "pct_cuda_time": 1.396903536461417, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 475.258, "pct_cuda_time": 1.3947435911620318, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.441, "cuda_time_us": 66.815, "pct_cuda_time": 0.1960825342098211, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.1960825342098211, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.445, "cuda_time_us": 233.405, "pct_cuda_time": 0.6849755877758481, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.573, "pct_cuda_time": 0.6825339104808908, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2373.163, "cuda_time_us": 1047.345, "pct_cuda_time": 3.073652051066154, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.171, "cuda_time_us": 14.208, "pct_cuda_time": 0.041696335344655215, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.208, "pct_cuda_time": 0.041696335344655215, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1696.458, "cuda_time_us": 239.80400000000003, "pct_cuda_time": 0.703754786105694, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.958, "cuda_time_us": 104.60600000000001, "pct_cuda_time": 0.30698809509170916, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 103.87, "pct_cuda_time": 0.30482814979232387, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.718, "cuda_time_us": 19.519, "pct_cuda_time": 0.05728257105801838, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.519, "pct_cuda_time": 0.05728257105801838, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 745.57, "cuda_time_us": 32.896, "pct_cuda_time": 0.09654016381600351, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.392, "pct_cuda_time": 0.021693363659043594, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.224, "pct_cuda_time": 0.07109037354933333, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003756426607626596, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.127, "cuda_time_us": 82.783, "pct_cuda_time": 0.2429439561399629, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.015, "pct_cuda_time": 0.24069010017538695, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.124, "cuda_time_us": 15.296, "pct_cuda_time": 0.044889297961137824, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.296, "pct_cuda_time": 0.044889297961137824, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.287, "cuda_time_us": 778.037, "pct_cuda_time": 2.283311631654667, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.488, "cuda_time_us": 478.041, "pct_cuda_time": 1.4029108843253324, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.305, "pct_cuda_time": 1.4007509390259472, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 93.997, "cuda_time_us": 66.335, "pct_cuda_time": 0.1946738742319611, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.335, "pct_cuda_time": 0.1946738742319611, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.876, "cuda_time_us": 233.661, "pct_cuda_time": 0.6857268730973736, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.861, "pct_cuda_time": 0.6833791064676068, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2636.038, "cuda_time_us": 1049.94, "pct_cuda_time": 3.08126761907146, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.025, "cuda_time_us": 14.624, "pct_cuda_time": 0.042917173992133864, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.624, "pct_cuda_time": 0.042917173992133864, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1907.423, "cuda_time_us": 238.462, "pct_cuda_time": 0.6998164075842604, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.532, "cuda_time_us": 105.022, "pct_cuda_time": 0.3082089337391878, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.286, "pct_cuda_time": 0.3060489884398025, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.154, "cuda_time_us": 19.072, "pct_cuda_time": 0.05597075645363628, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.072, "pct_cuda_time": 0.05597075645363628, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 758.532, "cuda_time_us": 33.952999999999996, "pct_cuda_time": 0.09964215047558264, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.873, "pct_cuda_time": 0.023104958345190774, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.384, "pct_cuda_time": 0.07155992687528666, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.00497726525510524, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 206.03, "cuda_time_us": 80.415, "pct_cuda_time": 0.23599456691585372, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 79.679, "pct_cuda_time": 0.2338346216164684, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.678, "cuda_time_us": 15.583, "pct_cuda_time": 0.0457315592395666, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.583, "pct_cuda_time": 0.0457315592395666, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 500.24, "cuda_time_us": 781.2710000000001, "pct_cuda_time": 2.2928024782554988, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.746, "cuda_time_us": 480.442, "pct_cuda_time": 1.4099571189229196, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 479.706, "pct_cuda_time": 1.4077971736235344, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.501, "cuda_time_us": 67.296, "pct_cuda_time": 0.1974941288959683, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.296, "pct_cuda_time": 0.1974941288959683, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 178.535, "cuda_time_us": 233.53300000000002, "pct_cuda_time": 0.6853512304366108, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.733, "pct_cuda_time": 0.6830034638068442, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2663.165, "cuda_time_us": 1046.385, "pct_cuda_time": 3.070834731110434, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.406, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1949.943, "cuda_time_us": 238.364, "pct_cuda_time": 0.699528806172114, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.842, "cuda_time_us": 105.055, "pct_cuda_time": 0.3083057791126657, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.319, "pct_cuda_time": 0.30614583381328037, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.574, "cuda_time_us": 18.719, "pct_cuda_time": 0.05493480442825176, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.719, "pct_cuda_time": 0.05493480442825176, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 950.073, "cuda_time_us": 33.022999999999996, "pct_cuda_time": 0.09691287176847896, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.456, "pct_cuda_time": 0.021881184989424922, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.095, "pct_cuda_time": 0.07071179618028346, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 212.032, "cuda_time_us": 81.56700000000001, "pct_cuda_time": 0.23937535086271766, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.831, "pct_cuda_time": 0.23721540556333234, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 87.105, "cuda_time_us": 15.231, "pct_cuda_time": 0.04469854192246928, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.231, "pct_cuda_time": 0.04469854192246928, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 479.674, "cuda_time_us": 777.91, "pct_cuda_time": 2.2829389237021913, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.531, "cuda_time_us": 478.26599999999996, "pct_cuda_time": 1.4035711936899544, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.53, "pct_cuda_time": 1.401411248390569, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 103.763, "cuda_time_us": 66.623, "pct_cuda_time": 0.19551907021867712, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.623, "pct_cuda_time": 0.19551907021867712, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.867, "cuda_time_us": 233.021, "pct_cuda_time": 0.6838486597935601, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.189, "pct_cuda_time": 0.6814069824986029, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2560.456, "cuda_time_us": 1045.584, "pct_cuda_time": 3.0684840297723803, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.75, "cuda_time_us": 14.911, "pct_cuda_time": 0.043759435270562634, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.911, "pct_cuda_time": 0.043759435270562634, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1871.404, "cuda_time_us": 238.46, "pct_cuda_time": 0.699810538167686, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.955, "cuda_time_us": 105.055, "pct_cuda_time": 0.3083057791126657, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.287, "pct_cuda_time": 0.3060519231480897, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 592.424, "cuda_time_us": 19.135, "pct_cuda_time": 0.05615564307573041, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.135, "pct_cuda_time": 0.05615564307573041, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 778.478, "cuda_time_us": 33.472, "pct_cuda_time": 0.0982305557894355, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.712, "pct_cuda_time": 0.02263247031095024, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.352, "pct_cuda_time": 0.07146601621009599, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.408, "pct_cuda_time": 0.004132069268389256, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.632, "cuda_time_us": 80.798, "pct_cuda_time": 0.23711856018985444, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0021570105910980845, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.063, "pct_cuda_time": 0.23496154959875637, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.914, "cuda_time_us": 15.263, "pct_cuda_time": 0.044792452587659945, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.263, "pct_cuda_time": 0.044792452587659945, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.876, "cuda_time_us": 776.95, "pct_cuda_time": 2.2801216037464718, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.688, "cuda_time_us": 477.018, "pct_cuda_time": 1.3999086777475185, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.282, "pct_cuda_time": 1.397748732448133, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.811, "cuda_time_us": 66.815, "pct_cuda_time": 0.1960825342098211, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.1960825342098211, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.346, "cuda_time_us": 233.11700000000002, "pct_cuda_time": 0.6841303917891322, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.317, "pct_cuda_time": 0.6817826251593656, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2392.617, "cuda_time_us": 1044.435, "pct_cuda_time": 3.065112049950378, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.969, "cuda_time_us": 14.751, "pct_cuda_time": 0.04328988194460931, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.751, "pct_cuda_time": 0.04328988194460931, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1718.863, "cuda_time_us": 237.50099999999998, "pct_cuda_time": 0.6969961529202532, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.446, "cuda_time_us": 104.575, "pct_cuda_time": 0.3068971191348057, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 103.807, "pct_cuda_time": 0.3046432631702297, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 498.658, "cuda_time_us": 19.008, "pct_cuda_time": 0.05578293512325495, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.008, "pct_cuda_time": 0.05578293512325495, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 733.883, "cuda_time_us": 32.736, "pct_cuda_time": 0.0960706104900502, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02159945299385293, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.096, "pct_cuda_time": 0.07071473088857066, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003756426607626596, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 187.713, "cuda_time_us": 81.182, "pct_cuda_time": 0.23824548817214247, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.703, "pct_cuda_time": 0.0020630999259074194, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.479, "pct_cuda_time": 0.236182388246235, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.774, "cuda_time_us": 15.2, "pct_cuda_time": 0.044607565965565824, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.2, "pct_cuda_time": 0.044607565965565824, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 455.008, "cuda_time_us": 776.983, "pct_cuda_time": 2.2802184491199493, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.321, "cuda_time_us": 478.746, "pct_cuda_time": 1.4049798536678144, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.978, "pct_cuda_time": 1.4027259977032385, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.589, "cuda_time_us": 66.144, "pct_cuda_time": 0.19411334494910437, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.144, "pct_cuda_time": 0.19411334494910437, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.698, "cuda_time_us": 232.093, "pct_cuda_time": 0.6811252505030309, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 231.261, "pct_cuda_time": 0.6786835732080736, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2466.786, "cuda_time_us": 1046.928, "pct_cuda_time": 3.072428277710389, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 92.461, "cuda_time_us": 14.4, "pct_cuda_time": 0.04225979933579921, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.4, "pct_cuda_time": 0.04225979933579921, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1741.631, "cuda_time_us": 238.07500000000002, "pct_cuda_time": 0.6986806754771109, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.489, "cuda_time_us": 105.599, "pct_cuda_time": 0.30990226042090696, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.863, "pct_cuda_time": 0.30774231512152167, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 503.93, "cuda_time_us": 18.815, "pct_cuda_time": 0.05521653642382375, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.815, "pct_cuda_time": 0.05521653642382375, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 722.888, "cuda_time_us": 33.022, "pct_cuda_time": 0.09690993706019176, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.583, "pct_cuda_time": 0.022253892941900374, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.967, "pct_cuda_time": 0.0703361535195208, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 206.236, "cuda_time_us": 80.63900000000001, "pct_cuda_time": 0.23665194157218836, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 79.903, "pct_cuda_time": 0.23449199627280307, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.298, "cuda_time_us": 15.104, "pct_cuda_time": 0.04432583396999383, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.104, "pct_cuda_time": 0.04432583396999383, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.999, "cuda_time_us": 779.349, "pct_cuda_time": 2.2871619689274847, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.156, "cuda_time_us": 478.457, "pct_cuda_time": 1.4041317229728112, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.721, "pct_cuda_time": 1.401971777673426, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.917, "cuda_time_us": 66.623, "pct_cuda_time": 0.19551907021867712, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.623, "pct_cuda_time": 0.19551907021867712, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.319, "cuda_time_us": 234.269, "pct_cuda_time": 0.6875111757359961, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.469, "pct_cuda_time": 0.6851634091062294, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2468.669, "cuda_time_us": 1048.179, "pct_cuda_time": 3.076099597777686, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.998, "cuda_time_us": 14.368, "pct_cuda_time": 0.04216588867060854, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.368, "pct_cuda_time": 0.04216588867060854, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1780.09, "cuda_time_us": 239.902, "pct_cuda_time": 0.7040423875178403, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 155.624, "cuda_time_us": 105.375, "pct_cuda_time": 0.3092448857645723, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.638, "pct_cuda_time": 0.3070820057568998, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 522.517, "cuda_time_us": 18.944, "pct_cuda_time": 0.05559511379287362, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.944, "pct_cuda_time": 0.05559511379287362, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 756.598, "cuda_time_us": 32.896, "pct_cuda_time": 0.09654016381600351, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.264, "pct_cuda_time": 0.021317720998280932, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.224, "pct_cuda_time": 0.07109037354933333, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.408, "pct_cuda_time": 0.004132069268389256, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.445, "cuda_time_us": 82.687, "pct_cuda_time": 0.2426622241443909, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.951, "pct_cuda_time": 0.24050227884500558, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.654, "cuda_time_us": 15.232, "pct_cuda_time": 0.04470147663075649, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.232, "pct_cuda_time": 0.04470147663075649, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.985, "cuda_time_us": 778.677, "pct_cuda_time": 2.2851898449584804, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.306, "cuda_time_us": 478.137, "pct_cuda_time": 1.4031926163209045, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0021570105910980845, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.402, "pct_cuda_time": 1.4010356057298063, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.308, "cuda_time_us": 67.007, "pct_cuda_time": 0.1966459982009651, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.007, "pct_cuda_time": 0.1966459982009651, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.807, "cuda_time_us": 233.53300000000002, "pct_cuda_time": 0.6853512304366108, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.733, "pct_cuda_time": 0.6830034638068442, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2427.148, "cuda_time_us": 1050.0030000000002, "pct_cuda_time": 3.0814525056935542, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.021, "cuda_time_us": 14.752, "pct_cuda_time": 0.04329281665289652, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.752, "pct_cuda_time": 0.04329281665289652, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1706.331, "cuda_time_us": 242.01300000000003, "pct_cuda_time": 0.7102375567121371, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.48, "cuda_time_us": 105.598, "pct_cuda_time": 0.30989932571261974, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.83, "pct_cuda_time": 0.3076454697480438, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 490.441, "cuda_time_us": 19.072, "pct_cuda_time": 0.05597075645363628, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.072, "pct_cuda_time": 0.05597075645363628, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 706.736, "cuda_time_us": 33.024, "pct_cuda_time": 0.09691580647676619, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.584, "pct_cuda_time": 0.02225682765018758, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.16, "pct_cuda_time": 0.070902552218952, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003756426607626596, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 197.319, "cuda_time_us": 84.319, "pct_cuda_time": 0.24745166806911478, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.583, "pct_cuda_time": 0.24529172276972952, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 104.766, "cuda_time_us": 15.456, "pct_cuda_time": 0.04535885128709115, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.04535885128709115, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 475.92, "cuda_time_us": 777.7820000000002, "pct_cuda_time": 2.2825632810414294, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.421, "cuda_time_us": 477.69100000000003, "pct_cuda_time": 1.4018837364248096, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.954, "pct_cuda_time": 1.399720856417137, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.626, "cuda_time_us": 66.815, "pct_cuda_time": 0.1960825342098211, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.815, "pct_cuda_time": 0.1960825342098211, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.331, "cuda_time_us": 233.276, "pct_cuda_time": 0.6845970104067983, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.476, "pct_cuda_time": 0.6822492437770317, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2479.365, "cuda_time_us": 1047.377, "pct_cuda_time": 3.0737459617313445, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.863, "cuda_time_us": 14.56, "pct_cuda_time": 0.04272935266175253, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.56, "pct_cuda_time": 0.04272935266175253, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1773.732, "cuda_time_us": 238.33100000000002, "pct_cuda_time": 0.6994319607986362, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.424, "cuda_time_us": 105.246, "pct_cuda_time": 0.3088663083955224, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0021570105910980845, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.511, "pct_cuda_time": 0.30670929780442435, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 552.82, "cuda_time_us": 18.783, "pct_cuda_time": 0.05512262575863309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.783, "pct_cuda_time": 0.05512262575863309, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 728.075, "cuda_time_us": 32.863, "pct_cuda_time": 0.09644331844252564, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.584, "pct_cuda_time": 0.02225682765018758, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.0, "pct_cuda_time": 0.07043299889299869, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003753491899339387, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 192.481, "cuda_time_us": 81.43900000000001, "pct_cuda_time": 0.238999708201955, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.703, "pct_cuda_time": 0.2368397629025697, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.13, "cuda_time_us": 15.359, "pct_cuda_time": 0.045074184583231945, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.359, "pct_cuda_time": 0.045074184583231945, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 484.854, "cuda_time_us": 779.127, "pct_cuda_time": 2.286510463687724, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 179.565, "cuda_time_us": 478.554, "pct_cuda_time": 1.4044163896766704, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.818, "pct_cuda_time": 1.402256444377285, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.623, "cuda_time_us": 66.655, "pct_cuda_time": 0.1956129808838678, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.655, "pct_cuda_time": 0.1956129808838678, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.134, "cuda_time_us": 233.918, "pct_cuda_time": 0.6864810931271861, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.118, "pct_cuda_time": 0.6841333264974193, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2741.257, "cuda_time_us": 1049.33, "pct_cuda_time": 3.0794774470162625, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.75, "cuda_time_us": 15.04, "pct_cuda_time": 0.0441380126396125, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.0441380126396125, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2030.452, "cuda_time_us": 241.245, "pct_cuda_time": 0.7079837007475611, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.454, "cuda_time_us": 105.471, "pct_cuda_time": 0.30952661776014434, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.735, "pct_cuda_time": 0.307366672460759, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 493.587, "cuda_time_us": 18.816, "pct_cuda_time": 0.05521947113211095, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.816, "pct_cuda_time": 0.05521947113211095, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1015.464, "cuda_time_us": 33.407, "pct_cuda_time": 0.09803979975076693, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.552, "pct_cuda_time": 0.022162916984996915, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.159, "pct_cuda_time": 0.07089961751066479, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.00497726525510524, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 221.51, "cuda_time_us": 83.551, "pct_cuda_time": 0.24519781210453886, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.815, "pct_cuda_time": 0.24303786680515355, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.422, "cuda_time_us": 15.616, "pct_cuda_time": 0.04582840461304447, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.616, "pct_cuda_time": 0.04582840461304447, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 457.817, "cuda_time_us": 777.4289999999999, "pct_cuda_time": 2.2815273290160443, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.399, "cuda_time_us": 477.27299999999997, "pct_cuda_time": 1.4006570283607565, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.505, "pct_cuda_time": 1.3984031723961807, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.801, "cuda_time_us": 66.975, "pct_cuda_time": 0.19655208753577444, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.975, "pct_cuda_time": 0.19655208753577444, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.38, "cuda_time_us": 233.18099999999998, "pct_cuda_time": 0.6843182131195135, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.349, "pct_cuda_time": 0.6818765358245562, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2486.405, "cuda_time_us": 1046.225, "pct_cuda_time": 3.0703651777844803, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.669, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1771.905, "cuda_time_us": 238.14, "pct_cuda_time": 0.6988714315157794, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 152.866, "cuda_time_us": 105.566, "pct_cuda_time": 0.3098054150474291, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.83, "pct_cuda_time": 0.3076454697480438, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.486, "cuda_time_us": 18.784, "pct_cuda_time": 0.05512556046692029, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.784, "pct_cuda_time": 0.05512556046692029, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.376, "cuda_time_us": 33.022999999999996, "pct_cuda_time": 0.09691287176847896, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02159945299385293, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.128, "pct_cuda_time": 0.07080864155376133, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.535, "pct_cuda_time": 0.0045047772208647065, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 205.543, "cuda_time_us": 80.76700000000001, "pct_cuda_time": 0.23702758423295103, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.031, "pct_cuda_time": 0.23486763893356571, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 85.718, "cuda_time_us": 15.072, "pct_cuda_time": 0.04423192330480317, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.072, "pct_cuda_time": 0.04423192330480317, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 490.143, "cuda_time_us": 778.1329999999999, "pct_cuda_time": 2.283593363650239, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 179.764, "cuda_time_us": 478.61699999999996, "pct_cuda_time": 1.4046012762987643, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.881, "pct_cuda_time": 1.402441330999379, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.076, "cuda_time_us": 66.207, "pct_cuda_time": 0.19429823157119847, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.207, "pct_cuda_time": 0.19429823157119847, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 157.217, "cuda_time_us": 233.309, "pct_cuda_time": 0.6846938557802762, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.509, "pct_cuda_time": 0.6823460891505095, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2416.881, "cuda_time_us": 1047.2179999999998, "pct_cuda_time": 3.0732793431136782, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.146, "cuda_time_us": 14.304, "pct_cuda_time": 0.041978067340227215, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.304, "pct_cuda_time": 0.041978067340227215, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1737.06, "cuda_time_us": 239.326, "pct_cuda_time": 0.7023519955444084, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 162.944, "cuda_time_us": 105.79, "pct_cuda_time": 0.3104627897037638, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 105.022, "pct_cuda_time": 0.3082089337391878, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 511.793, "cuda_time_us": 18.752, "pct_cuda_time": 0.05503164980172963, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.752, "pct_cuda_time": 0.05503164980172963, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.769, "cuda_time_us": 33.535999999999994, "pct_cuda_time": 0.09841837711981681, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.712, "pct_cuda_time": 0.02263247031095024, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.576, "pct_cuda_time": 0.07212339086643065, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003662515942435931, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.199, "cuda_time_us": 81.24799999999999, "pct_cuda_time": 0.23843917891909816, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.511, "pct_cuda_time": 0.23627629891142565, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.403, "cuda_time_us": 15.52, "pct_cuda_time": 0.04554667261747247, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.52, "pct_cuda_time": 0.04554667261747247, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 454.311, "cuda_time_us": 778.068, "pct_cuda_time": 2.2834026076115705, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.985, "cuda_time_us": 476.953, "pct_cuda_time": 1.3997179217088498, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.767, "pct_cuda_time": 0.0022509212562887492, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.186, "pct_cuda_time": 1.3974670004525611, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.461, "cuda_time_us": 66.879, "pct_cuda_time": 0.19627035554020245, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.879, "pct_cuda_time": 0.19627035554020245, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 150.69, "cuda_time_us": 234.236, "pct_cuda_time": 0.6874143303625182, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.404, "pct_cuda_time": 0.6849726530675609, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2426.68, "cuda_time_us": 1049.071, "pct_cuda_time": 3.078717357569875, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.123, "cuda_time_us": 14.719, "pct_cuda_time": 0.04319597127941865, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.719, "pct_cuda_time": 0.04319597127941865, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1716.5, "cuda_time_us": 238.972, "pct_cuda_time": 0.7013131088107366, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.372, "cuda_time_us": 105.69500000000001, "pct_cuda_time": 0.310183992416479, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.959, "pct_cuda_time": 0.30802404711709364, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 471.93, "cuda_time_us": 18.943, "pct_cuda_time": 0.05559217908458642, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.943, "pct_cuda_time": 0.05559217908458642, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 752.975, "cuda_time_us": 33.12, "pct_cuda_time": 0.09719753847233817, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.52, "pct_cuda_time": 0.02206900631980625, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.32, "pct_cuda_time": 0.07137210554490533, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003756426607626596, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 196.247, "cuda_time_us": 81.214, "pct_cuda_time": 0.23833939883733307, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.478, "pct_cuda_time": 0.2361794535379478, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.557, "cuda_time_us": 15.711, "pct_cuda_time": 0.046107201900329256, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.711, "pct_cuda_time": 0.046107201900329256, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 484.656, "cuda_time_us": 779.669, "pct_cuda_time": 2.288101075579391, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.184, "cuda_time_us": 478.649, "pct_cuda_time": 1.4046951869639552, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0021570105910980845, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.914, "pct_cuda_time": 1.402538176372857, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.349, "cuda_time_us": 66.367, "pct_cuda_time": 0.19476778489715182, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.367, "pct_cuda_time": 0.19476778489715182, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 172.013, "cuda_time_us": 234.65300000000002, "pct_cuda_time": 0.6886381037182842, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.853, "pct_cuda_time": 0.6862903370885175, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2488.876, "cuda_time_us": 1046.513, "pct_cuda_time": 3.0712103737711964, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.898, "cuda_time_us": 14.399, "pct_cuda_time": 0.042256864627512, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.399, "pct_cuda_time": 0.042256864627512, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1802.2, "cuda_time_us": 238.429, "pct_cuda_time": 0.6997195622107826, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.79, "cuda_time_us": 105.055, "pct_cuda_time": 0.3083057791126657, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.287, "pct_cuda_time": 0.3060519231480897, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 541.346, "cuda_time_us": 18.72, "pct_cuda_time": 0.05493773913653896, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.72, "pct_cuda_time": 0.05493773913653896, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 760.125, "cuda_time_us": 33.28, "pct_cuda_time": 0.0976670917982915, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.52, "pct_cuda_time": 0.02206900631980625, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.16, "pct_cuda_time": 0.070902552218952, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004695533259533246, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 189.987, "cuda_time_us": 81.374, "pct_cuda_time": 0.2388089521632864, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 80.606, "pct_cuda_time": 0.23655509619871046, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.31, "cuda_time_us": 15.2, "pct_cuda_time": 0.044607565965565824, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.2, "pct_cuda_time": 0.044607565965565824, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 456.7, "cuda_time_us": 778.4849999999999, "pct_cuda_time": 2.284626380967336, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.432, "cuda_time_us": 478.073, "pct_cuda_time": 1.403004794990523, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.337, "pct_cuda_time": 1.4008448496911379, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.488, "cuda_time_us": 66.847, "pct_cuda_time": 0.19617644487501176, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.847, "pct_cuda_time": 0.19617644487501176, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.352, "cuda_time_us": 233.565, "pct_cuda_time": 0.6854451411018014, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.765, "pct_cuda_time": 0.6830973744720349, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2387.581, "cuda_time_us": 1046.8990000000001, "pct_cuda_time": 3.0723431711700595, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.469, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1725.127, "cuda_time_us": 237.82000000000002, "pct_cuda_time": 0.6979323248638728, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.172, "cuda_time_us": 105.311, "pct_cuda_time": 0.309057064434191, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.575, "pct_cuda_time": 0.3068971191348057, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 501.031, "cuda_time_us": 19.167, "pct_cuda_time": 0.056249553740921075, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.167, "pct_cuda_time": 0.056249553740921075, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 747.061, "cuda_time_us": 32.896, "pct_cuda_time": 0.09654016381600351, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.296, "pct_cuda_time": 0.021411631663471598, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.16, "pct_cuda_time": 0.070902552218952, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.00422597993357992, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 181.527, "cuda_time_us": 80.446, "pct_cuda_time": 0.23608554287275715, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 79.678, "pct_cuda_time": 0.23383168690818118, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.268, "cuda_time_us": 15.712, "pct_cuda_time": 0.046110136608616466, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.046110136608616466, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 451.023, "cuda_time_us": 778.4870000000001, "pct_cuda_time": 2.284632250383911, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 153.814, "cuda_time_us": 478.235, "pct_cuda_time": 1.4034802177330508, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.499, "pct_cuda_time": 1.4013202724336657, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.861, "cuda_time_us": 67.135, "pct_cuda_time": 0.19702164086172774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.135, "pct_cuda_time": 0.19702164086172774, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.015, "cuda_time_us": 233.11700000000002, "pct_cuda_time": 0.6841303917891322, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.317, "pct_cuda_time": 0.6817826251593656, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2348.758, "cuda_time_us": 1054.74, "pct_cuda_time": 3.0953542188500593, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.186, "cuda_time_us": 15.2, "pct_cuda_time": 0.044607565965565824, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.2, "pct_cuda_time": 0.044607565965565824, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1665.899, "cuda_time_us": 241.59799999999998, "pct_cuda_time": 0.7090196527729455, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.121, "cuda_time_us": 105.503, "pct_cuda_time": 0.309620528425335, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.767, "pct_cuda_time": 0.3074605831259497, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 470.383, "cuda_time_us": 19.264, "pct_cuda_time": 0.05653422044478027, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.264, "pct_cuda_time": 0.05653422044478027, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 710.187, "cuda_time_us": 33.568, "pct_cuda_time": 0.09851228778500747, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.84, "pct_cuda_time": 0.023008112971712902, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.448, "pct_cuda_time": 0.07174774820566798, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003756426607626596, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 191.073, "cuda_time_us": 83.263, "pct_cuda_time": 0.24435261611782289, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.527, "pct_cuda_time": 0.24219267081843757, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.842, "cuda_time_us": 15.456, "pct_cuda_time": 0.04535885128709115, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.456, "pct_cuda_time": 0.04535885128709115, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.503, "cuda_time_us": 782.486, "pct_cuda_time": 2.2963681488244565, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.995, "cuda_time_us": 481.018, "pct_cuda_time": 1.4116475108963515, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 480.282, "pct_cuda_time": 1.4094875655969663, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.829, "cuda_time_us": 67.039, "pct_cuda_time": 0.19673990886615575, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.039, "pct_cuda_time": 0.19673990886615575, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.076, "cuda_time_us": 234.429, "pct_cuda_time": 0.6879807290619494, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.629, "pct_cuda_time": 0.6856329624321829, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2745.229, "cuda_time_us": 1048.7839999999999, "pct_cuda_time": 3.0778750962914465, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.29, "cuda_time_us": 15.04, "pct_cuda_time": 0.0441380126396125, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.04, "pct_cuda_time": 0.0441380126396125, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 2050.494, "cuda_time_us": 240.572, "pct_cuda_time": 0.7060086420702699, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 151.656, "cuda_time_us": 105.694, "pct_cuda_time": 0.31018105770819177, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.926, "pct_cuda_time": 0.3079272017436158, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 549.114, "cuda_time_us": 18.367, "pct_cuda_time": 0.053901787111154446, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.367, "pct_cuda_time": 0.053901787111154446, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 947.902, "cuda_time_us": 32.544, "pct_cuda_time": 0.0955071464989062, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.36, "pct_cuda_time": 0.02159945299385293, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.904, "pct_cuda_time": 0.07015126689742668, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003756426607626596, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 226.158, "cuda_time_us": 83.967, "pct_cuda_time": 0.2464186507520175, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.231, "pct_cuda_time": 0.24425870545263217, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.272, "cuda_time_us": 15.679, "pct_cuda_time": 0.046013291235138594, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.679, "pct_cuda_time": 0.046013291235138594, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 467.765, "cuda_time_us": 777.4929999999999, "pct_cuda_time": 2.2817151503464257, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.517, "cuda_time_us": 477.818, "pct_cuda_time": 1.402256444377285, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.082, "pct_cuda_time": 1.4000964990779, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.941, "cuda_time_us": 66.526, "pct_cuda_time": 0.1952344035148179, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.526, "pct_cuda_time": 0.1952344035148179, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.95, "cuda_time_us": 233.149, "pct_cuda_time": 0.6842243024543229, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.801, "pct_cuda_time": 0.0023507013380538308, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.348, "pct_cuda_time": 0.681873601116269, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2429.305, "cuda_time_us": 1046.452, "pct_cuda_time": 3.071031356565677, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.452, "cuda_time_us": 14.72, "pct_cuda_time": 0.04319890598770586, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.72, "pct_cuda_time": 0.04319890598770586, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1756.897, "cuda_time_us": 239.358, "pct_cuda_time": 0.7024459062095991, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.259, "cuda_time_us": 105.15100000000001, "pct_cuda_time": 0.3085875111082377, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.415, "pct_cuda_time": 0.3064275658088524, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.189, "cuda_time_us": 19.648, "pct_cuda_time": 0.057661148427068244, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.648, "pct_cuda_time": 0.057661148427068244, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 746.074, "cuda_time_us": 33.888, "pct_cuda_time": 0.09945139443691413, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.968, "pct_cuda_time": 0.02338375563247556, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.224, "pct_cuda_time": 0.07109037354933333, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.00497726525510524, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 188.649, "cuda_time_us": 80.671, "pct_cuda_time": 0.23674585223737904, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 79.935, "pct_cuda_time": 0.23458590693799372, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.004, "cuda_time_us": 15.712, "pct_cuda_time": 0.046110136608616466, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.046110136608616466, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.864, "cuda_time_us": 776.662, "pct_cuda_time": 2.2792764077597556, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.534, "cuda_time_us": 477.402, "pct_cuda_time": 1.4010356057298063, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 476.634, "pct_cuda_time": 1.3987817497652304, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.749, "cuda_time_us": 66.431, "pct_cuda_time": 0.19495560622753313, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.431, "pct_cuda_time": 0.19495560622753313, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 149.526, "cuda_time_us": 232.829, "pct_cuda_time": 0.6832851958024162, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.029, "pct_cuda_time": 0.6809374291726495, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2422.933, "cuda_time_us": 1052.2720000000002, "pct_cuda_time": 3.0881113587972298, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.887, "cuda_time_us": 14.528, "pct_cuda_time": 0.042635441996561864, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.528, "pct_cuda_time": 0.042635441996561864, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1719.823, "cuda_time_us": 241.43400000000003, "pct_cuda_time": 0.7085383606138436, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.956, "cuda_time_us": 105.662, "pct_cuda_time": 0.3100871470430011, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.926, "pct_cuda_time": 0.3079272017436158, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.641, "cuda_time_us": 19.103, "pct_cuda_time": 0.05606173241053974, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.103, "pct_cuda_time": 0.05606173241053974, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 716.941, "cuda_time_us": 32.862, "pct_cuda_time": 0.09644038373423844, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.423, "pct_cuda_time": 0.02178433961594705, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.967, "pct_cuda_time": 0.0703361535195208, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 206.399, "cuda_time_us": 83.807, "pct_cuda_time": 0.24594909742606416, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.071, "pct_cuda_time": 0.24378915212667887, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 89.146, "cuda_time_us": 14.976, "pct_cuda_time": 0.043950191309231175, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.976, "pct_cuda_time": 0.043950191309231175, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.023, "cuda_time_us": 781.3340000000001, "pct_cuda_time": 2.2929873648775927, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.955, "cuda_time_us": 479.99399999999997, "pct_cuda_time": 1.4086423696102501, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 479.226, "pct_cuda_time": 1.4063885136456744, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.691, "cuda_time_us": 66.943, "pct_cuda_time": 0.19645817687058376, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.943, "pct_cuda_time": 0.19645817687058376, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 152.496, "cuda_time_us": 234.397, "pct_cuda_time": 0.6878868183967588, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.565, "pct_cuda_time": 0.6854451411018014, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2565.48, "cuda_time_us": 1050.548, "pct_cuda_time": 3.0830519217100822, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.244, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.88, "pct_cuda_time": 0.04366845931365918, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1880.551, "cuda_time_us": 240.51, "pct_cuda_time": 0.705826690156463, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 153.516, "cuda_time_us": 105.856, "pct_cuda_time": 0.3106564804507195, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 105.119, "pct_cuda_time": 0.30849360044304697, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 538.979, "cuda_time_us": 18.783, "pct_cuda_time": 0.05512262575863309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 18.783, "pct_cuda_time": 0.05512262575863309, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 835.703, "cuda_time_us": 32.704, "pct_cuda_time": 0.09597669982485954, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.392, "pct_cuda_time": 0.021693363659043594, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.064, "pct_cuda_time": 0.07062082022338001, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.003662515942435931, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 201.783, "cuda_time_us": 83.167, "pct_cuda_time": 0.2440708841222509, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.431, "pct_cuda_time": 0.24191093882286557, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.143, "cuda_time_us": 15.391, "pct_cuda_time": 0.04516809524842261, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.391, "pct_cuda_time": 0.04516809524842261, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.979, "cuda_time_us": 779.767, "pct_cuda_time": 2.288388676991538, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 159.94, "cuda_time_us": 478.71500000000003, "pct_cuda_time": 1.404888877710911, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.978, "pct_cuda_time": 1.4027259977032385, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.697, "cuda_time_us": 67.135, "pct_cuda_time": 0.19702164086172774, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.135, "pct_cuda_time": 0.19702164086172774, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 153.163, "cuda_time_us": 233.917, "pct_cuda_time": 0.6864781584188988, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 233.117, "pct_cuda_time": 0.6841303917891322, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2373.833, "cuda_time_us": 1049.267, "pct_cuda_time": 3.0792925603941685, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.078, "cuda_time_us": 14.847, "pct_cuda_time": 0.0435716139401813, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.847, "pct_cuda_time": 0.0435716139401813, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1687.249, "cuda_time_us": 240.31799999999998, "pct_cuda_time": 0.7052632261653189, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 149.996, "cuda_time_us": 106.07900000000001, "pct_cuda_time": 0.31131092039876695, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 105.343, "pct_cuda_time": 0.30915097509938166, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.759, "cuda_time_us": 19.008, "pct_cuda_time": 0.05578293512325495, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.008, "pct_cuda_time": 0.05578293512325495, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 703.041, "cuda_time_us": 33.056, "pct_cuda_time": 0.09700971714195684, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.648, "pct_cuda_time": 0.022444648980568912, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 24.096, "pct_cuda_time": 0.07071473088857066, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.0038503372728172615, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 186.52, "cuda_time_us": 82.175, "pct_cuda_time": 0.24115965350134022, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 81.439, "pct_cuda_time": 0.23899970820195496, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.59, "cuda_time_us": 15.904, "pct_cuda_time": 0.04667360059976046, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.904, "pct_cuda_time": 0.04667360059976046, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 471.94, "cuda_time_us": 778.198, "pct_cuda_time": 2.2837841196889075, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 174.155, "cuda_time_us": 477.75399999999996, "pct_cuda_time": 1.4020686230469035, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 477.018, "pct_cuda_time": 1.3999086777475185, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.605, "cuda_time_us": 67.199, "pct_cuda_time": 0.19720946219210908, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 67.199, "pct_cuda_time": 0.19720946219210908, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.528, "cuda_time_us": 233.245, "pct_cuda_time": 0.6845060344498949, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.413, "pct_cuda_time": 0.6820643571549376, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2426.364, "cuda_time_us": 1047.8899999999999, "pct_cuda_time": 3.0752514670826825, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.105, "cuda_time_us": 14.4, "pct_cuda_time": 0.04225979933579921, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.4, "pct_cuda_time": 0.04225979933579921, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1689.2, "cuda_time_us": 237.66000000000003, "pct_cuda_time": 0.6974627715379195, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 147.747, "cuda_time_us": 105.182, "pct_cuda_time": 0.3086784870651411, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 104.414, "pct_cuda_time": 0.30642463110056517, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[1536, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.941, "cuda_time_us": 19.2, "pct_cuda_time": 0.05634639911439894, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 19.2, "pct_cuda_time": 0.05634639911439894, "trace": "_C::rotary_embedding(int64[1536], bfloat16[1536, 4096], bfloat16[1536, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 714.276, "cuda_time_us": 32.864, "pct_cuda_time": 0.09644625315081284, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 7.616, "pct_cuda_time": 0.022350738315378246, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[1536], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, false, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 23.776, "pct_cuda_time": 0.06977562423666402, "trace": "_vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_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": 0.004319890598770585, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], None, None, bfloat16[1536, 32, 128], int32[7], int32[7], 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[1536, 32, 128], bfloat16[1536, 8, 128], bfloat16[1536, 8, 128], bfloat16[1536, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 204.614, "cuda_time_us": 80.414, "pct_cuda_time": 0.23599163220756647, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.0021570105910980845, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize128x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 79.679, "pct_cuda_time": 0.2338346216164684, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[1536, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 109.315, "cuda_time_us": 15.328, "pct_cuda_time": 0.044983208626328486, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 15.328, "pct_cuda_time": 0.044983208626328486, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 485.639, "cuda_time_us": 780.502, "pct_cuda_time": 2.2905456875826355, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.163, "cuda_time_us": 480.53799999999995, "pct_cuda_time": 1.4102388509184915, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 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": 479.77, "pct_cuda_time": 1.4079849949539156, "trace": "mm(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[1536, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[1536, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.705, "cuda_time_us": 66.495, "pct_cuda_time": 0.19514342755791447, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 66.495, "pct_cuda_time": 0.19514342755791447, "trace": "_C::silu_and_mul(bfloat16[1536, 14336], bfloat16[1536, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 158.556, "cuda_time_us": 233.469, "pct_cuda_time": 0.6851634091062294, "trace": "" }, "children": [ { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.832, "pct_cuda_time": 0.002441677294957287, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize256x128x64_warpgroupsize2x1x1_execute_segment_k_on_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 232.637, "pct_cuda_time": 0.6827217318112722, "trace": "mm(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[1536, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[1536, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.184, "cuda_time_us": 14.656, "pct_cuda_time": 0.043011084657324526, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 14.656, "pct_cuda_time": 0.043011084657324526, "trace": "_C::fused_add_rms_norm(bfloat16[1536, 4096], bfloat16[1536, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 471.522, "cuda_time_us": 361.02, "pct_cuda_time": 1.0594883858479325, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 4.801, "pct_cuda_time": 0.014089534486886945, "trace": "index_select(bfloat16[1536, 4096], 0, int64[6])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 355.483, "pct_cuda_time": 1.0432389060616605, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3389.325, "cuda_time_us": 115.038, "pct_cuda_time": 0.3376029719438659, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.737, "pct_cuda_time": 0.002162880007672501, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.0021599452993852925, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "copy_(int32[6], int32[6], True) <- _to_copy(int32[6], 3, 0, None, None, True, None) <- to(int32[6], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.002347766629766623, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.0022538559645759577, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.48, "pct_cuda_time": 0.013147493126693086, "trace": "copy_(float32[6, 128256], bfloat16[6, 128256], False) <- _to_copy(bfloat16[6, 128256], 6, None, None, None, False, None) <- to(bfloat16[6, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 5.503, "pct_cuda_time": 0.016149699704507154, "trace": "div_(float32[6, 128256], bfloat16[6, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 34.495, "pct_cuda_time": 0.10123276236724955, "trace": "_softmax(float32[6, 128256], -1, False) <- softmax(float32[6, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 27.295, "pct_cuda_time": 0.08010286269934996, "trace": "_log_softmax(float32[6, 128256], -1, False) <- log_softmax(float32[6, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.824, "pct_cuda_time": 0.005352907915867899, "trace": "copy_(int64[6], int32[6], False) <- _to_copy(int32[6], 4, None, None, None, False, None) <- to(int32[6], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 5.856, "pct_cuda_time": 0.017185651729891678, "trace": "index(float32[6, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 27.616, "pct_cuda_time": 0.08104490405954382, "trace": "argmax(float32[6, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.007418942550062527, "trace": "copy_(int64[6], int64[6], False) <- _to_copy(int64[6], 4, 0, None, None, False, None) <- to(int64[6], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] }, "decode_1": { "metadata": { "num_running_seqs": 6 }, "summary_stats": [ { "entry": { "name": "LlamaForCausalLM", "cuda_time_us": 6442.813, "pct_cuda_time": 93.26521093633288, "invocations": 1 }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cuda_time_us": 6.175, "pct_cuda_time": 0.08938838943980766, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 6.175, "pct_cuda_time": 0.08938838943980766, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cuda_time_us": 6433.278, "pct_cuda_time": 93.12718368235579, "invocations": 32 }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 212.51200000000009, "pct_cuda_time": 3.0762923751631437, "invocations": 64 }, "children": [ { "entry": { "name": "void vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, int)", "cuda_time_us": 4.223, "pct_cuda_time": 0.06113152528004984, "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": 208.28900000000007, "pct_cuda_time": 3.015160849883094, "invocations": 63 }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cuda_time_us": 1889.6299999999997, "pct_cuda_time": 27.354005236784406, "invocations": 32 }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cuda_time_us": 672.1489999999999, "pct_cuda_time": 9.729929809486197, "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": 672.1489999999999, "pct_cuda_time": 9.729929809486197, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cuda_time_us": 116.54100000000001, "pct_cuda_time": 1.687030330964312, "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": 116.54100000000001, "pct_cuda_time": 1.687030330964312, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "Attention", "cuda_time_us": 531.0980000000001, "pct_cuda_time": 7.688096332745421, "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": 75.039, "pct_cuda_time": 1.086253498813559, "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": 408.22100000000006, "pct_cuda_time": 5.9093470000822235, "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": 47.838, "pct_cuda_time": 0.6924958338496388, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cuda_time_us": 569.8419999999999, "pct_cuda_time": 8.24894876358848, "invocations": 32 }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cuda_time_us": 501.7509999999999, "pct_cuda_time": 7.263273488228814, "invocations": 32 }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cuda_time_us": 68.09100000000002, "pct_cuda_time": 0.9856752753596673, "invocations": 32 }, "children": [] } ] } ] }, { "entry": { "name": "LlamaMLP", "cuda_time_us": 4331.136, "pct_cuda_time": 62.69688607040823, "invocations": 32 }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cuda_time_us": 2657.0509999999995, "pct_cuda_time": 38.463078469543376, "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": 2657.0509999999995, "pct_cuda_time": 38.463078469543376, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cuda_time_us": 285.1180000000001, "pct_cuda_time": 4.127326124744792, "invocations": 32 }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cuda_time_us": 285.1180000000001, "pct_cuda_time": 4.127326124744792, "invocations": 32 }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cuda_time_us": 1388.9669999999999, "pct_cuda_time": 20.10648147612005, "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": 1388.9669999999999, "pct_cuda_time": 20.10648147612005, "invocations": 32 }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "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": 3.36, "pct_cuda_time": 0.04863886453728806, "invocations": 1 }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cuda_time_us": 347.19599999999997, "pct_cuda_time": 5.025958098776268, "invocations": 1 }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cuda_time_us": 4.673, "pct_cuda_time": 0.06764565892343663, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memset (Device)", "cuda_time_us": 0.736, "pct_cuda_time": 0.010654227470072622, "invocations": 1 }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cuda_time_us": 341.787, "pct_cuda_time": 4.947658212382759, "invocations": 1 }, "children": [] } ] }, { "entry": { "name": "Sampler", "cuda_time_us": 118.04700000000001, "pct_cuda_time": 1.7088309648908464, "invocations": 1 }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cuda_time_us": 5.4399999999999995, "pct_cuda_time": 0.07874863782227588, "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": 4.288, "pct_cuda_time": 0.06207245569520571, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cuda_time_us": 5.568, "pct_cuda_time": 0.08060154694750592, "invocations": 1 }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cuda_time_us": 35.616, "pct_cuda_time": 0.5155719640952534, "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": 28.16, "pct_cuda_time": 0.4076400075506047, "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": 1.728, "pct_cuda_time": 0.02501427319060529, "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": 5.983, "pct_cuda_time": 0.08660902575196262, "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": 28.608, "pct_cuda_time": 0.4141251894889097, "invocations": 1 }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cuda_time_us": 2.656, "pct_cuda_time": 0.03844786434852294, "invocations": 1 }, "children": [] } ] } ], "model_stats": [ { "entry": { "name": "LlamaForCausalLM", "cpu_time_us": 82646.489, "cuda_time_us": 6442.813, "pct_cuda_time": 93.26521093633288, "trace": "" }, "children": [ { "entry": { "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", "cpu_time_us": 313.016, "cuda_time_us": 6.175, "pct_cuda_time": 0.08938838943980766, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 6.175, "pct_cuda_time": 0.08938838943980766, "trace": "index_select(bfloat16[128256, 4096], 0, int64[6]) <- embedding(bfloat16[128256, 4096], int64[6], -1, False, False)" }, "children": [] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 4292.328, "cuda_time_us": 208.891, "pct_cuda_time": 3.0238753131126903, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 282.746, "cuda_time_us": 4.223, "pct_cuda_time": 0.06113152528004984, "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": 4.223, "pct_cuda_time": 0.06113152528004984, "trace": "_C::rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 3149.864, "cuda_time_us": 64.89500000000001, "pct_cuda_time": 0.9394104506390799, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 564.213, "cuda_time_us": 25.631, "pct_cuda_time": 0.3710305764747709, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 25.631, "pct_cuda_time": 0.3710305764747709, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 919.933, "cuda_time_us": 3.424, "pct_cuda_time": 0.04956531909990307, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.424, "pct_cuda_time": 0.04956531909990307, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 1085.682, "cuda_time_us": 16.896, "pct_cuda_time": 0.2445840045303628, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.176, "pct_cuda_time": 0.031499455128910364, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.024, "pct_cuda_time": 0.18853350349215464, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.696, "pct_cuda_time": 0.024551045909297776, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 283.788, "cuda_time_us": 18.944000000000003, "pct_cuda_time": 0.27423055053404316, "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": 16.832, "pct_cuda_time": 0.2436575499677478, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 115.837, "cuda_time_us": 3.359, "pct_cuda_time": 0.048624388684747194, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.359, "pct_cuda_time": 0.048624388684747194, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 600.151, "cuda_time_us": 136.414, "pct_cuda_time": 1.9747089485088132, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 219.615, "cuda_time_us": 83.615, "pct_cuda_time": 1.2103984102039704, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.615, "pct_cuda_time": 1.2103984102039704, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 140.003, "cuda_time_us": 8.832, "pct_cuda_time": 0.12785072964087146, "trace": "" }, "children": [ { "entry": { "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": 0.12785072964087146, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 165.527, "cuda_time_us": 43.967, "pct_cuda_time": 0.6364598086639713, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.967, "pct_cuda_time": 0.6364598086639713, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2669.861, "cuda_time_us": 200.28500000000003, "pct_cuda_time": 2.8992961261460537, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.984, "cuda_time_us": 3.2, "pct_cuda_time": 0.04632272813075053, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04632272813075053, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1894.837, "cuda_time_us": 59.071, "pct_cuda_time": 0.8551030854411138, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 150.528, "cuda_time_us": 21.12, "pct_cuda_time": 0.3057300056629535, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.12, "pct_cuda_time": 0.3057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 583.521, "cuda_time_us": 3.776, "pct_cuda_time": 0.05466081919428562, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05466081919428562, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 805.513, "cuda_time_us": 16.543, "pct_cuda_time": 0.23947402858343936, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.799, "pct_cuda_time": 0.18527643667046123, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 184.079, "cuda_time_us": 17.631999999999998, "pct_cuda_time": 0.2552382320004354, "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": 0.22466523143414005, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 88.157, "cuda_time_us": 3.392, "pct_cuda_time": 0.04910209181859555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04910209181859555, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 508.312, "cuda_time_us": 134.622, "pct_cuda_time": 1.9487682207555934, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 184.511, "cuda_time_us": 82.815, "pct_cuda_time": 1.1988177281712828, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.815, "pct_cuda_time": 1.1988177281712828, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.912, "cuda_time_us": 8.704, "pct_cuda_time": 0.12599782051564146, "trace": "" }, "children": [ { "entry": { "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": 0.12599782051564146, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.255, "cuda_time_us": 43.103, "pct_cuda_time": 0.6239526720686688, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6239526720686688, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2561.779, "cuda_time_us": 201.726, "pct_cuda_time": 2.9201558296574315, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.106, "cuda_time_us": 3.169, "pct_cuda_time": 0.045873976701983885, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.169, "pct_cuda_time": 0.045873976701983885, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1819.167, "cuda_time_us": 58.975, "pct_cuda_time": 0.8537134035971915, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.944, "cuda_time_us": 20.896, "pct_cuda_time": 0.30248741469380097, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.896, "pct_cuda_time": 0.30248741469380097, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 537.452, "cuda_time_us": 3.615, "pct_cuda_time": 0.052330206935207235, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.615, "pct_cuda_time": 0.052330206935207235, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 734.191, "cuda_time_us": 16.96, "pct_cuda_time": 0.24551045909297783, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.56, "pct_cuda_time": 0.037058182504600426, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.832, "pct_cuda_time": 0.18575413980430963, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.02269813678406776, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 241.47, "cuda_time_us": 17.503999999999998, "pct_cuda_time": 0.2533853228752054, "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": 0.221885867746295, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.176, "pct_cuda_time": 0.031499455128910364, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.281, "cuda_time_us": 3.553, "pct_cuda_time": 0.051432704077673946, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.553, "pct_cuda_time": 0.051432704077673946, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 484.903, "cuda_time_us": 136.029, "pct_cuda_time": 1.9691357452805824, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.265, "cuda_time_us": 83.743, "pct_cuda_time": 1.2122513193292004, "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": 83.743, "pct_cuda_time": 1.2122513193292004, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 105.346, "cuda_time_us": 8.991, "pct_cuda_time": 0.13015239019486813, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.991, "pct_cuda_time": 0.13015239019486813, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.704, "cuda_time_us": 43.295, "pct_cuda_time": 0.6267320357565138, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.295, "pct_cuda_time": 0.6267320357565138, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2414.204, "cuda_time_us": 201.308, "pct_cuda_time": 2.914104923295352, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.822, "cuda_time_us": 3.265, "pct_cuda_time": 0.047263658545906403, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.265, "pct_cuda_time": 0.047263658545906403, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1736.447, "cuda_time_us": 58.653999999999996, "pct_cuda_time": 0.8490666549315753, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.789, "cuda_time_us": 20.8, "pct_cuda_time": 0.30109773284987845, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30109773284987845, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.457, "cuda_time_us": 3.68, "pct_cuda_time": 0.05327113735036311, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05327113735036311, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 742.268, "cuda_time_us": 16.544, "pct_cuda_time": 0.23948850443598024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.0347420460980629, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.672, "pct_cuda_time": 0.1834380033977721, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.783, "cuda_time_us": 17.63, "pct_cuda_time": 0.25520928029535367, "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": 15.487, "pct_cuda_time": 0.2241875283002917, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.143, "pct_cuda_time": 0.031021751995061995, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.814, "cuda_time_us": 3.328, "pct_cuda_time": 0.04817563725598055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04817563725598055, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 459.519, "cuda_time_us": 136.06099999999998, "pct_cuda_time": 1.9695989725618896, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 160.87, "cuda_time_us": 83.294, "pct_cuda_time": 1.2057516615383546, "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": 83.294, "pct_cuda_time": 1.2057516615383546, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.367, "cuda_time_us": 8.928, "pct_cuda_time": 0.12924041148479398, "trace": "" }, "children": [ { "entry": { "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": 0.12924041148479398, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.528, "cuda_time_us": 43.839, "pct_cuda_time": 0.6346068995387414, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.6346068995387414, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2412.72, "cuda_time_us": 201.437, "pct_cuda_time": 2.9159723082731235, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.594, "cuda_time_us": 3.168, "pct_cuda_time": 0.04585950084944303, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04585950084944303, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1702.473, "cuda_time_us": 59.264, "pct_cuda_time": 0.8578969249814999, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.876, "cuda_time_us": 21.439, "pct_cuda_time": 0.3103478026234877, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.439, "pct_cuda_time": 0.3103478026234877, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.213, "cuda_time_us": 3.585, "pct_cuda_time": 0.051895931358981455, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.585, "pct_cuda_time": 0.051895931358981455, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 741.061, "cuda_time_us": 16.576, "pct_cuda_time": 0.23995173171728776, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.736, "pct_cuda_time": 0.1843644579603871, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.02177168222145275, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 163.255, "cuda_time_us": 17.664, "pct_cuda_time": 0.2557014592817429, "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": 0.22512845871544757, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.786, "cuda_time_us": 3.295, "pct_cuda_time": 0.047697934122132184, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.047697934122132184, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 487.03, "cuda_time_us": 135.71, "pct_cuda_time": 1.9645179483200481, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.705, "cuda_time_us": 83.775, "pct_cuda_time": 1.212714546610508, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.775, "pct_cuda_time": 1.212714546610508, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 101.375, "cuda_time_us": 8.928, "pct_cuda_time": 0.12924041148479398, "trace": "" }, "children": [ { "entry": { "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": 0.12924041148479398, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 151.672, "cuda_time_us": 43.007, "pct_cuda_time": 0.6225629902247463, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.007, "pct_cuda_time": 0.6225629902247463, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2664.938, "cuda_time_us": 201.406, "pct_cuda_time": 2.9155235568443567, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.709, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1881.878, "cuda_time_us": 60.352000000000004, "pct_cuda_time": 0.8736466525459551, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.824, "cuda_time_us": 20.992, "pct_cuda_time": 0.3038770965377235, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.992, "pct_cuda_time": 0.3038770965377235, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 534.717, "cuda_time_us": 3.68, "pct_cuda_time": 0.05327113735036311, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05327113735036311, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 877.847, "cuda_time_us": 16.736, "pct_cuda_time": 0.24226786812382525, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.96, "pct_cuda_time": 0.18760704892953967, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 165.621, "cuda_time_us": 18.944000000000003, "pct_cuda_time": 0.27423055053404316, "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": 16.8, "pct_cuda_time": 0.24319432268644028, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.144, "pct_cuda_time": 0.031036227847602856, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 81.492, "cuda_time_us": 3.392, "pct_cuda_time": 0.04910209181859555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04910209181859555, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 507.393, "cuda_time_us": 134.398, "pct_cuda_time": 1.9455256297864403, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 194.96, "cuda_time_us": 82.783, "pct_cuda_time": 1.1983545008899754, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.783, "pct_cuda_time": 1.1983545008899754, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.684, "cuda_time_us": 8.768, "pct_cuda_time": 0.12692427507825646, "trace": "" }, "children": [ { "entry": { "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": 0.12692427507825646, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.226, "cuda_time_us": 42.847, "pct_cuda_time": 0.6202468538182088, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6202468538182088, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2366.36, "cuda_time_us": 200.54000000000002, "pct_cuda_time": 2.9029874685439725, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.99, "cuda_time_us": 3.424, "pct_cuda_time": 0.04956531909990307, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04956531909990307, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1692.562, "cuda_time_us": 58.559, "pct_cuda_time": 0.8476914489401938, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.246, "cuda_time_us": 20.479, "pct_cuda_time": 0.2964509841842625, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.479, "pct_cuda_time": 0.2964509841842625, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 503.464, "cuda_time_us": 3.552, "pct_cuda_time": 0.05141822822513309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05141822822513309, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.68, "cuda_time_us": 16.767000000000003, "pct_cuda_time": 0.24271661955259194, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.367, "pct_cuda_time": 0.03426434296421453, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.832, "pct_cuda_time": 0.18575413980430963, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.02269813678406776, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 158.607, "cuda_time_us": 17.761, "pct_cuda_time": 0.2571056169782063, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.22651814055937008, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030587476418836208, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.254, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.712, "cuda_time_us": 135.293, "pct_cuda_time": 1.95848151781051, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 162.399, "cuda_time_us": 83.518, "pct_cuda_time": 1.208994252507507, "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": 83.518, "pct_cuda_time": 1.208994252507507, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.716, "cuda_time_us": 8.544, "pct_cuda_time": 0.12368168410910392, "trace": "" }, "children": [ { "entry": { "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": 0.12368168410910392, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.127, "cuda_time_us": 43.231, "pct_cuda_time": 0.6258055811938988, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6258055811938988, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2421.012, "cuda_time_us": 200.38200000000003, "pct_cuda_time": 2.900700283842517, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.513, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1718.712, "cuda_time_us": 58.783, "pct_cuda_time": 0.8509340399093464, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.752, "cuda_time_us": 20.991, "pct_cuda_time": 0.3038626206851826, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.991, "pct_cuda_time": 0.3038626206851826, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.22, "cuda_time_us": 3.712, "pct_cuda_time": 0.05373436463167061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05373436463167061, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 757.535, "cuda_time_us": 16.416, "pct_cuda_time": 0.23763559531075024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.64, "pct_cuda_time": 0.1829747761164646, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.888, "cuda_time_us": 17.664, "pct_cuda_time": 0.2557014592817429, "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": 0.22512845871544757, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.471, "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04863886453728806, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.299, "cuda_time_us": 134.97500000000002, "pct_cuda_time": 1.9538781967025167, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.581, "cuda_time_us": 82.399, "pct_cuda_time": 1.1927957735142853, "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": 82.399, "pct_cuda_time": 1.1927957735142853, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.41, "cuda_time_us": 8.608, "pct_cuda_time": 0.12460813867171894, "trace": "" }, "children": [ { "entry": { "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": 0.12460813867171894, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.747, "cuda_time_us": 43.968, "pct_cuda_time": 0.6364742845165123, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.968, "pct_cuda_time": 0.6364742845165123, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2325.067, "cuda_time_us": 200.79700000000003, "pct_cuda_time": 2.9067077626469735, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.954, "cuda_time_us": 3.488, "pct_cuda_time": 0.05049177366251808, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.05049177366251808, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1648.988, "cuda_time_us": 58.879000000000005, "pct_cuda_time": 0.8523237217532689, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.456, "cuda_time_us": 20.863, "pct_cuda_time": 0.3020097115599526, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.863, "pct_cuda_time": 0.3020097115599526, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.451, "cuda_time_us": 3.552, "pct_cuda_time": 0.05141822822513309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05141822822513309, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 726.487, "cuda_time_us": 16.928, "pct_cuda_time": 0.24504723181167032, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 13.12, "pct_cuda_time": 0.18992318533607716, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.083, "cuda_time_us": 17.536, "pct_cuda_time": 0.2538485501565129, "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": 0.22327554959021756, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.574, "cuda_time_us": 3.231, "pct_cuda_time": 0.04677147955951717, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04677147955951717, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 455.584, "cuda_time_us": 135.199, "pct_cuda_time": 1.9571207876716692, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 156.992, "cuda_time_us": 83.167, "pct_cuda_time": 1.2039132282656655, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.167, "pct_cuda_time": 1.2039132282656655, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.963, "cuda_time_us": 8.768, "pct_cuda_time": 0.12692427507825646, "trace": "" }, "children": [ { "entry": { "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": 0.12692427507825646, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 142.688, "cuda_time_us": 43.264, "pct_cuda_time": 0.6262832843277472, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6262832843277472, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2436.975, "cuda_time_us": 200.191, "pct_cuda_time": 2.8979353960072123, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.49, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1750.132, "cuda_time_us": 58.592, "pct_cuda_time": 0.8481691520740422, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 225.661, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 477.999, "cuda_time_us": 3.616, "pct_cuda_time": 0.0523446827877481, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0523446827877481, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.942, "cuda_time_us": 16.32, "pct_cuda_time": 0.2362459134668277, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.576, "pct_cuda_time": 0.1820483215538496, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.344, "cuda_time_us": 17.792, "pct_cuda_time": 0.257554368406973, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.68, "pct_cuda_time": 0.2269813678406776, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.85, "cuda_time_us": 3.392, "pct_cuda_time": 0.04910209181859555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04910209181859555, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 460.661, "cuda_time_us": 134.911, "pct_cuda_time": 1.9529517421399016, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.815, "cuda_time_us": 82.911, "pct_cuda_time": 1.2002074100152054, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.911, "pct_cuda_time": 1.2002074100152054, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.386, "cuda_time_us": 8.768, "pct_cuda_time": 0.12692427507825646, "trace": "" }, "children": [ { "entry": { "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": 0.12692427507825646, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.705, "cuda_time_us": 43.232, "pct_cuda_time": 0.6258200570464396, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6258200570464396, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2394.894, "cuda_time_us": 201.21200000000002, "pct_cuda_time": 2.91271524145143, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.489, "cuda_time_us": 3.359, "pct_cuda_time": 0.048624388684747194, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.359, "pct_cuda_time": 0.048624388684747194, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1643.526, "cuda_time_us": 58.815, "pct_cuda_time": 0.8513972671906539, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 154.393, "cuda_time_us": 20.64, "pct_cuda_time": 0.2987815964433409, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2987815964433409, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 456.143, "cuda_time_us": 3.616, "pct_cuda_time": 0.0523446827877481, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0523446827877481, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 716.41, "cuda_time_us": 16.672, "pct_cuda_time": 0.24134141356121028, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.768, "pct_cuda_time": 0.18482768524169463, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.568, "pct_cuda_time": 0.02269813678406776, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 162.565, "cuda_time_us": 17.887, "pct_cuda_time": 0.25892957439835457, "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": 0.22650366470682923, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.24, "pct_cuda_time": 0.032425909691525374, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 100.893, "cuda_time_us": 3.232, "pct_cuda_time": 0.04678595541205804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04678595541205804, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 506.345, "cuda_time_us": 135.806, "pct_cuda_time": 1.965907630163971, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 172.504, "cuda_time_us": 82.815, "pct_cuda_time": 1.1988177281712828, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.815, "pct_cuda_time": 1.1988177281712828, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 124.984, "cuda_time_us": 8.992, "pct_cuda_time": 0.130166866047409, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.992, "pct_cuda_time": 0.130166866047409, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.676, "cuda_time_us": 43.999, "pct_cuda_time": 0.636923035945279, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.999, "pct_cuda_time": 0.636923035945279, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2475.111, "cuda_time_us": 200.925, "pct_cuda_time": 2.9085606717722037, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.858, "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04863886453728806, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1789.961, "cuda_time_us": 58.494, "pct_cuda_time": 0.8467505185250379, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.294, "cuda_time_us": 20.8, "pct_cuda_time": 0.30109773284987845, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30109773284987845, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 502.813, "cuda_time_us": 3.647, "pct_cuda_time": 0.052793434216514744, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.647, "pct_cuda_time": 0.052793434216514744, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 731.463, "cuda_time_us": 16.512, "pct_cuda_time": 0.23902527715467273, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.736, "pct_cuda_time": 0.1843644579603871, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 159.3, "cuda_time_us": 17.535, "pct_cuda_time": 0.25383407430397203, "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": 15.423, "pct_cuda_time": 0.22326107373767667, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.953, "cuda_time_us": 3.328, "pct_cuda_time": 0.04817563725598055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04817563725598055, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.504, "cuda_time_us": 135.743, "pct_cuda_time": 1.9649956514538967, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.505, "cuda_time_us": 83.103, "pct_cuda_time": 1.2029867737030502, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.103, "pct_cuda_time": 1.2029867737030502, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 98.394, "cuda_time_us": 9.152, "pct_cuda_time": 0.1324830024539465, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.152, "pct_cuda_time": 0.1324830024539465, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.684, "cuda_time_us": 43.488, "pct_cuda_time": 0.6295258752968996, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.488, "pct_cuda_time": 0.6295258752968996, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2619.405, "cuda_time_us": 200.82799999999997, "pct_cuda_time": 2.9071565140757394, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.505, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1908.218, "cuda_time_us": 58.910999999999994, "pct_cuda_time": 0.8527869490345763, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 139.085, "cuda_time_us": 20.736, "pct_cuda_time": 0.3001712782872634, "trace": "" }, "children": [ { "entry": { "name": "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": 0.3001712782872634, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 479.015, "cuda_time_us": 3.712, "pct_cuda_time": 0.05373436463167061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05373436463167061, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 957.862, "cuda_time_us": 16.703, "pct_cuda_time": 0.2417901649899769, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.895, "pct_cuda_time": 0.18666611851438378, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 182.202, "cuda_time_us": 17.759999999999998, "pct_cuda_time": 0.2570911411256654, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.22651814055937008, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.509, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 480.339, "cuda_time_us": 135.325, "pct_cuda_time": 1.958944745091817, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 175.988, "cuda_time_us": 83.422, "pct_cuda_time": 1.2076045706635845, "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": 83.422, "pct_cuda_time": 1.2076045706635845, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.346, "cuda_time_us": 8.864, "pct_cuda_time": 0.12831395692217898, "trace": "" }, "children": [ { "entry": { "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": 0.12831395692217898, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 144.79, "cuda_time_us": 43.039, "pct_cuda_time": 0.6230262175060538, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.039, "pct_cuda_time": 0.6230262175060538, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2450.562, "cuda_time_us": 200.797, "pct_cuda_time": 2.906707762646973, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.065, "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04863886453728806, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1720.651, "cuda_time_us": 58.495, "pct_cuda_time": 0.8467649943775788, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.529, "cuda_time_us": 20.895, "pct_cuda_time": 0.3024729388412601, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.895, "pct_cuda_time": 0.3024729388412601, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 536.01, "cuda_time_us": 3.616, "pct_cuda_time": 0.0523446827877481, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0523446827877481, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 734.303, "cuda_time_us": 16.384, "pct_cuda_time": 0.23717236802944272, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.64, "pct_cuda_time": 0.1829747761164646, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.703, "cuda_time_us": 17.6, "pct_cuda_time": 0.2547750047191279, "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": 0.22420200415283253, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.419, "cuda_time_us": 3.328, "pct_cuda_time": 0.04817563725598055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04817563725598055, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 504.44, "cuda_time_us": 135.614, "pct_cuda_time": 1.9631282664761256, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.717, "cuda_time_us": 83.775, "pct_cuda_time": 1.212714546610508, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.775, "pct_cuda_time": 1.212714546610508, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 139.236, "cuda_time_us": 8.896, "pct_cuda_time": 0.1287771842034865, "trace": "" }, "children": [ { "entry": { "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": 0.1287771842034865, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.971, "cuda_time_us": 42.943, "pct_cuda_time": 0.6216365356621312, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6216365356621312, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2401.993, "cuda_time_us": 200.702, "pct_cuda_time": 2.9053325566555914, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.246, "cuda_time_us": 3.263, "pct_cuda_time": 0.04723470684082468, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.263, "pct_cuda_time": 0.04723470684082468, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1733.595, "cuda_time_us": 58.753, "pct_cuda_time": 0.8504997643331206, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 178.824, "cuda_time_us": 20.704, "pct_cuda_time": 0.29970805100595593, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29970805100595593, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 489.571, "cuda_time_us": 3.68, "pct_cuda_time": 0.05327113735036311, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05327113735036311, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 755.511, "cuda_time_us": 16.448, "pct_cuda_time": 0.2380988225920577, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.64, "pct_cuda_time": 0.1829747761164646, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 150.53, "cuda_time_us": 17.921, "pct_cuda_time": 0.2594217533847438, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.808, "pct_cuda_time": 0.2288342769659076, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.113, "pct_cuda_time": 0.030587476418836208, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 74.722, "cuda_time_us": 3.232, "pct_cuda_time": 0.04678595541205804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04678595541205804, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.203, "cuda_time_us": 135.454, "pct_cuda_time": 1.9608121300695884, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.963, "cuda_time_us": 83.294, "pct_cuda_time": 1.2057516615383546, "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": 83.294, "pct_cuda_time": 1.2057516615383546, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.748, "cuda_time_us": 8.704, "pct_cuda_time": 0.12599782051564146, "trace": "" }, "children": [ { "entry": { "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": 0.12599782051564146, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.412, "cuda_time_us": 43.456, "pct_cuda_time": 0.6290626480155923, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.456, "pct_cuda_time": 0.6290626480155923, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2419.542, "cuda_time_us": 200.192, "pct_cuda_time": 2.897949871859753, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.068, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1701.745, "cuda_time_us": 58.433, "pct_cuda_time": 0.8458674915200455, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.938, "cuda_time_us": 20.448, "pct_cuda_time": 0.2960022327554959, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2960022327554959, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 482.834, "cuda_time_us": 3.681, "pct_cuda_time": 0.05328561320290397, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.681, "pct_cuda_time": 0.05328561320290397, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 759.048, "cuda_time_us": 16.544, "pct_cuda_time": 0.23948850443598024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.768, "pct_cuda_time": 0.18482768524169463, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 173.653, "cuda_time_us": 17.759999999999998, "pct_cuda_time": 0.2570911411256654, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.648, "pct_cuda_time": 0.22651814055937008, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.687, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 491.034, "cuda_time_us": 135.167, "pct_cuda_time": 1.9566575603903613, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 181.606, "cuda_time_us": 82.303, "pct_cuda_time": 1.1914060916703626, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.303, "pct_cuda_time": 1.1914060916703626, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 102.325, "cuda_time_us": 8.961, "pct_cuda_time": 0.12971811461864236, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.961, "pct_cuda_time": 0.12971811461864236, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.765, "cuda_time_us": 43.903, "pct_cuda_time": 0.6355333541013564, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6355333541013564, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2421.435, "cuda_time_us": 199.836, "pct_cuda_time": 2.892796468355207, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.706, "cuda_time_us": 3.2, "pct_cuda_time": 0.04632272813075053, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04632272813075053, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1721.585, "cuda_time_us": 58.59, "pct_cuda_time": 0.8481402003689605, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 143.304, "cuda_time_us": 20.672, "pct_cuda_time": 0.2992448237246484, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.672, "pct_cuda_time": 0.2992448237246484, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 508.262, "cuda_time_us": 3.552, "pct_cuda_time": 0.05141822822513309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05141822822513309, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 746.765, "cuda_time_us": 16.735, "pct_cuda_time": 0.24225339227128437, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.927, "pct_cuda_time": 0.18712934579569127, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.62, "cuda_time_us": 17.631, "pct_cuda_time": 0.25522375614789455, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.519, "pct_cuda_time": 0.22465075558159922, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 84.678, "cuda_time_us": 3.392, "pct_cuda_time": 0.04910209181859555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04910209181859555, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.947, "cuda_time_us": 134.654, "pct_cuda_time": 1.9492314480369006, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 168.823, "cuda_time_us": 82.847, "pct_cuda_time": 1.1992809554525903, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.847, "pct_cuda_time": 1.1992809554525903, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.252, "cuda_time_us": 8.864, "pct_cuda_time": 0.12831395692217898, "trace": "" }, "children": [ { "entry": { "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": 0.12831395692217898, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.06, "cuda_time_us": 42.943, "pct_cuda_time": 0.6216365356621312, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6216365356621312, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2341.103, "cuda_time_us": 200.86, "pct_cuda_time": 2.9076197413570473, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.97, "cuda_time_us": 3.233, "pct_cuda_time": 0.046800431264598895, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.233, "pct_cuda_time": 0.046800431264598895, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1652.85, "cuda_time_us": 58.876999999999995, "pct_cuda_time": 0.8522947700481872, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 140.204, "cuda_time_us": 20.704, "pct_cuda_time": 0.29970805100595593, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29970805100595593, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.007, "cuda_time_us": 3.775, "pct_cuda_time": 0.054646343341744764, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.775, "pct_cuda_time": 0.054646343341744764, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 713.362, "cuda_time_us": 16.672, "pct_cuda_time": 0.24134141356121028, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03613172794198541, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.704, "pct_cuda_time": 0.1839012306790796, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.506, "cuda_time_us": 17.726, "pct_cuda_time": 0.25659896213927624, "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": 15.583, "pct_cuda_time": 0.2255772101442142, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.143, "pct_cuda_time": 0.031021751995061995, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.811, "cuda_time_us": 3.392, "pct_cuda_time": 0.04910209181859555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04910209181859555, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 468.463, "cuda_time_us": 135.358, "pct_cuda_time": 1.959422448225666, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.542, "cuda_time_us": 83.743, "pct_cuda_time": 1.2122513193292004, "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": 83.743, "pct_cuda_time": 1.2122513193292004, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.558, "cuda_time_us": 8.544, "pct_cuda_time": 0.12368168410910392, "trace": "" }, "children": [ { "entry": { "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": 0.12368168410910392, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.485, "cuda_time_us": 43.071, "pct_cuda_time": 0.6234894447873612, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.071, "pct_cuda_time": 0.6234894447873612, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2555.618, "cuda_time_us": 201.757, "pct_cuda_time": 2.9206045810861987, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.457, "cuda_time_us": 3.424, "pct_cuda_time": 0.04956531909990307, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04956531909990307, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1844.164, "cuda_time_us": 59.55, "pct_cuda_time": 0.8620370188081856, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 145.312, "cuda_time_us": 21.375, "pct_cuda_time": 0.3094213480608727, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.375, "pct_cuda_time": 0.3094213480608727, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 524.197, "cuda_time_us": 3.583, "pct_cuda_time": 0.051866979653899734, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.583, "pct_cuda_time": 0.051866979653899734, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 790.378, "cuda_time_us": 16.449, "pct_cuda_time": 0.23811329844459858, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.641, "pct_cuda_time": 0.18298925196900545, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 160.518, "cuda_time_us": 18.143, "pct_cuda_time": 0.26263539264881464, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.839, "pct_cuda_time": 0.22928302839467424, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.304, "pct_cuda_time": 0.03335236425414038, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 96.678, "cuda_time_us": 3.265, "pct_cuda_time": 0.047263658545906403, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.265, "pct_cuda_time": 0.047263658545906403, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 477.813, "cuda_time_us": 135.518, "pct_cuda_time": 1.961738584632203, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 170.536, "cuda_time_us": 83.487, "pct_cuda_time": 1.2085455010787405, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.487, "pct_cuda_time": 1.2085455010787405, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.047, "cuda_time_us": 8.896, "pct_cuda_time": 0.1287771842034865, "trace": "" }, "children": [ { "entry": { "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": 0.1287771842034865, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 154.661, "cuda_time_us": 43.135, "pct_cuda_time": 0.6244158993499762, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6244158993499762, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2614.03, "cuda_time_us": 200.76500000000001, "pct_cuda_time": 2.906244535365666, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 70.54, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1747.629, "cuda_time_us": 59.071999999999996, "pct_cuda_time": 0.8551175612936547, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 142.283, "cuda_time_us": 20.672, "pct_cuda_time": 0.2992448237246484, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.672, "pct_cuda_time": 0.2992448237246484, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 486.259, "cuda_time_us": 3.584, "pct_cuda_time": 0.05188145550644059, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05188145550644059, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 806.303, "cuda_time_us": 16.416, "pct_cuda_time": 0.23763559531075024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.672, "pct_cuda_time": 0.1834380033977721, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 157.715, "cuda_time_us": 18.4, "pct_cuda_time": 0.2663556867518155, "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": 16.288, "pct_cuda_time": 0.23578268618552023, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 77.717, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 646.37, "cuda_time_us": 135.133, "pct_cuda_time": 1.9561653814039723, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 157.492, "cuda_time_us": 82.495, "pct_cuda_time": 1.1941854553582079, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.495, "pct_cuda_time": 1.1941854553582079, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 92.612, "cuda_time_us": 8.799, "pct_cuda_time": 0.1273730265070231, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.799, "pct_cuda_time": 0.1273730265070231, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 344.706, "cuda_time_us": 43.839, "pct_cuda_time": 0.6346068995387414, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.6346068995387414, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2489.418, "cuda_time_us": 201.30700000000002, "pct_cuda_time": 2.9140904474428115, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.956, "cuda_time_us": 3.2, "pct_cuda_time": 0.04632272813075053, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04632272813075053, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1774.243, "cuda_time_us": 58.87700000000001, "pct_cuda_time": 0.8522947700481873, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 135.948, "cuda_time_us": 20.767, "pct_cuda_time": 0.3006200297160301, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.767, "pct_cuda_time": 0.3006200297160301, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.28, "cuda_time_us": 3.583, "pct_cuda_time": 0.051866979653899734, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.583, "pct_cuda_time": 0.051866979653899734, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 775.566, "cuda_time_us": 16.672, "pct_cuda_time": 0.24134141356121028, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.672, "pct_cuda_time": 0.1834380033977721, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.664, "pct_cuda_time": 0.024087818627990275, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.035, "cuda_time_us": 17.855, "pct_cuda_time": 0.25846634711704713, "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": 15.743, "pct_cuda_time": 0.22789334655075175, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 86.586, "cuda_time_us": 3.424, "pct_cuda_time": 0.04956531909990307, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04956531909990307, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 490.344, "cuda_time_us": 135.80599999999998, "pct_cuda_time": 1.9659076301639704, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 179.248, "cuda_time_us": 82.143, "pct_cuda_time": 1.1890899552638252, "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": 82.143, "pct_cuda_time": 1.1890899552638252, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 104.481, "cuda_time_us": 9.472, "pct_cuda_time": 0.13711527526702158, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.472, "pct_cuda_time": 0.13711527526702158, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 148.176, "cuda_time_us": 44.191, "pct_cuda_time": 0.6397023996331239, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.191, "pct_cuda_time": 0.6397023996331239, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2432.595, "cuda_time_us": 200.989, "pct_cuda_time": 2.909487126334818, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.411, "cuda_time_us": 3.392, "pct_cuda_time": 0.04910209181859555, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04910209181859555, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1714.556, "cuda_time_us": 58.751000000000005, "pct_cuda_time": 0.8504708126280389, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 144.1, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 516.097, "cuda_time_us": 3.712, "pct_cuda_time": 0.05373436463167061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05373436463167061, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.119, "cuda_time_us": 16.703000000000003, "pct_cuda_time": 0.24179016498997696, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.432, "pct_cuda_time": 0.035205273379370405, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.8, "pct_cuda_time": 0.18529091252300212, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.021293979087604383, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 161.927, "cuda_time_us": 17.472, "pct_cuda_time": 0.2529220955938979, "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": 0.221885867746295, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.144, "pct_cuda_time": 0.031036227847602856, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.471, "cuda_time_us": 3.328, "pct_cuda_time": 0.04817563725598055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04817563725598055, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 489.062, "cuda_time_us": 135.518, "pct_cuda_time": 1.961738584632203, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 167.333, "cuda_time_us": 83.135, "pct_cuda_time": 1.2034500009843578, "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": 83.135, "pct_cuda_time": 1.2034500009843578, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 117.796, "cuda_time_us": 9.088, "pct_cuda_time": 0.1315565478913315, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.088, "pct_cuda_time": 0.1315565478913315, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.609, "cuda_time_us": 43.295, "pct_cuda_time": 0.6267320357565138, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.295, "pct_cuda_time": 0.6267320357565138, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2417.404, "cuda_time_us": 200.447, "pct_cuda_time": 2.901641214257672, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 66.133, "cuda_time_us": 3.296, "pct_cuda_time": 0.04771240997467304, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04771240997467304, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1732.831, "cuda_time_us": 58.49700000000001, "pct_cuda_time": 0.8467939460826607, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 166.387, "cuda_time_us": 20.832, "pct_cuda_time": 0.30156096013118594, "trace": "" }, "children": [ { "entry": { "name": "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": 0.30156096013118594, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 481.259, "cuda_time_us": 3.553, "pct_cuda_time": 0.051432704077673946, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.553, "pct_cuda_time": 0.051432704077673946, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 755.201, "cuda_time_us": 16.512, "pct_cuda_time": 0.23902527715467273, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.672, "pct_cuda_time": 0.1834380033977721, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.02177168222145275, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 169.808, "cuda_time_us": 17.6, "pct_cuda_time": 0.2547750047191279, "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": 0.22420200415283253, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.009, "cuda_time_us": 3.232, "pct_cuda_time": 0.04678595541205804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04678595541205804, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.967, "cuda_time_us": 135.422, "pct_cuda_time": 1.9603489027882806, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.588, "cuda_time_us": 83.327, "pct_cuda_time": 1.2062293646722029, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.327, "pct_cuda_time": 1.2062293646722029, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.308, "cuda_time_us": 8.864, "pct_cuda_time": 0.12831395692217898, "trace": "" }, "children": [ { "entry": { "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": 0.12831395692217898, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 141.78, "cuda_time_us": 43.231, "pct_cuda_time": 0.6258055811938988, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6258055811938988, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2508.49, "cuda_time_us": 200.157, "pct_cuda_time": 2.897443217020823, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.387, "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04863886453728806, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1802.086, "cuda_time_us": 58.367000000000004, "pct_cuda_time": 0.8449120852523488, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 146.568, "cuda_time_us": 20.544, "pct_cuda_time": 0.2973919145994184, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2973919145994184, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 531.589, "cuda_time_us": 3.616, "pct_cuda_time": 0.0523446827877481, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0523446827877481, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 796.909, "cuda_time_us": 16.449, "pct_cuda_time": 0.23811329844459858, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.641, "pct_cuda_time": 0.18298925196900545, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 162.76, "cuda_time_us": 17.758, "pct_cuda_time": 0.2570621894205837, "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": 0.22650366470682923, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030558524713754497, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 80.592, "cuda_time_us": 3.328, "pct_cuda_time": 0.04817563725598055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04817563725598055, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 463.868, "cuda_time_us": 135.102, "pct_cuda_time": 1.9557166299752058, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 158.967, "cuda_time_us": 83.231, "pct_cuda_time": 1.2048396828282804, "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": 1.2048396828282804, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.317, "cuda_time_us": 8.96, "pct_cuda_time": 0.1297036387661015, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.96, "pct_cuda_time": 0.1297036387661015, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.76, "cuda_time_us": 42.911, "pct_cuda_time": 0.6211733083808237, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6211733083808237, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2355.937, "cuda_time_us": 199.997, "pct_cuda_time": 2.895127080614286, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 72.242, "cuda_time_us": 3.233, "pct_cuda_time": 0.046800431264598895, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.233, "pct_cuda_time": 0.046800431264598895, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1668.353, "cuda_time_us": 58.655, "pct_cuda_time": 0.8490811307841163, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.477, "cuda_time_us": 20.704, "pct_cuda_time": 0.29970805100595593, "trace": "" }, "children": [ { "entry": { "name": "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": 0.29970805100595593, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 483.848, "cuda_time_us": 3.712, "pct_cuda_time": 0.05373436463167061, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05373436463167061, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 735.998, "cuda_time_us": 16.639, "pct_cuda_time": 0.24086371042736188, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.831, "pct_cuda_time": 0.18573966395176875, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.02177168222145275, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.472, "cuda_time_us": 17.6, "pct_cuda_time": 0.2547750047191279, "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": 0.22420200415283253, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.971, "cuda_time_us": 3.232, "pct_cuda_time": 0.04678595541205804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04678595541205804, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.141, "cuda_time_us": 134.877, "pct_cuda_time": 1.9524595631535124, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.248, "cuda_time_us": 83.391, "pct_cuda_time": 1.207155819234818, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 83.391, "pct_cuda_time": 1.207155819234818, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.701, "cuda_time_us": 8.799, "pct_cuda_time": 0.1273730265070231, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 8.799, "pct_cuda_time": 0.1273730265070231, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.033, "cuda_time_us": 42.687, "pct_cuda_time": 0.6179307174116712, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6179307174116712, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2423.663, "cuda_time_us": 202.173, "pct_cuda_time": 2.9266265357431958, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.826, "cuda_time_us": 3.2, "pct_cuda_time": 0.04632272813075053, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04632272813075053, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1752.138, "cuda_time_us": 59.743, "pct_cuda_time": 0.8648308583485715, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 174.581, "cuda_time_us": 21.408, "pct_cuda_time": 0.30989905119472105, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.408, "pct_cuda_time": 0.30989905119472105, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 523.423, "cuda_time_us": 3.968, "pct_cuda_time": 0.05744018288213066, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.968, "pct_cuda_time": 0.05744018288213066, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 739.064, "cuda_time_us": 16.672, "pct_cuda_time": 0.24134141356121028, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.704, "pct_cuda_time": 0.1839012306790796, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.632, "pct_cuda_time": 0.023624591346682766, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 156.483, "cuda_time_us": 17.695, "pct_cuda_time": 0.2561502107105096, "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": 15.583, "pct_cuda_time": 0.2255772101442142, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.55, "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04863886453728806, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.003, "cuda_time_us": 135.87, "pct_cuda_time": 1.9668340847265857, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 154.977, "cuda_time_us": 82.783, "pct_cuda_time": 1.1983545008899754, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.783, "pct_cuda_time": 1.1983545008899754, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 95.77, "cuda_time_us": 9.152, "pct_cuda_time": 0.1324830024539465, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.152, "pct_cuda_time": 0.1324830024539465, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 143.142, "cuda_time_us": 43.935, "pct_cuda_time": 0.6359965813826639, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.935, "pct_cuda_time": 0.6359965813826639, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2353.061, "cuda_time_us": 201.30900000000003, "pct_cuda_time": 2.9141193991478938, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.23, "cuda_time_us": 3.232, "pct_cuda_time": 0.04678595541205804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04678595541205804, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1675.513, "cuda_time_us": 59.35900000000001, "pct_cuda_time": 0.8592721309728817, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.572, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 484.416, "cuda_time_us": 3.648, "pct_cuda_time": 0.0528079100690556, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.0528079100690556, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 723.87, "cuda_time_us": 16.543000000000003, "pct_cuda_time": 0.2394740285834394, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.335, "pct_cuda_time": 0.03380111568290703, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.736, "pct_cuda_time": 0.1843644579603871, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 170.511, "cuda_time_us": 18.304000000000002, "pct_cuda_time": 0.26496600490789307, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 16.192, "pct_cuda_time": 0.23439300434159765, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.144, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 453.3, "cuda_time_us": 135.454, "pct_cuda_time": 1.9608121300695884, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 163.668, "cuda_time_us": 82.207, "pct_cuda_time": 1.19001640982644, "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": 1.19001640982644, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 97.193, "cuda_time_us": 9.056, "pct_cuda_time": 0.131093320610024, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.056, "pct_cuda_time": 0.131093320610024, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.516, "cuda_time_us": 44.191, "pct_cuda_time": 0.6397023996331239, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 44.191, "pct_cuda_time": 0.6397023996331239, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2653.033, "cuda_time_us": 200.348, "pct_cuda_time": 2.9002081048561275, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 71.374, "cuda_time_us": 3.2, "pct_cuda_time": 0.04632272813075053, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04632272813075053, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1957.889, "cuda_time_us": 58.463, "pct_cuda_time": 0.8463017670962714, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 161.129, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.864, "pct_cuda_time": 0.3020241874124935, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 504.018, "cuda_time_us": 3.647, "pct_cuda_time": 0.052793434216514744, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.647, "pct_cuda_time": 0.052793434216514744, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 971.916, "cuda_time_us": 16.352, "pct_cuda_time": 0.2367091407481352, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.032889136972832876, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.608, "pct_cuda_time": 0.1825115488351571, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 167.796, "cuda_time_us": 17.6, "pct_cuda_time": 0.2547750047191279, "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": 0.22420200415283253, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.112, "pct_cuda_time": 0.03057300056629535, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 83.901, "cuda_time_us": 3.328, "pct_cuda_time": 0.04817563725598055, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04817563725598055, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 465.631, "cuda_time_us": 135.357, "pct_cuda_time": 1.9594079723731248, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 169.367, "cuda_time_us": 82.654, "pct_cuda_time": 1.1964871159122044, "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": 82.654, "pct_cuda_time": 1.1964871159122044, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 100.445, "cuda_time_us": 8.832, "pct_cuda_time": 0.12785072964087146, "trace": "" }, "children": [ { "entry": { "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": 0.12785072964087146, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.929, "cuda_time_us": 43.871, "pct_cuda_time": 0.6350701268200489, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6350701268200489, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2349.604, "cuda_time_us": 201.566, "pct_cuda_time": 2.917839693250894, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 68.646, "cuda_time_us": 3.2, "pct_cuda_time": 0.04632272813075053, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04632272813075053, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1687.336, "cuda_time_us": 59.455000000000005, "pct_cuda_time": 0.860661812816804, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.243, "cuda_time_us": 21.183, "pct_cuda_time": 0.30664198437302764, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 21.183, "pct_cuda_time": 0.30664198437302764, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 507.614, "cuda_time_us": 3.744, "pct_cuda_time": 0.05419759191297812, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05419759191297812, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 729.975, "cuda_time_us": 16.705000000000002, "pct_cuda_time": 0.24181911669505868, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03336684010668124, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.896, "pct_cuda_time": 0.18668059436692466, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.504, "pct_cuda_time": 0.02177168222145275, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 154.932, "cuda_time_us": 17.823, "pct_cuda_time": 0.2580031198357396, "trace": "" }, "children": [ { "entry": { "name": "void cutlass::Kernel2(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_64x4_tn_align8::Params)", "cpu_time_us": 0, "cuda_time_us": 15.712, "pct_cuda_time": 0.22744459512198512, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030558524713754497, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 76.393, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 449.535, "cuda_time_us": 135.647, "pct_cuda_time": 1.9636059696099737, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 161.502, "cuda_time_us": 83.008, "pct_cuda_time": 1.2016115677116685, "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": 83.008, "pct_cuda_time": 1.2016115677116685, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 94.678, "cuda_time_us": 9.696, "pct_cuda_time": 0.14035786623617408, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.696, "pct_cuda_time": 0.14035786623617408, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.12, "cuda_time_us": 42.943, "pct_cuda_time": 0.6216365356621312, "trace": "" }, "children": [ { "entry": { "name": "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": 0.6216365356621312, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2426.34, "cuda_time_us": 199.07, "pct_cuda_time": 2.8817079653089084, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 67.965, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1715.806, "cuda_time_us": 58.015, "pct_cuda_time": 0.8398165851579662, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 137.085, "cuda_time_us": 20.576, "pct_cuda_time": 0.2978551418807259, "trace": "" }, "children": [ { "entry": { "name": "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": 0.2978551418807259, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 475.777, "cuda_time_us": 3.585, "pct_cuda_time": 0.051895931358981455, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.585, "pct_cuda_time": 0.051895931358981455, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 737.763, "cuda_time_us": 16.319000000000003, "pct_cuda_time": 0.2362314376142869, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03335236425414038, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.544, "pct_cuda_time": 0.18158509427254207, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.471, "pct_cuda_time": 0.021293979087604383, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 215.326, "cuda_time_us": 17.535, "pct_cuda_time": 0.25383407430397203, "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": 0.22327554959021756, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030558524713754497, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 82.85, "cuda_time_us": 3.552, "pct_cuda_time": 0.05141822822513309, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.552, "pct_cuda_time": 0.05141822822513309, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 487.778, "cuda_time_us": 134.239, "pct_cuda_time": 1.9432239692324438, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 164.863, "cuda_time_us": 82.239, "pct_cuda_time": 1.1904796371077477, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 82.239, "pct_cuda_time": 1.1904796371077477, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 99.232, "cuda_time_us": 8.896, "pct_cuda_time": 0.1287771842034865, "trace": "" }, "children": [ { "entry": { "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": 0.1287771842034865, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 146.26, "cuda_time_us": 43.104, "pct_cuda_time": 0.6239671479212096, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.104, "pct_cuda_time": 0.6239671479212096, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2461.535, "cuda_time_us": 200.348, "pct_cuda_time": 2.9002081048561275, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 78.762, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1768.523, "cuda_time_us": 58.65299999999999, "pct_cuda_time": 0.8490521790790345, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 138.152, "cuda_time_us": 20.959, "pct_cuda_time": 0.30339939340387506, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.959, "pct_cuda_time": 0.30339939340387506, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 528.107, "cuda_time_us": 3.552, "pct_cuda_time": 0.05141822822513309, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_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": 0.05141822822513309, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 792.752, "cuda_time_us": 16.607, "pct_cuda_time": 0.24040048314605436, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 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": 0.03381559153544788, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.799, "pct_cuda_time": 0.18527643667046123, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.472, "pct_cuda_time": 0.021308454940145244, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 151.775, "cuda_time_us": 17.535, "pct_cuda_time": 0.25383407430397203, "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": 0.22327554959021756, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030558524713754497, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 79.801, "cuda_time_us": 3.232, "pct_cuda_time": 0.04678595541205804, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04678595541205804, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 462.651, "cuda_time_us": 135.199, "pct_cuda_time": 1.9571207876716692, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 166.557, "cuda_time_us": 82.879, "pct_cuda_time": 1.199744182733898, "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": 1.199744182733898, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.461, "cuda_time_us": 9.12, "pct_cuda_time": 0.13201977517263902, "trace": "" }, "children": [ { "entry": { "name": "void vllm::act_and_mul_kernel(c10::BFloat16 const&)), true>(c10::BFloat16*, c10::BFloat16 const*, int)", "cpu_time_us": 0, "cuda_time_us": 9.12, "pct_cuda_time": 0.13201977517263902, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 145.928, "cuda_time_us": 43.2, "pct_cuda_time": 0.6253568297651322, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.2, "pct_cuda_time": 0.6253568297651322, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "LlamaDecoderLayer", "cpu_time_us": 2416.805, "cuda_time_us": 200.73, "pct_cuda_time": 2.9057378805267353, "trace": "" }, "children": [ { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 73.538, "cuda_time_us": 3.424, "pct_cuda_time": 0.04956531909990307, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04956531909990307, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaAttention", "cpu_time_us": 1741.465, "cuda_time_us": 58.781, "pct_cuda_time": 0.8509050882042646, "trace": "" }, "children": [ { "entry": { "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", "cpu_time_us": 136.696, "cuda_time_us": 20.863, "pct_cuda_time": 0.3020097115599526, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 20.863, "pct_cuda_time": 0.3020097115599526, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[6, 4096], bfloat16[6144, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Llama3RotaryEmbedding", "cpu_time_us": 525.891, "cuda_time_us": 3.583, "pct_cuda_time": 0.051866979653899734, "trace": "" }, "children": [ { "entry": { "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", "cpu_time_us": 0, "cuda_time_us": 3.583, "pct_cuda_time": 0.051866979653899734, "trace": "_C::rotary_embedding(int64[6], bfloat16[6, 4096], bfloat16[6, 1024], 128, bfloat16[131072, 128], True)" }, "children": [] } ] }, { "entry": { "name": "Attention", "cpu_time_us": 761.458, "cuda_time_us": 16.704, "pct_cuda_time": 0.2418046408425178, "trace": "" }, "children": [ { "entry": { "name": "void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0>(__nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", "cpu_time_us": 0, "cuda_time_us": 2.528, "pct_cuda_time": 0.03659495522329292, "trace": "_C_cache_ops::reshape_and_cache_flash(bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], int64[6], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void cutlass::device_kernel, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > > >(flash::enable_sm90_or_later, cute::C<1>, cute::C<1> >, cute::tuple, cute::C<128>, cute::C<128> >, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, true, true, false, true, true, true, false, false>, flash::CollectiveEpilogueFwd, cute::C<128>, cute::C<128> >, cute::tuple, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 256, true, true, false>, flash::VarlenDynamicPersistentTileScheduler<128, 256, 128, false, true, true> > >::Params)", "cpu_time_us": 0, "cuda_time_us": 12.736, "pct_cuda_time": 0.1843644579603871, "trace": "_vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] }, { "entry": { "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "cpu_time_us": 0, "cuda_time_us": 1.44, "pct_cuda_time": 0.020845227658837735, "trace": "fill_(int32[1], 0) <- zero_(int32[1]) <- zeros(None, 3, 0, None, None) <- _vllm_fa3_C::fwd(bfloat16[6, 1, 32, 128], bfloat16[28102, 16, 8, 128], bfloat16[28102, 16, 8, 128], None, None, bfloat16[6, 1, 32, 128], None, None, None, None, int32[6], None, None, int32[6, 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[6, 32, 128], bfloat16[6, 8, 128], bfloat16[6, 8, 128], bfloat16[6, 32, 128], None)" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", "cpu_time_us": 159.585, "cuda_time_us": 17.631, "pct_cuda_time": 0.25522375614789455, "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": 0.22466523143414005, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] }, { "entry": { "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, true, false, false>(cublasLt::cublasSplitKParams, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*)", "cpu_time_us": 0, "cuda_time_us": 2.111, "pct_cuda_time": 0.030558524713754497, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[6, 4096], bfloat16[4096, 4096], None)" }, "children": [] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 75.238, "cuda_time_us": 3.264, "pct_cuda_time": 0.04724918269336553, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04724918269336553, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] }, { "entry": { "name": "LlamaMLP", "cpu_time_us": 458.987, "cuda_time_us": 135.261, "pct_cuda_time": 1.9580182905292023, "trace": "" }, "children": [ { "entry": { "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", "cpu_time_us": 165.383, "cuda_time_us": 82.75, "pct_cuda_time": 1.1978767977561269, "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": 82.75, "pct_cuda_time": 1.1978767977561269, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[6, 4096], bfloat16[28672, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "SiluAndMul", "cpu_time_us": 96.198, "cuda_time_us": 8.672, "pct_cuda_time": 0.12553459323433394, "trace": "" }, "children": [ { "entry": { "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": 0.12553459323433394, "trace": "_C::silu_and_mul(bfloat16[6, 14336], bfloat16[6, 28672])" }, "children": [] } ] }, { "entry": { "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", "cpu_time_us": 140.919, "cuda_time_us": 43.839, "pct_cuda_time": 0.6346068995387414, "trace": "" }, "children": [ { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x64x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 43.839, "pct_cuda_time": 0.6346068995387414, "trace": "mm(bfloat16[6, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[6, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[6, 14336], bfloat16[4096, 14336], None)" }, "children": [] } ] } ] } ] }, { "entry": { "name": "RMSNorm(weight=bfloat16[4096])", "cpu_time_us": 69.85, "cuda_time_us": 3.36, "pct_cuda_time": 0.04863886453728806, "trace": "" }, "children": [ { "entry": { "name": "std::enable_if<(((8)>(0)))&&vllm::_typeConvert::exists, void>::type 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": 0.04863886453728806, "trace": "_C::fused_add_rms_norm(bfloat16[6, 4096], bfloat16[6, 4096], bfloat16[4096], 1e-05)" }, "children": [] } ] } ] }, { "entry": { "name": "LogitsProcessor", "cpu_time_us": 448.575, "cuda_time_us": 347.19599999999997, "pct_cuda_time": 5.025958098776268, "trace": "" }, "children": [ { "entry": { "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", "cpu_time_us": 0, "cuda_time_us": 4.673, "pct_cuda_time": 0.06764565892343663, "trace": "index_select(bfloat16[6, 4096], 0, int64[6])" }, "children": [] }, { "entry": { "name": "Memset (Device)", "cpu_time_us": 0, "cuda_time_us": 0.736, "pct_cuda_time": 0.010654227470072622, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] }, { "entry": { "name": "sm90_xmma_gemm_bf16bf16_bf16f32_f32_tn_n_tilesize64x128x64_warpgroupsize1x1x1_execute_segment_k_off_kernel__5x_cublas", "cpu_time_us": 0, "cuda_time_us": 341.787, "pct_cuda_time": 4.947658212382759, "trace": "mm(bfloat16[6, 4096], bfloat16[4096, 128256]) <- matmul(bfloat16[6, 4096], bfloat16[4096, 128256]) <- linear(bfloat16[6, 4096], bfloat16[128256, 4096], None)" }, "children": [] } ] }, { "entry": { "name": "Sampler", "cpu_time_us": 3445.263, "cuda_time_us": 118.04700000000001, "pct_cuda_time": 1.7088309648908464, "trace": "" }, "children": [ { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.769, "pct_cuda_time": 0.011131930603920986, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.735, "pct_cuda_time": 0.010639751617531763, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011117454751380127, "trace": "copy_(int32[6], int32[6], True) <- _to_copy(int32[6], 3, 0, None, None, True, None) <- to(int32[6], 3, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.011580682032687632, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.011580682032687632, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.8, "pct_cuda_time": 0.011580682032687632, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "Memcpy HtoD (Pinned -> Device)", "cpu_time_us": 0, "cuda_time_us": 0.768, "pct_cuda_time": 0.011117454751380127, "trace": "copy_(bfloat16[6], bfloat16[6], True) <- _to_copy(bfloat16[6], 15, 0, None, None, True, None) <- to(bfloat16[6], 15, 0, None, None, True, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 4.288, "pct_cuda_time": 0.06207245569520571, "trace": "copy_(float32[6, 128256], bfloat16[6, 128256], False) <- _to_copy(bfloat16[6, 128256], 6, None, None, None, False, None) <- to(bfloat16[6, 128256], 6, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 5.568, "pct_cuda_time": 0.08060154694750592, "trace": "div_(float32[6, 128256], bfloat16[6, 1])" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::SoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 35.616, "pct_cuda_time": 0.5155719640952534, "trace": "_softmax(float32[6, 128256], -1, False) <- softmax(float32[6, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::(anonymous namespace)::cunn_SoftMaxForward<4, float, float, float, at::native::(anonymous namespace)::LogSoftMaxForwardEpilogue>(float*, float const*, int)", "cpu_time_us": 0, "cuda_time_us": 28.16, "pct_cuda_time": 0.4076400075506047, "trace": "_log_softmax(float32[6, 128256], -1, False) <- log_softmax(float32[6, 128256], -1, 6)" }, "children": [] }, { "entry": { "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "cpu_time_us": 0, "cuda_time_us": 1.728, "pct_cuda_time": 0.02501427319060529, "trace": "copy_(int64[6], int32[6], False) <- _to_copy(int32[6], 4, None, None, None, False, None) <- to(int32[6], 4, False, False, None)" }, "children": [] }, { "entry": { "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "cpu_time_us": 0, "cuda_time_us": 5.983, "pct_cuda_time": 0.08660902575196262, "trace": "index(float32[6, 128256], None)" }, "children": [] }, { "entry": { "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4> >(at::native::ReduceOp, unsigned int, long, 4>)", "cpu_time_us": 0, "cuda_time_us": 28.608, "pct_cuda_time": 0.4141251894889097, "trace": "argmax(float32[6, 128256], -1, False)" }, "children": [] }, { "entry": { "name": "Memcpy DtoH (Device -> Pageable)", "cpu_time_us": 0, "cuda_time_us": 2.656, "pct_cuda_time": 0.03844786434852294, "trace": "copy_(int64[6], int64[6], False) <- _to_copy(int64[6], 4, 0, None, None, False, None) <- to(int64[6], 4, 0, None, None, False, False, None)" }, "children": [] } ] } ] } }