diff --git "a/MI300x_llama8b_pp1_tp1/profiling_bs1_pl256.json" "b/MI300x_llama8b_pp1_tp1/profiling_bs1_pl256.json" new file mode 100644--- /dev/null +++ "b/MI300x_llama8b_pp1_tp1/profiling_bs1_pl256.json" @@ -0,0 +1,18162 @@ +{ + "context": { + "python_version": "3.12.10 (main, Apr 9 2025, 08:55:05) [GCC 11.4.0]", + "torch_version": "2.7.0a0+git295f2ed", + "engine_args": { + "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", + "served_model_name": null, + "tokenizer": null, + "hf_config_path": null, + "task": "auto", + "skip_tokenizer_init": false, + "tokenizer_mode": "auto", + "trust_remote_code": false, + "allowed_local_media_path": "", + "download_dir": null, + "load_format": "dummy", + "config_format": "auto", + "dtype": "auto", + "kv_cache_dtype": "auto", + "seed": null, + "max_model_len": null, + "cuda_graph_sizes": [ + 512 + ], + "distributed_executor_backend": null, + "pipeline_parallel_size": 1, + "tensor_parallel_size": 1, + "data_parallel_size": 1, + "enable_expert_parallel": false, + "max_parallel_loading_workers": null, + "block_size": null, + "enable_prefix_caching": null, + "prefix_caching_hash_algo": "builtin", + "disable_sliding_window": false, + "disable_cascade_attn": 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": null, + "max_logprobs": 20, + "disable_log_stats": false, + "revision": null, + "code_revision": null, + "rope_scaling": {}, + "rope_theta": null, + "hf_token": null, + "hf_overrides": {}, + "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": {}, + "limit_mm_per_prompt": {}, + "mm_processor_kwargs": null, + "disable_mm_preprocessor_cache": false, + "enable_lora": null, + "enable_lora_bias": false, + "max_loras": 1, + "max_lora_rank": 16, + "fully_sharded_loras": false, + "max_cpu_loras": null, + "lora_dtype": "auto", + "lora_extra_vocab_size": 256, + "long_lora_scaling_factors": null, + "enable_prompt_adapter": null, + "max_prompt_adapters": 1, + "max_prompt_adapter_token": 0, + "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": {}, + "ignore_patterns": null, + "preemption_mode": null, + "scheduler_delay_factor": 0.0, + "enable_chunked_prefill": null, + "disable_chunked_mm_input": false, + "guided_decoding_backend": "xgrammar", + "guided_decoding_disable_fallback": false, + "guided_decoding_disable_any_whitespace": false, + "guided_decoding_disable_additional_properties": false, + "logits_processor_pattern": null, + "speculative_config": null, + "qlora_adapter_name_or_path": null, + "show_hidden_metrics_for_version": 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": {}, + "override_pooler_config": null, + "compilation_config": null, + "worker_cls": "auto", + "worker_extension_cls": "", + "kv_transfer_config": null, + "kv_events_config": null, + "generation_config": "auto", + "enable_sleep_mode": false, + "override_generation_config": {}, + "model_impl": "auto", + "calculate_kv_scales": false, + "additional_config": null, + "enable_reasoning": null, + "reasoning_parser": "", + "use_tqdm_on_load": true, + "pt_load_map_location": "cpu" + }, + "prompt_len": 0, + "batch_size": 1, + "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": 18191.011, + "pct_cuda_time": 97.49206882589682, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 7.21, + "pct_cuda_time": 0.03864094283900527, + "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": 7.21, + "pct_cuda_time": 0.03864094283900527, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 18173.827, + "pct_cuda_time": 97.39997368557152, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 622.5070000000002, + "pct_cuda_time": 3.3362354235618112, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", + "cuda_time_us": 5.166, + "pct_cuda_time": 0.027686423121539704, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 617.3410000000001, + "pct_cuda_time": 3.3085490004402716, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 6983.4929999999995, + "pct_cuda_time": 37.42701162685068, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 1738.7930000000006, + "pct_cuda_time": 9.318807340064149, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cuda_time_us": 1738.7930000000006, + "pct_cuda_time": 9.318807340064149, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 857.7579999999999, + "pct_cuda_time": 4.597028827697571, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 629.3709999999998, + "pct_cuda_time": 3.373022029892868, + "invocations": 128 + }, + "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": 228.38700000000006, + "pct_cuda_time": 1.2240067978047018, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 3009.6349999999998, + "pct_cuda_time": 16.129699584087327, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 179.31999999999996, + "pct_cuda_time": 0.9610393716907661, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cuda_time_us": 2830.3149999999996, + "pct_cuda_time": 15.16866021239656, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 1377.307, + "pct_cuda_time": 7.381475875001642, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cuda_time_us": 1377.307, + "pct_cuda_time": 7.381475875001642, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 10567.827000000001, + "pct_cuda_time": 56.63672663515903, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 6272.113, + "pct_cuda_time": 33.61447432909596, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cuda_time_us": 6272.113, + "pct_cuda_time": 33.61447432909596, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 456.39799999999997, + "pct_cuda_time": 2.4459984784793796, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cuda_time_us": 456.39799999999997, + "pct_cuda_time": 2.4459984784793796, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 3839.316000000001, + "pct_cuda_time": 20.57625382758369, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cuda_time_us": 3630.76, + "pct_cuda_time": 19.45852838032549, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cuda_time_us": 208.556, + "pct_cuda_time": 1.1177254472581948, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 9.974, + "pct_cuda_time": 0.05345419748630216, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 9.974, + "pct_cuda_time": 0.05345419748630216, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 241.549, + "pct_cuda_time": 1.2945466160636456, + "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": 4.566, + "pct_cuda_time": 0.0244708106800136, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cuda_time_us": 236.983, + "pct_cuda_time": 1.270075805383632, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 226.40500000000003, + "pct_cuda_time": 1.213384558039527, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cuda_time_us": 33.443, + "pct_cuda_time": 0.17923287813659544, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 2.443, + "pct_cuda_time": 0.013092901991080427, + "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)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", + "cuda_time_us": 14.141, + "pct_cuda_time": 0.07578662589270091, + "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": 70.985, + "pct_cuda_time": 0.3804337486028834, + "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": 40.42, + "pct_cuda_time": 0.2166250914774748, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "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": 8.251, + "pct_cuda_time": 0.04422003042505305, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", + "cuda_time_us": 46.949, + "pct_cuda_time": 0.2516163141953479, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Host)", + "cuda_time_us": 5.007, + "pct_cuda_time": 0.02683428582453528, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 34566.393, + "cuda_time_us": 18191.011, + "pct_cuda_time": 97.49206882589682, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 66.9, + "cuda_time_us": 7.21, + "pct_cuda_time": 0.03864094283900527, + "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": 7.21, + "pct_cuda_time": 0.03864094283900527, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[256]) <- embedding(bfloat16[128256, 4096], int64[256], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1477.095, + "cuda_time_us": 577.6560000000001, + "pct_cuda_time": 3.095863034203666, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 78.106, + "cuda_time_us": 5.166, + "pct_cuda_time": 0.027686423121539704, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", + "cpu_time_us": 0, + "cuda_time_us": 5.166, + "pct_cuda_time": 0.027686423121539704, + "trace": "_C::rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 1073.462, + "cuda_time_us": 219.038, + "pct_cuda_time": 1.1739021966116556, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 175.832, + "cuda_time_us": 60.218, + "pct_cuda_time": 0.3227295833396976, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 60.218, + "pct_cuda_time": 0.3227295833396976, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 95.663, + "cuda_time_us": 25.232999999999997, + "pct_cuda_time": 0.13523258122838, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.046, + "pct_cuda_time": 0.027043300633234478, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.046, + "pct_cuda_time": 0.02168394656402432, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.33, + "pct_cuda_time": 0.03928406532731049, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.029406775777756165, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 3.324, + "pct_cuda_time": 0.01781449292605458, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 632.474, + "cuda_time_us": 90.614, + "pct_cuda_time": 0.4856325096274097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.808, + "pct_cuda_time": 0.031127128433972626, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 84.806, + "pct_cuda_time": 0.454505381193437, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 115.382, + "cuda_time_us": 42.973, + "pct_cuda_time": 0.2303075224161683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.973, + "pct_cuda_time": 0.2303075224161683, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 47.071, + "cuda_time_us": 9.693, + "pct_cuda_time": 0.0519482189928541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.693, + "pct_cuda_time": 0.0519482189928541, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 248.731, + "cuda_time_us": 343.759, + "pct_cuda_time": 1.8423261954776164, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 94.331, + "cuda_time_us": 201.63, + "pct_cuda_time": 1.080606560974845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 201.63, + "pct_cuda_time": 1.080606560974845, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 32.409, + "cuda_time_us": 15.132, + "pct_cuda_time": 0.08109774577528818, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.132, + "pct_cuda_time": 0.08109774577528818, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 99.368, + "cuda_time_us": 126.997, + "pct_cuda_time": 0.680621888727483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 120.108, + "pct_cuda_time": 0.6437012985446943, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.889, + "pct_cuda_time": 0.03692059018278881, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1081.305, + "cuda_time_us": 573.1500000000001, + "pct_cuda_time": 3.0717137847678053, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.027, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 757.28, + "cuda_time_us": 214.901, + "pct_cuda_time": 1.1517305488273333, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 90.395, + "cuda_time_us": 57.095, + "pct_cuda_time": 0.30599232058155423, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 57.095, + "pct_cuda_time": 0.30599232058155423, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 79.95, + "cuda_time_us": 25.032000000000004, + "pct_cuda_time": 0.1341553510604688, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.02403670300040758, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.023613314028939975, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.0336996183871935, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.808, + "pct_cuda_time": 0.031127128433972626, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.021678587209955108, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 446.306, + "cuda_time_us": 90.733, + "pct_cuda_time": 0.4862702727616457, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.168, + "pct_cuda_time": 0.03305649589888828, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 84.565, + "pct_cuda_time": 0.4532137768627573, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 102.924, + "cuda_time_us": 42.041, + "pct_cuda_time": 0.2253126044236644, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.041, + "pct_cuda_time": 0.2253126044236644, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.947, + "cuda_time_us": 9.453, + "pct_cuda_time": 0.05066197401624366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.453, + "pct_cuda_time": 0.05066197401624366, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 223.474, + "cuda_time_us": 339.463, + "pct_cuda_time": 1.8193024103962896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 87.891, + "cuda_time_us": 199.267, + "pct_cuda_time": 1.0679424073093013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 199.267, + "pct_cuda_time": 1.0679424073093013, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.03, + "cuda_time_us": 14.841, + "pct_cuda_time": 0.07953817374114801, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 14.841, + "pct_cuda_time": 0.07953817374114801, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 91.357, + "cuda_time_us": 125.355, + "pct_cuda_time": 0.67182182934584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 118.506, + "pct_cuda_time": 0.6351156133258196, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.849, + "pct_cuda_time": 0.0367062160200204, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1058.56, + "cuda_time_us": 585.37, + "pct_cuda_time": 3.137205091493553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.17, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.05023322569070685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.05023322569070685, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 727.636, + "cuda_time_us": 217.20600000000002, + "pct_cuda_time": 1.1640838599568626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.964, + "cuda_time_us": 55.031, + "pct_cuda_time": 0.29493061378270446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 55.031, + "pct_cuda_time": 0.29493061378270446, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.52, + "cuda_time_us": 26.154, + "pct_cuda_time": 0.1401685463261226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.02382768819170838, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.034776848555104746, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.767, + "pct_cuda_time": 0.030907394917135008, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.02597142981939245, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 435.26, + "cuda_time_us": 90.614, + "pct_cuda_time": 0.4856325096274097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.528, + "pct_cuda_time": 0.029626509294593775, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 85.086, + "pct_cuda_time": 0.45600600033281585, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 99.899, + "cuda_time_us": 45.407, + "pct_cuda_time": 0.24335219022062582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 45.407, + "pct_cuda_time": 0.24335219022062582, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 40.24, + "cuda_time_us": 9.894, + "pct_cuda_time": 0.053025449160765355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.894, + "pct_cuda_time": 0.053025449160765355, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 229.863, + "cuda_time_us": 348.897, + "pct_cuda_time": 1.869862556685218, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 90.786, + "cuda_time_us": 208.911, + "pct_cuda_time": 1.1196280179527642, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 208.911, + "pct_cuda_time": 1.1196280179527642, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.19, + "cuda_time_us": 15.422, + "pct_cuda_time": 0.08265195845535912, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.422, + "pct_cuda_time": 0.08265195845535912, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 93.37, + "cuda_time_us": 124.56400000000001, + "pct_cuda_time": 0.6675825802770947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 118.155, + "pct_cuda_time": 0.6332344800475268, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.03434810022956793, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1089.667, + "cuda_time_us": 570.447, + "pct_cuda_time": 3.0572274507187296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.244, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 769.859, + "cuda_time_us": 213.48000000000002, + "pct_cuda_time": 1.1441149066949856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 86.499, + "cuda_time_us": 54.5, + "pct_cuda_time": 0.29208479677195387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.5, + "pct_cuda_time": 0.29208479677195387, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 93.77, + "cuda_time_us": 26.675, + "pct_cuda_time": 0.14296076979618108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.03606309353171518, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.727, + "pct_cuda_time": 0.030693020754366602, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.087, + "pct_cuda_time": 0.027263034150072092, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 445.675, + "cuda_time_us": 89.291, + "pct_cuda_time": 0.47854208419384464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.028549279126682536, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 83.964, + "pct_cuda_time": 0.4499928050671621, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 101.682, + "cuda_time_us": 43.014, + "pct_cuda_time": 0.23052725593300594, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 43.014, + "pct_cuda_time": 0.23052725593300594, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.636, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.05151947066731728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.05151947066731728, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 224.455, + "cuda_time_us": 338.021, + "pct_cuda_time": 1.8115742218284885, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.738, + "cuda_time_us": 199.657, + "pct_cuda_time": 1.0700325553962935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 199.657, + "pct_cuda_time": 1.0700325553962935, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 25.949, + "cuda_time_us": 15.252, + "pct_cuda_time": 0.0817408682635934, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.252, + "pct_cuda_time": 0.0817408682635934, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 92.848, + "cuda_time_us": 123.11200000000001, + "pct_cuda_time": 0.6598007981686016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 116.983, + "pct_cuda_time": 0.6269533170784125, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.129, + "pct_cuda_time": 0.03284748109018908, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1053.903, + "cuda_time_us": 575.104, + "pct_cuda_time": 3.0821859626190418, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.001, + "cuda_time_us": 9.292, + "pct_cuda_time": 0.04979911801110083, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.292, + "pct_cuda_time": 0.04979911801110083, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 737.821, + "cuda_time_us": 217.247, + "pct_cuda_time": 1.1643035934737003, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 83.494, + "cuda_time_us": 54.371, + "pct_cuda_time": 0.2913934400970257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.371, + "pct_cuda_time": 0.2913934400970257, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 76.183, + "cuda_time_us": 26.634, + "pct_cuda_time": 0.14274103627934348, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.485, + "pct_cuda_time": 0.02403670300040758, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.33, + "pct_cuda_time": 0.03928406532731049, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.029192401614987755, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 439.896, + "cuda_time_us": 90.935, + "pct_cuda_time": 0.4873528622836261, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.608, + "pct_cuda_time": 0.03005525762013059, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 85.327, + "pct_cuda_time": 0.4572976046634955, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 101.06, + "cuda_time_us": 45.307, + "pct_cuda_time": 0.24281625481370486, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 45.307, + "pct_cuda_time": 0.24281625481370486, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.797, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 220.229, + "cuda_time_us": 339.23199999999997, + "pct_cuda_time": 1.8180643996063017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.217, + "cuda_time_us": 198.856, + "pct_cuda_time": 1.065739712786856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 198.856, + "pct_cuda_time": 1.065739712786856, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.459, + "cuda_time_us": 14.931, + "pct_cuda_time": 0.08002051560737693, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 14.931, + "pct_cuda_time": 0.08002051560737693, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.635, + "cuda_time_us": 125.445, + "pct_cuda_time": 0.6723041712120688, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 119.077, + "pct_cuda_time": 0.6381758044993385, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.03412836671273032, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1045.631, + "cuda_time_us": 580.3309999999999, + "pct_cuda_time": 3.110199306338802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.991, + "cuda_time_us": 10.255, + "pct_cuda_time": 0.05496017597975022, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.255, + "pct_cuda_time": 0.05496017597975022, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 722.569, + "cuda_time_us": 216.13299999999998, + "pct_cuda_time": 1.1583332730406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 86.979, + "cuda_time_us": 57.504, + "pct_cuda_time": 0.30818429639586115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 57.504, + "pct_cuda_time": 0.30818429639586115, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.349, + "cuda_time_us": 25.954, + "pct_cuda_time": 0.13909667551228055, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.02382768819170838, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.328, + "pct_cuda_time": 0.033913992549961906, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.031770250922277844, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 427.388, + "cuda_time_us": 89.692, + "pct_cuda_time": 0.4806911851755979, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.028549279126682536, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 84.365, + "pct_cuda_time": 0.4521419060489153, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 98.958, + "cuda_time_us": 42.983, + "pct_cuda_time": 0.2303611159568604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.983, + "pct_cuda_time": 0.2303611159568604, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.88, + "cuda_time_us": 10.414, + "pct_cuda_time": 0.05581231327675463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.414, + "pct_cuda_time": 0.05581231327675463, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 221.811, + "cuda_time_us": 343.529, + "pct_cuda_time": 1.841093544041698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 86.389, + "cuda_time_us": 202.692, + "pct_cuda_time": 1.0862981949963464, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 202.692, + "pct_cuda_time": 1.0862981949963464, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 25.798, + "cuda_time_us": 15.402, + "pct_cuda_time": 0.08254477137397491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.402, + "pct_cuda_time": 0.08254477137397491, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.756, + "cuda_time_us": 125.435, + "pct_cuda_time": 0.6722505776713767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 119.307, + "pct_cuda_time": 0.6394084559352569, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.032842121736119875, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1040.864, + "cuda_time_us": 581.314, + "pct_cuda_time": 3.115467551388836, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.57, + "cuda_time_us": 9.734, + "pct_cuda_time": 0.052167952509691716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.734, + "pct_cuda_time": 0.052167952509691716, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 717.18, + "cuda_time_us": 217.195, + "pct_cuda_time": 1.1640249070621012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.654, + "cuda_time_us": 57.404, + "pct_cuda_time": 0.30764836098894016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 57.404, + "pct_cuda_time": 0.30764836098894016, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 72.178, + "cuda_time_us": 25.994, + "pct_cuda_time": 0.13931104967504895, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.02382768819170838, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.609, + "pct_cuda_time": 0.035419971043409965, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.607, + "pct_cuda_time": 0.03004989826606138, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 427.338, + "cuda_time_us": 90.213, + "pct_cuda_time": 0.48348340864565637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.02897802745221935, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 84.806, + "pct_cuda_time": 0.454505381193437, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 99.819, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.23358208775245576, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 43.584, + "pct_cuda_time": 0.23358208775245576, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.476, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 232.157, + "cuda_time_us": 344.53099999999995, + "pct_cuda_time": 1.846463616819046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 84.716, + "cuda_time_us": 206.047, + "pct_cuda_time": 1.1042788278985463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 206.047, + "pct_cuda_time": 1.1042788278985463, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 25.979, + "cuda_time_us": 15.051, + "pct_cuda_time": 0.08066363809568215, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.051, + "pct_cuda_time": 0.08066363809568215, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 104.536, + "cuda_time_us": 123.43299999999999, + "pct_cuda_time": 0.661521150824818, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 117.505, + "pct_cuda_time": 0.6297508999025401, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.031770250922277844, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1068.154, + "cuda_time_us": 568.855, + "pct_cuda_time": 3.048695359040547, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.436, + "cuda_time_us": 9.573, + "pct_cuda_time": 0.05130509650454888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.573, + "pct_cuda_time": 0.05130509650454888, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 744.201, + "cuda_time_us": 211.80800000000002, + "pct_cuda_time": 1.1351540666912663, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 90.996, + "cuda_time_us": 54.5, + "pct_cuda_time": 0.29208479677195387, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.5, + "pct_cuda_time": 0.29208479677195387, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 76.324, + "cuda_time_us": 25.754, + "pct_cuda_time": 0.13802480469843853, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.02382768819170838, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.03434274087549872, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.648, + "pct_cuda_time": 0.030269631782898994, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 438.765, + "cuda_time_us": 90.173, + "pct_cuda_time": 0.48326903448288794, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.028549279126682536, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 84.846, + "pct_cuda_time": 0.4547197553562055, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 102.103, + "cuda_time_us": 41.381, + "pct_cuda_time": 0.22177543073798575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 41.381, + "pct_cuda_time": 0.22177543073798575, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.689, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.05044759985347526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.05044759985347526, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 222.983, + "cuda_time_us": 338.061, + "pct_cuda_time": 1.8117885959912567, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 86.239, + "cuda_time_us": 198.976, + "pct_cuda_time": 1.0663828352751612, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 198.976, + "pct_cuda_time": 1.0663828352751612, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.72, + "cuda_time_us": 15.052, + "pct_cuda_time": 0.08066899744975135, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.052, + "pct_cuda_time": 0.08066899744975135, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 92.589, + "cuda_time_us": 124.033, + "pct_cuda_time": 0.6647367632663441, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 117.224, + "pct_cuda_time": 0.6282449214090922, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.809, + "pct_cuda_time": 0.036491841857251996, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1044.039, + "cuda_time_us": 586.942, + "pct_cuda_time": 3.145629996090351, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.42, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.05023322569070685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.05023322569070685, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 727.837, + "cuda_time_us": 217.16600000000003, + "pct_cuda_time": 1.163869485794094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 81.932, + "cuda_time_us": 57.194, + "pct_cuda_time": 0.3065228966344061, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 57.194, + "pct_cuda_time": 0.3065228966344061, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.523, + "cuda_time_us": 27.116, + "pct_cuda_time": 0.14532424494070276, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.02382768819170838, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.169, + "pct_cuda_time": 0.03842120932216765, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 6.168, + "pct_cuda_time": 0.03305649589888828, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.767, + "pct_cuda_time": 0.025548040847924846, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 430.352, + "cuda_time_us": 89.973, + "pct_cuda_time": 0.48219716366904597, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.527, + "pct_cuda_time": 0.029621149940524567, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 84.446, + "pct_cuda_time": 0.45257601372852135, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 104.135, + "cuda_time_us": 42.883, + "pct_cuda_time": 0.2298251805499394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.883, + "pct_cuda_time": 0.2298251805499394, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.75, + "cuda_time_us": 9.653, + "pct_cuda_time": 0.0517338448300857, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.653, + "pct_cuda_time": 0.0517338448300857, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 221.42, + "cuda_time_us": 350.75, + "pct_cuda_time": 1.8797934397754643, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 87.961, + "cuda_time_us": 208.641, + "pct_cuda_time": 1.1181809923540775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 208.641, + "pct_cuda_time": 1.1181809923540775, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 25.959, + "cuda_time_us": 16.653, + "pct_cuda_time": 0.08924932331455682, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 16.653, + "pct_cuda_time": 0.08924932331455682, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.405, + "cuda_time_us": 125.456, + "pct_cuda_time": 0.6723631241068302, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 119.047, + "pct_cuda_time": 0.6380150238772622, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.409, + "pct_cuda_time": 0.03434810022956793, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1038.109, + "cuda_time_us": 579.1790000000001, + "pct_cuda_time": 3.104025330451073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.839, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.05023322569070685, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.373, + "pct_cuda_time": 0.05023322569070685, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 728.046, + "cuda_time_us": 216.504, + "pct_cuda_time": 1.1603215934002769, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.303, + "cuda_time_us": 57.475, + "pct_cuda_time": 0.3080288751278541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 57.475, + "pct_cuda_time": 0.3080288751278541, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.433, + "cuda_time_us": 26.354, + "pct_cuda_time": 0.1412404171399646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.605, + "pct_cuda_time": 0.024679825488712802, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.406, + "pct_cuda_time": 0.023613314028939975, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.21, + "pct_cuda_time": 0.03864094283900527, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.527, + "pct_cuda_time": 0.029621149940524567, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 435.27, + "cuda_time_us": 91.294, + "pct_cuda_time": 0.4892768703944726, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.029406775777756165, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 85.807, + "pct_cuda_time": 0.45987009461671646, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.29, + "cuda_time_us": 41.381, + "pct_cuda_time": 0.22177543073798575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 41.381, + "pct_cuda_time": 0.22177543073798575, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.827, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.053239823323533754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.053239823323533754, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 218.436, + "cuda_time_us": 343.36800000000005, + "pct_cuda_time": 1.8402306880365555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 83.895, + "cuda_time_us": 204.073, + "pct_cuda_time": 1.0936994629659256, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 204.073, + "pct_cuda_time": 1.0936994629659256, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.639, + "cuda_time_us": 16.674, + "pct_cuda_time": 0.08936186975001024, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 16.674, + "pct_cuda_time": 0.08936186975001024, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.805, + "cuda_time_us": 122.62100000000001, + "pct_cuda_time": 0.6571693553206194, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 116.373, + "pct_cuda_time": 0.6236841110961943, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.248, + "pct_cuda_time": 0.033485244224425094, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1061.555, + "cuda_time_us": 575.774, + "pct_cuda_time": 3.085776729845412, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 32.889, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.05044759985347526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.05044759985347526, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 722.007, + "cuda_time_us": 216.35399999999998, + "pct_cuda_time": 1.1595176902898954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.053, + "cuda_time_us": 57.896, + "pct_cuda_time": 0.3102851631909916, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 57.896, + "pct_cuda_time": 0.3102851631909916, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 76.013, + "cuda_time_us": 26.113999999999997, + "pct_cuda_time": 0.13995417216335415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.02425107716317599, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.02339893986617157, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.969, + "pct_cuda_time": 0.03734933850832562, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.488, + "pct_cuda_time": 0.029412135131825373, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 428.74, + "cuda_time_us": 92.215, + "pct_cuda_time": 0.4942128354922151, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.030478646591598192, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 86.528, + "pct_cuda_time": 0.46373418890061696, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.36, + "cuda_time_us": 40.129, + "pct_cuda_time": 0.21506551944333463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 40.129, + "pct_cuda_time": 0.21506551944333463, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 60.14, + "cuda_time_us": 9.693, + "pct_cuda_time": 0.0519482189928541, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.693, + "pct_cuda_time": 0.0519482189928541, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 225.656, + "cuda_time_us": 340.314, + "pct_cuda_time": 1.8238632207091872, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 88.522, + "cuda_time_us": 200.028, + "pct_cuda_time": 1.0720208757559702, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 200.028, + "pct_cuda_time": 1.0720208757559702, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.71, + "cuda_time_us": 15.252, + "pct_cuda_time": 0.0817408682635934, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.252, + "pct_cuda_time": 0.0817408682635934, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 92.368, + "cuda_time_us": 125.034, + "pct_cuda_time": 0.6701014766896235, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 118.786, + "pct_cuda_time": 0.6366162324651984, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.248, + "pct_cuda_time": 0.033485244224425094, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1057.519, + "cuda_time_us": 577.828, + "pct_cuda_time": 3.0967848431035696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.301, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 738.522, + "cuda_time_us": 212.658, + "pct_cuda_time": 1.1397095176500947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 84.075, + "cuda_time_us": 54.951, + "pct_cuda_time": 0.29450186545716767, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.951, + "pct_cuda_time": 0.29450186545716767, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 74.772, + "cuda_time_us": 26.394999999999996, + "pct_cuda_time": 0.14146015065680223, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.366, + "pct_cuda_time": 0.02339893986617157, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.29, + "pct_cuda_time": 0.039069691164542085, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.447, + "pct_cuda_time": 0.029192401614987755, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 430.633, + "cuda_time_us": 90.813, + "pct_cuda_time": 0.4866990210871825, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.028763653289450943, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 85.446, + "pct_cuda_time": 0.45793536779773153, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 108.011, + "cuda_time_us": 40.499, + "pct_cuda_time": 0.21704848044894237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 40.499, + "pct_cuda_time": 0.21704848044894237, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.507, + "cuda_time_us": 9.494, + "pct_cuda_time": 0.050881707533081286, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.494, + "pct_cuda_time": 0.050881707533081286, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 223.213, + "cuda_time_us": 346.343, + "pct_cuda_time": 1.8561747663924555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 87.711, + "cuda_time_us": 205.536, + "pct_cuda_time": 1.10154019796918, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 205.536, + "pct_cuda_time": 1.10154019796918, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.559, + "cuda_time_us": 15.722, + "pct_cuda_time": 0.08425976467612217, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.722, + "pct_cuda_time": 0.08425976467612217, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.736, + "cuda_time_us": 125.085, + "pct_cuda_time": 0.6703748037471532, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 118.356, + "pct_cuda_time": 0.634311710215438, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.03606309353171518, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1042.326, + "cuda_time_us": 572.538, + "pct_cuda_time": 3.068433860077448, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.941, + "cuda_time_us": 9.253, + "pct_cuda_time": 0.049590103202401634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.253, + "pct_cuda_time": 0.049590103202401634, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 727.496, + "cuda_time_us": 215.862, + "pct_cuda_time": 1.156880888087844, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 83.515, + "cuda_time_us": 55.411, + "pct_cuda_time": 0.2969671683290043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 55.411, + "pct_cuda_time": 0.2969671683290043, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.513, + "cuda_time_us": 25.994, + "pct_cuda_time": 0.13931104967504895, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.024899559005550416, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.03584871936894678, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.028334904963914127, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.02597142981939245, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 433.878, + "cuda_time_us": 91.574, + "pct_cuda_time": 0.4907774895338514, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.527, + "pct_cuda_time": 0.029621149940524567, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 86.047, + "pct_cuda_time": 0.46115633959332686, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 98.998, + "cuda_time_us": 42.883, + "pct_cuda_time": 0.2298251805499394, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.883, + "pct_cuda_time": 0.2298251805499394, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.259, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.05151947066731728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.05151947066731728, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 219.888, + "cuda_time_us": 337.81, + "pct_cuda_time": 1.810443398119885, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 84.125, + "cuda_time_us": 200.699, + "pct_cuda_time": 1.0756170023364104, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 200.699, + "pct_cuda_time": 1.0756170023364104, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.02, + "cuda_time_us": 14.1, + "pct_cuda_time": 0.07556689237586328, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 14.1, + "pct_cuda_time": 0.07556689237586328, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 91.387, + "cuda_time_us": 123.01100000000001, + "pct_cuda_time": 0.6592595034076113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 116.763, + "pct_cuda_time": 0.6257742591831862, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.248, + "pct_cuda_time": 0.033485244224425094, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1039.783, + "cuda_time_us": 580.462, + "pct_cuda_time": 3.1109013817218694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.46, + "cuda_time_us": 9.894, + "pct_cuda_time": 0.053025449160765355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.894, + "pct_cuda_time": 0.053025449160765355, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 729.399, + "cuda_time_us": 217.106, + "pct_cuda_time": 1.1635479245499414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 83.304, + "cuda_time_us": 54.22, + "pct_cuda_time": 0.290584177632575, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.22, + "pct_cuda_time": 0.290584177632575, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 74.291, + "cuda_time_us": 26.595, + "pct_cuda_time": 0.14253202147064425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.209, + "pct_cuda_time": 0.03863558348493606, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.488, + "pct_cuda_time": 0.029412135131825373, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 433.107, + "cuda_time_us": 91.695, + "pct_cuda_time": 0.49142597137622585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.327, + "pct_cuda_time": 0.028549279126682536, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 86.368, + "pct_cuda_time": 0.46287669224954325, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 102.914, + "cuda_time_us": 44.596, + "pct_cuda_time": 0.2390057540704964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 44.596, + "pct_cuda_time": 0.2390057540704964, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.088, + "cuda_time_us": 9.573, + "pct_cuda_time": 0.05130509650454888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.573, + "pct_cuda_time": 0.05130509650454888, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 216.653, + "cuda_time_us": 343.889, + "pct_cuda_time": 1.8430229115066137, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.718, + "cuda_time_us": 205.736, + "pct_cuda_time": 1.1026120687830219, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 205.736, + "pct_cuda_time": 1.1026120687830219, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.529, + "cuda_time_us": 15.292, + "pct_cuda_time": 0.0819552424263618, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 15.292, + "pct_cuda_time": 0.0819552424263618, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 86.87, + "cuda_time_us": 122.861, + "pct_cuda_time": 0.6584556002972298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 116.453, + "pct_cuda_time": 0.6241128594217311, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.03434274087549872, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1081.094, + "cuda_time_us": 574.472, + "pct_cuda_time": 3.0787988508473, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.36, + "cuda_time_us": 9.774, + "pct_cuda_time": 0.05238232667246012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.774, + "pct_cuda_time": 0.05238232667246012, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 732.113, + "cuda_time_us": 216.043, + "pct_cuda_time": 1.1578509311743712, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 81.322, + "cuda_time_us": 54.961, + "pct_cuda_time": 0.29455545899785973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.961, + "pct_cuda_time": 0.29455545899785973, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.8, + "cuda_time_us": 26.354, + "pct_cuda_time": 0.1412404171399646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.023607954674870767, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.01, + "pct_cuda_time": 0.03756907202516324, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.030478646591598192, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 441.94, + "cuda_time_us": 91.775, + "pct_cuda_time": 0.49185471970176264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.608, + "pct_cuda_time": 0.03005525762013059, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 86.167, + "pct_cuda_time": 0.46179946208163203, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 99.389, + "cuda_time_us": 42.953, + "pct_cuda_time": 0.23020033533478415, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.953, + "pct_cuda_time": 0.23020033533478415, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 45.808, + "cuda_time_us": 9.573, + "pct_cuda_time": 0.05130509650454888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.573, + "pct_cuda_time": 0.05130509650454888, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 248.471, + "cuda_time_us": 339.082, + "pct_cuda_time": 1.81726049649592, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 103.855, + "cuda_time_us": 205.416, + "pct_cuda_time": 1.1008970754808747, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 205.416, + "pct_cuda_time": 1.1008970754808747, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 34.742, + "cuda_time_us": 13.058, + "pct_cuda_time": 0.0699824454357463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.058, + "pct_cuda_time": 0.0699824454357463, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 91.717, + "cuda_time_us": 120.608, + "pct_cuda_time": 0.6463809755792993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 113.839, + "pct_cuda_time": 0.6101035078848157, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.03627746769448359, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1048.225, + "cuda_time_us": 565.97, + "pct_cuda_time": 3.033233622550876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.673, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.05001885152793845, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 727.205, + "cuda_time_us": 215.573, + "pct_cuda_time": 1.1553320347618423, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 84.636, + "cuda_time_us": 53.509, + "pct_cuda_time": 0.28677367688936656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 53.509, + "pct_cuda_time": 0.28677367688936656, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.663, + "cuda_time_us": 26.955, + "pct_cuda_time": 0.1444613889355599, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.25, + "pct_cuda_time": 0.03885531700177368, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.647, + "pct_cuda_time": 0.03026427242882979, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 431.063, + "cuda_time_us": 92.216, + "pct_cuda_time": 0.4942181948462843, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.247, + "pct_cuda_time": 0.028120530801145724, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 86.969, + "pct_cuda_time": 0.46609766404513864, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 101.281, + "cuda_time_us": 42.893, + "pct_cuda_time": 0.2298787740906315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.893, + "pct_cuda_time": 0.2298787740906315, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.917, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 225.166, + "cuda_time_us": 331.21000000000004, + "pct_cuda_time": 1.7750716612630981, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 86.67, + "cuda_time_us": 195.561, + "pct_cuda_time": 1.0480806411288086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 195.561, + "pct_cuda_time": 1.0480806411288086, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.051, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.07599564070140011, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 14.18, + "pct_cuda_time": 0.07599564070140011, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 92.197, + "cuda_time_us": 121.46900000000001, + "pct_cuda_time": 0.6509953794328892, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 115.061, + "pct_cuda_time": 0.6166526385573906, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.03434274087549872, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1037.689, + "cuda_time_us": 571.789, + "pct_cuda_time": 3.0644197038796097, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.412, + "cuda_time_us": 9.814, + "pct_cuda_time": 0.052596700835228535, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.814, + "pct_cuda_time": 0.052596700835228535, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 725.533, + "cuda_time_us": 218.196, + "pct_cuda_time": 1.1693896204853806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 84.967, + "cuda_time_us": 54.781, + "pct_cuda_time": 0.2935907752654019, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.781, + "pct_cuda_time": 0.2935907752654019, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.322, + "cuda_time_us": 26.875, + "pct_cuda_time": 0.1440326406100231, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.608, + "pct_cuda_time": 0.03541461168934075, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 6.049, + "pct_cuda_time": 0.03241873276465227, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.02597142981939245, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 431.143, + "cuda_time_us": 93.217, + "pct_cuda_time": 0.4995829082695637, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.727, + "pct_cuda_time": 0.030693020754366602, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 87.49, + "pct_cuda_time": 0.4688898875151971, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 98.087, + "cuda_time_us": 43.323, + "pct_cuda_time": 0.23218329634039186, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 43.323, + "pct_cuda_time": 0.23218329634039186, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.295, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.05044759985347526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.413, + "pct_cuda_time": 0.05044759985347526, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 217.144, + "cuda_time_us": 334.366, + "pct_cuda_time": 1.7919857827055252, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 83.394, + "cuda_time_us": 201.941, + "pct_cuda_time": 1.0822733200903694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 201.941, + "pct_cuda_time": 1.0822733200903694, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.279, + "cuda_time_us": 13.179, + "pct_cuda_time": 0.07063092727812073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.179, + "pct_cuda_time": 0.07063092727812073, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 89.424, + "cuda_time_us": 119.246, + "pct_cuda_time": 0.6390815353370349, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 112.797, + "pct_cuda_time": 0.6045190609446986, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.449, + "pct_cuda_time": 0.03456247439233634, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1030.018, + "cuda_time_us": 567.894, + "pct_cuda_time": 3.0435450197800358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.07, + "cuda_time_us": 9.773, + "pct_cuda_time": 0.05237696731839091, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.773, + "pct_cuda_time": 0.05237696731839091, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 724.872, + "cuda_time_us": 217.718, + "pct_cuda_time": 1.166827849240298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.924, + "cuda_time_us": 54.661, + "pct_cuda_time": 0.2929476527770967, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.661, + "pct_cuda_time": 0.2929476527770967, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.059, + "cuda_time_us": 27.236, + "pct_cuda_time": 0.14596736742900798, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.89, + "pct_cuda_time": 0.04228530360606818, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.448, + "pct_cuda_time": 0.02919776096905697, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 428.44, + "cuda_time_us": 93.49799999999999, + "pct_cuda_time": 0.5010888867630117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.407, + "pct_cuda_time": 0.02897802745221935, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 88.091, + "pct_cuda_time": 0.4721108593107924, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 98.016, + "cuda_time_us": 42.323, + "pct_cuda_time": 0.2268239422711817, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.323, + "pct_cuda_time": 0.2268239422711817, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.745, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.05151947066731728, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.05151947066731728, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 214.44, + "cuda_time_us": 330.79, + "pct_cuda_time": 1.77282073255403, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 84.236, + "cuda_time_us": 196.953, + "pct_cuda_time": 1.0555408619931492, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 196.953, + "pct_cuda_time": 1.0555408619931492, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.57, + "cuda_time_us": 13.619, + "pct_cuda_time": 0.0729890430685732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.619, + "pct_cuda_time": 0.0729890430685732, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 86.399, + "cuda_time_us": 120.218, + "pct_cuda_time": 0.6442908274923074, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 113.649, + "pct_cuda_time": 0.6090852306116658, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.569, + "pct_cuda_time": 0.03520559688064156, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1051.77, + "cuda_time_us": 569.144, + "pct_cuda_time": 3.050244212366549, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.211, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.053239823323533754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.053239823323533754, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 732.494, + "cuda_time_us": 222.833, + "pct_cuda_time": 1.1942409453043081, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.073, + "cuda_time_us": 53.919, + "pct_cuda_time": 0.2889710120577427, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 53.919, + "pct_cuda_time": 0.2889710120577427, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.169, + "cuda_time_us": 26.715000000000003, + "pct_cuda_time": 0.1431751439589495, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.129, + "pct_cuda_time": 0.038206835159399245, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.488, + "pct_cuda_time": 0.029412135131825373, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.026400178144929263, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 440.788, + "cuda_time_us": 94.05799999999999, + "pct_cuda_time": 0.5040901250417694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.029406775777756165, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 88.571, + "pct_cuda_time": 0.47468334926401323, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 101.0, + "cuda_time_us": 48.141, + "pct_cuda_time": 0.25800466424584645, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 48.141, + "pct_cuda_time": 0.25800466424584645, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.698, + "cuda_time_us": 10.134, + "pct_cuda_time": 0.054311694137375785, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.134, + "pct_cuda_time": 0.054311694137375785, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 224.606, + "cuda_time_us": 326.243, + "pct_cuda_time": 1.7484517496013312, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 86.009, + "cuda_time_us": 194.64, + "pct_cuda_time": 1.0431446760310659, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 194.64, + "pct_cuda_time": 1.0431446760310659, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.89, + "cuda_time_us": 13.219, + "pct_cuda_time": 0.07084530144088913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.219, + "pct_cuda_time": 0.07084530144088913, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 93.25, + "cuda_time_us": 118.384, + "pct_cuda_time": 0.6344617721293758, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 112.016, + "pct_cuda_time": 0.6003334054166456, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.03412836671273032, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1039.582, + "cuda_time_us": 566.03, + "pct_cuda_time": 3.033555183795028, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.512, + "cuda_time_us": 9.654, + "pct_cuda_time": 0.05173920418415491, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.654, + "pct_cuda_time": 0.05173920418415491, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 719.764, + "cuda_time_us": 220.63000000000002, + "pct_cuda_time": 1.1824342882898382, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 83.975, + "cuda_time_us": 54.45, + "pct_cuda_time": 0.2918168290684934, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.45, + "pct_cuda_time": 0.2918168290684934, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 74.26, + "cuda_time_us": 26.994999999999997, + "pct_cuda_time": 0.14467576309832833, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.21, + "pct_cuda_time": 0.03864094283900527, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.807, + "pct_cuda_time": 0.031121769079903414, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 424.514, + "cuda_time_us": 95.10000000000001, + "pct_cuda_time": 0.5096745719818865, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.528, + "pct_cuda_time": 0.029626509294593775, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 89.572, + "pct_cuda_time": 0.4800480626872927, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.661, + "cuda_time_us": 44.085, + "pct_cuda_time": 0.23626712414113, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 44.085, + "pct_cuda_time": 0.23626712414113, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.727, + "cuda_time_us": 9.694, + "pct_cuda_time": 0.05195357834692332, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.694, + "pct_cuda_time": 0.05195357834692332, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 225.767, + "cuda_time_us": 326.052, + "pct_cuda_time": 1.7474281129741118, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.267, + "cuda_time_us": 196.272, + "pct_cuda_time": 1.051891141872017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 196.272, + "pct_cuda_time": 1.051891141872017, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.251, + "cuda_time_us": 12.978, + "pct_cuda_time": 0.06955369711020949, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 12.978, + "pct_cuda_time": 0.06955369711020949, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 94.101, + "cuda_time_us": 116.80199999999999, + "pct_cuda_time": 0.6259832739918854, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 110.234, + "pct_cuda_time": 0.590783036465313, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.568, + "pct_cuda_time": 0.035200237526572344, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1040.914, + "cuda_time_us": 559.521, + "pct_cuda_time": 2.998671148158539, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.002, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.053239823323533754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.934, + "pct_cuda_time": 0.053239823323533754, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 726.825, + "cuda_time_us": 221.98299999999998, + "pct_cuda_time": 1.1896854943454793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.312, + "cuda_time_us": 54.18, + "pct_cuda_time": 0.2903698034698066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.18, + "pct_cuda_time": 0.2903698034698066, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.28, + "cuda_time_us": 27.876999999999995, + "pct_cuda_time": 0.14940271338737168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.024042062354476788, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.851, + "pct_cuda_time": 0.04207628879736899, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.767, + "pct_cuda_time": 0.030907394917135008, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.02683428582453528, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 435.66, + "cuda_time_us": 95.901, + "pct_cuda_time": 0.5139674145913238, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.030478646591598192, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 90.214, + "pct_cuda_time": 0.4834887679997256, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.671, + "cuda_time_us": 44.025, + "pct_cuda_time": 0.2359455628969774, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 44.025, + "pct_cuda_time": 0.2359455628969774, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.227, + "cuda_time_us": 9.734, + "pct_cuda_time": 0.052167952509691716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.734, + "pct_cuda_time": 0.052167952509691716, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 219.748, + "cuda_time_us": 317.87, + "pct_cuda_time": 1.7035778779798345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 83.165, + "cuda_time_us": 187.478, + "pct_cuda_time": 1.0047609821873829, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 187.478, + "pct_cuda_time": 1.0047609821873829, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.85, + "cuda_time_us": 13.179, + "pct_cuda_time": 0.07063092727812073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.179, + "pct_cuda_time": 0.07063092727812073, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 91.476, + "cuda_time_us": 117.21300000000001, + "pct_cuda_time": 0.6281859685143308, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 110.724, + "pct_cuda_time": 0.593409119959226, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.034776848555104746, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1038.139, + "cuda_time_us": 566.0409999999999, + "pct_cuda_time": 3.0336141366897893, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.121, + "cuda_time_us": 9.734, + "pct_cuda_time": 0.052167952509691716, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.734, + "pct_cuda_time": 0.052167952509691716, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 723.76, + "cuda_time_us": 223.214, + "pct_cuda_time": 1.1962828592046773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 84.496, + "cuda_time_us": 54.029, + "pct_cuda_time": 0.2895605410053559, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.029, + "pct_cuda_time": 0.2895605410053559, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 73.73, + "cuda_time_us": 28.076999999999998, + "pct_cuda_time": 0.1504745842012137, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 8.372, + "pct_cuda_time": 0.04486851226742748, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.767, + "pct_cuda_time": 0.030907394917135008, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 429.371, + "cuda_time_us": 96.663, + "pct_cuda_time": 0.5180512423920619, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.648, + "pct_cuda_time": 0.030269631782898994, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 91.015, + "pct_cuda_time": 0.487781610609163, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.339, + "cuda_time_us": 44.445, + "pct_cuda_time": 0.23819649160604564, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 44.445, + "pct_cuda_time": 0.23819649160604564, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.597, + "cuda_time_us": 9.814, + "pct_cuda_time": 0.052596700835228535, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.814, + "pct_cuda_time": 0.052596700835228535, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 219.338, + "cuda_time_us": 323.279, + "pct_cuda_time": 1.732566624140192, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.347, + "cuda_time_us": 192.917, + "pct_cuda_time": 1.0339105089698168, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 192.917, + "pct_cuda_time": 1.0339105089698168, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.78, + "cuda_time_us": 13.579, + "pct_cuda_time": 0.07277466890580479, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.579, + "pct_cuda_time": 0.07277466890580479, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 89.073, + "cuda_time_us": 116.78299999999999, + "pct_cuda_time": 0.6258814462645703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 110.094, + "pct_cuda_time": 0.5900327268956236, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.03584871936894678, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1059.412, + "cuda_time_us": 564.609, + "pct_cuda_time": 3.0259395416626806, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.3, + "cuda_time_us": 9.252, + "pct_cuda_time": 0.04958474384833243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.252, + "pct_cuda_time": 0.04958474384833243, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 740.936, + "cuda_time_us": 222.79500000000002, + "pct_cuda_time": 1.1940372898496783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 85.658, + "cuda_time_us": 54.441, + "pct_cuda_time": 0.29176859488187046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 54.441, + "pct_cuda_time": 0.29176859488187046, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 86.018, + "cuda_time_us": 27.115000000000002, + "pct_cuda_time": 0.14531888558663358, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.024042062354476788, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.33, + "pct_cuda_time": 0.03928406532731049, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.687, + "pct_cuda_time": 0.030478646591598192, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 431.955, + "cuda_time_us": 96.743, + "pct_cuda_time": 0.5184799907175988, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.568, + "pct_cuda_time": 0.02984088345736218, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 91.175, + "pct_cuda_time": 0.4886391072602365, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.99, + "cuda_time_us": 44.496, + "pct_cuda_time": 0.2384698186635754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 44.496, + "pct_cuda_time": 0.2384698186635754, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 40.741, + "cuda_time_us": 10.495, + "pct_cuda_time": 0.056246420956360656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.495, + "pct_cuda_time": 0.056246420956360656, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 223.534, + "cuda_time_us": 322.067, + "pct_cuda_time": 1.7260710870083094, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 87.0, + "cuda_time_us": 193.327, + "pct_cuda_time": 1.036107844138193, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 193.327, + "pct_cuda_time": 1.036107844138193, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.801, + "cuda_time_us": 13.259, + "pct_cuda_time": 0.07105967560365754, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.259, + "pct_cuda_time": 0.07105967560365754, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.275, + "cuda_time_us": 115.481, + "pct_cuda_time": 0.6189035672664587, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 109.393, + "pct_cuda_time": 0.5862758196931073, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.088, + "pct_cuda_time": 0.03262774757335147, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1053.263, + "cuda_time_us": 555.864, + "pct_cuda_time": 2.979071990327438, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 33.35, + "cuda_time_us": 9.853, + "pct_cuda_time": 0.05280571564392773, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.853, + "pct_cuda_time": 0.05280571564392773, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 737.011, + "cuda_time_us": 220.25000000000003, + "pct_cuda_time": 1.1803977337435383, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 85.387, + "cuda_time_us": 51.727, + "pct_cuda_time": 0.27722330793803407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 51.727, + "pct_cuda_time": 0.27722330793803407, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 76.344, + "cuda_time_us": 26.955000000000002, + "pct_cuda_time": 0.14446138893555993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.024465451325944392, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.05, + "pct_cuda_time": 0.03778344618793164, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.031770250922277844, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 440.117, + "cuda_time_us": 97.66300000000001, + "pct_cuda_time": 0.5234105964612722, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.727, + "pct_cuda_time": 0.030693020754366602, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 91.936, + "pct_cuda_time": 0.49271757570690555, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 99.148, + "cuda_time_us": 43.905, + "pct_cuda_time": 0.2353024404086722, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 43.905, + "pct_cuda_time": 0.2353024404086722, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.286, + "cuda_time_us": 9.894, + "pct_cuda_time": 0.053025449160765355, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.894, + "pct_cuda_time": 0.053025449160765355, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 224.315, + "cuda_time_us": 315.867, + "pct_cuda_time": 1.6928430917792066, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.487, + "cuda_time_us": 186.217, + "pct_cuda_time": 0.9980028367061088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 186.217, + "pct_cuda_time": 0.9980028367061088, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.512, + "cuda_time_us": 13.218, + "pct_cuda_time": 0.07083994208681993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.218, + "pct_cuda_time": 0.07083994208681993, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 92.428, + "cuda_time_us": 116.432, + "pct_cuda_time": 0.6240003129862777, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 109.343, + "pct_cuda_time": 0.5860078519896468, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 7.089, + "pct_cuda_time": 0.037992460996630846, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1073.152, + "cuda_time_us": 565.168, + "pct_cuda_time": 3.028935420587369, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.363, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 751.081, + "cuda_time_us": 223.914, + "pct_cuda_time": 1.2000344070531241, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 89.103, + "cuda_time_us": 53.509, + "pct_cuda_time": 0.28677367688936656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 53.509, + "pct_cuda_time": 0.28677367688936656, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 74.892, + "cuda_time_us": 27.876, + "pct_cuda_time": 0.14939735403330248, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.024256436517245194, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.41, + "pct_cuda_time": 0.039712813652847304, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 6.248, + "pct_cuda_time": 0.033485244224425094, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.026614552307697666, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 447.869, + "cuda_time_us": 97.18299999999999, + "pct_cuda_time": 0.5208381065080512, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.728, + "pct_cuda_time": 0.030698380108435806, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 91.455, + "pct_cuda_time": 0.4901397263996154, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 102.123, + "cuda_time_us": 45.346, + "pct_cuda_time": 0.243025269622404, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 45.346, + "pct_cuda_time": 0.243025269622404, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.388, + "cuda_time_us": 10.775, + "pct_cuda_time": 0.057747040095739506, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.775, + "pct_cuda_time": 0.057747040095739506, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 226.098, + "cuda_time_us": 320.625, + "pct_cuda_time": 1.7183428984405082, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 86.199, + "cuda_time_us": 192.607, + "pct_cuda_time": 1.0322491092083619, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 192.607, + "pct_cuda_time": 1.0322491092083619, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.57, + "cuda_time_us": 14.06, + "pct_cuda_time": 0.07535251821309488, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 14.06, + "pct_cuda_time": 0.07535251821309488, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 94.992, + "cuda_time_us": 113.958, + "pct_cuda_time": 0.6107412710190517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 107.469, + "pct_cuda_time": 0.5759644224639469, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.034776848555104746, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1092.201, + "cuda_time_us": 552.309, + "pct_cuda_time": 2.9600194866113956, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.152, + "cuda_time_us": 9.774, + "pct_cuda_time": 0.05238232667246012, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.774, + "pct_cuda_time": 0.05238232667246012, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 749.829, + "cuda_time_us": 218.03599999999997, + "pct_cuda_time": 1.1685321238343067, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 89.694, + "cuda_time_us": 51.516, + "pct_cuda_time": 0.2760924842294307, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 51.516, + "pct_cuda_time": 0.2760924842294307, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.012, + "cuda_time_us": 27.355, + "pct_cuda_time": 0.146605130563244, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.024465451325944392, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.37, + "pct_cuda_time": 0.0394984394900789, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.848, + "pct_cuda_time": 0.03134150259674103, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.02597142981939245, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 448.309, + "cuda_time_us": 98.185, + "pct_cuda_time": 0.5262081792853998, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.030912754271204216, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 92.417, + "pct_cuda_time": 0.49529542501419555, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.43, + "cuda_time_us": 40.98, + "pct_cuda_time": 0.21962632975623245, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 40.98, + "pct_cuda_time": 0.21962632975623245, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 39.098, + "cuda_time_us": 10.535, + "pct_cuda_time": 0.05646079511912906, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.535, + "pct_cuda_time": 0.05646079511912906, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 245.116, + "cuda_time_us": 313.964, + "pct_cuda_time": 1.6826442409854994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 86.549, + "cuda_time_us": 186.267, + "pct_cuda_time": 0.9982708044095694, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 186.267, + "pct_cuda_time": 0.9982708044095694, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 42.523, + "cuda_time_us": 13.299, + "pct_cuda_time": 0.07127404976642596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.299, + "pct_cuda_time": 0.07127404976642596, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 97.075, + "cuda_time_us": 114.39800000000001, + "pct_cuda_time": 0.6130993868095043, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 107.79, + "pct_cuda_time": 0.5776847751201635, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.608, + "pct_cuda_time": 0.03541461168934075, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1061.775, + "cuda_time_us": 549.547, + "pct_cuda_time": 2.945216950672237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.004, + "cuda_time_us": 10.295, + "pct_cuda_time": 0.05517455014251862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.295, + "pct_cuda_time": 0.05517455014251862, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 744.201, + "cuda_time_us": 218.11700000000002, + "pct_cuda_time": 1.168966231513913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 86.379, + "cuda_time_us": 51.796, + "pct_cuda_time": 0.2775931033688096, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 51.796, + "pct_cuda_time": 0.2775931033688096, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 77.876, + "cuda_time_us": 26.755, + "pct_cuda_time": 0.1433895181217179, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.926, + "pct_cuda_time": 0.026400178144929263, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 6.689, + "pct_cuda_time": 0.03584871936894678, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 6.289, + "pct_cuda_time": 0.03370497774126271, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.285, + "pct_cuda_time": 0.022964832186565545, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 442.641, + "cuda_time_us": 98.265, + "pct_cuda_time": 0.5266369276109366, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.031770250922277844, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 92.337, + "pct_cuda_time": 0.4948666766886588, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.801, + "cuda_time_us": 41.301, + "pct_cuda_time": 0.22134668241244895, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 41.301, + "pct_cuda_time": 0.22134668241244895, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.217, + "cuda_time_us": 9.974, + "pct_cuda_time": 0.05345419748630216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.974, + "pct_cuda_time": 0.05345419748630216, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 221.771, + "cuda_time_us": 311.16100000000006, + "pct_cuda_time": 1.6676219715295038, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.718, + "cuda_time_us": 182.942, + "pct_cuda_time": 0.9804509521294456, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 182.942, + "pct_cuda_time": 0.9804509521294456, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 27.921, + "cuda_time_us": 13.419, + "pct_cuda_time": 0.07191717225473117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.419, + "pct_cuda_time": 0.07191717225473117, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 89.454, + "cuda_time_us": 114.80000000000001, + "pct_cuda_time": 0.6152538471453267, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 108.311, + "pct_cuda_time": 0.580476998590222, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.489, + "pct_cuda_time": 0.034776848555104746, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1045.05, + "cuda_time_us": 556.697, + "pct_cuda_time": 2.98353633226709, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.563, + "cuda_time_us": 10.054, + "pct_cuda_time": 0.05388294581183898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.054, + "pct_cuda_time": 0.05388294581183898, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 729.459, + "cuda_time_us": 220.10000000000002, + "pct_cuda_time": 1.1795938306331568, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 82.693, + "cuda_time_us": 50.654, + "pct_cuda_time": 0.27147272102177156, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 50.654, + "pct_cuda_time": 0.27147272102177156, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.232, + "cuda_time_us": 27.396, + "pct_cuda_time": 0.1468248640800816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.691, + "pct_cuda_time": 0.041218792146295355, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.487, + "pct_cuda_time": 0.029406775777756165, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.02532830733108723, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 435.55, + "cuda_time_us": 99.06700000000001, + "pct_cuda_time": 0.5309351295744432, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.688, + "pct_cuda_time": 0.030484005945667404, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 93.379, + "pct_cuda_time": 0.5004511236287759, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.771, + "cuda_time_us": 42.983, + "pct_cuda_time": 0.2303611159568604, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.983, + "pct_cuda_time": 0.2303611159568604, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.648, + "cuda_time_us": 10.495, + "pct_cuda_time": 0.056246420956360656, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.495, + "pct_cuda_time": 0.056246420956360656, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 221.02, + "cuda_time_us": 316.048, + "pct_cuda_time": 1.6938131348657335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 87.681, + "cuda_time_us": 188.891, + "pct_cuda_time": 1.012333749487177, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 188.891, + "pct_cuda_time": 1.012333749487177, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 25.929, + "cuda_time_us": 13.299, + "pct_cuda_time": 0.07127404976642596, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.299, + "pct_cuda_time": 0.07127404976642596, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 89.915, + "cuda_time_us": 113.858, + "pct_cuda_time": 0.6102053356121308, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 107.089, + "pct_cuda_time": 0.573927867917647, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.03627746769448359, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1044.239, + "cuda_time_us": 552.17, + "pct_cuda_time": 2.959274536395775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.863, + "cuda_time_us": 10.254, + "pct_cuda_time": 0.054954816625681004, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.254, + "pct_cuda_time": 0.054954816625681004, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 732.324, + "cuda_time_us": 221.51299999999998, + "pct_cuda_time": 1.1871665979329507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 84.106, + "cuda_time_us": 51.747, + "pct_cuda_time": 0.27733049501941825, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 51.747, + "pct_cuda_time": 0.27733049501941825, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 74.191, + "cuda_time_us": 27.836, + "pct_cuda_time": 0.14918297987053405, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.024899559005550416, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.37, + "pct_cuda_time": 0.0394984394900789, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 6.008, + "pct_cuda_time": 0.03219899924781466, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.026828926470466075, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 437.453, + "cuda_time_us": 99.427, + "pct_cuda_time": 0.5328644970393589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.768, + "pct_cuda_time": 0.030912754271204216, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 93.659, + "pct_cuda_time": 0.5019517427681547, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.691, + "cuda_time_us": 42.503, + "pct_cuda_time": 0.22778862600363955, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.503, + "pct_cuda_time": 0.22778862600363955, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 36.905, + "cuda_time_us": 10.174, + "pct_cuda_time": 0.054526068300144184, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.174, + "pct_cuda_time": 0.054526068300144184, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 218.316, + "cuda_time_us": 310.229, + "pct_cuda_time": 1.6626270535369994, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 84.767, + "cuda_time_us": 182.001, + "pct_cuda_time": 0.9754077999503188, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 182.001, + "pct_cuda_time": 0.9754077999503188, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.75, + "cuda_time_us": 13.659, + "pct_cuda_time": 0.07320341723134162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.659, + "pct_cuda_time": 0.07320341723134162, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 89.484, + "cuda_time_us": 114.569, + "pct_cuda_time": 0.6140158363553391, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 108.0, + "pct_cuda_time": 0.5788102394746976, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.569, + "pct_cuda_time": 0.03520559688064156, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1067.593, + "cuda_time_us": 551.428, + "pct_cuda_time": 2.9552978956764213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 34.351, + "cuda_time_us": 10.094, + "pct_cuda_time": 0.05409731997460738, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.094, + "pct_cuda_time": 0.05409731997460738, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 748.006, + "cuda_time_us": 219.389, + "pct_cuda_time": 1.1757833298899483, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 81.932, + "cuda_time_us": 49.063, + "pct_cuda_time": 0.26294598869765823, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 49.063, + "pct_cuda_time": 0.26294598869765823, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 74.621, + "cuda_time_us": 27.916, + "pct_cuda_time": 0.1496117281960709, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.026185803982160853, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.025113933168318822, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.53, + "pct_cuda_time": 0.04035593614115252, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.031770250922277844, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.026185803982160853, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 436.341, + "cuda_time_us": 99.547, + "pct_cuda_time": 0.533507619527664, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.808, + "pct_cuda_time": 0.031127128433972626, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 93.739, + "pct_cuda_time": 0.5023804910936914, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 120.931, + "cuda_time_us": 42.863, + "pct_cuda_time": 0.2297179934685552, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 42.863, + "pct_cuda_time": 0.2297179934685552, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 40.04, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.854, + "pct_cuda_time": 0.052811074997996935, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 223.173, + "cuda_time_us": 312.091, + "pct_cuda_time": 1.6726061708138686, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.107, + "cuda_time_us": 184.734, + "pct_cuda_time": 0.9900549146214701, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 184.734, + "pct_cuda_time": 0.9900549146214701, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.64, + "cuda_time_us": 13.5, + "pct_cuda_time": 0.0723512799343372, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.5, + "pct_cuda_time": 0.0723512799343372, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 93.349, + "cuda_time_us": 113.857, + "pct_cuda_time": 0.6101999762580614, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 106.928, + "pct_cuda_time": 0.5730650119125043, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.929, + "pct_cuda_time": 0.037134964345557214, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1049.467, + "cuda_time_us": 552.862, + "pct_cuda_time": 2.9629832094116684, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.342, + "cuda_time_us": 10.295, + "pct_cuda_time": 0.05517455014251862, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.295, + "pct_cuda_time": 0.05517455014251862, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 737.111, + "cuda_time_us": 221.022, + "pct_cuda_time": 1.1845351550849683, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 84.286, + "cuda_time_us": 50.374, + "pct_cuda_time": 0.26997210188239273, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 50.374, + "pct_cuda_time": 0.26997210188239273, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 72.628, + "cuda_time_us": 28.197, + "pct_cuda_time": 0.15111770668951896, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.026185803982160853, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.531, + "pct_cuda_time": 0.04036129549522173, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 6.008, + "pct_cuda_time": 0.03219899924781466, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.026828926470466075, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 440.988, + "cuda_time_us": 100.549, + "pct_cuda_time": 0.5388776923050126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.608, + "pct_cuda_time": 0.03005525762013059, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 94.941, + "pct_cuda_time": 0.508822434684882, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 102.934, + "cuda_time_us": 41.902, + "pct_cuda_time": 0.22456765420804425, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 41.902, + "pct_cuda_time": 0.22456765420804425, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.556, + "cuda_time_us": 10.414, + "pct_cuda_time": 0.05581231327675463, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.414, + "pct_cuda_time": 0.05581231327675463, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 218.046, + "cuda_time_us": 311.131, + "pct_cuda_time": 1.667461190907427, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 84.376, + "cuda_time_us": 184.184, + "pct_cuda_time": 0.9871072698834046, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 184.184, + "pct_cuda_time": 0.9871072698834046, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.119, + "cuda_time_us": 13.339, + "pct_cuda_time": 0.07148842392919436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.339, + "pct_cuda_time": 0.07148842392919436, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 88.932, + "cuda_time_us": 113.60799999999999, + "pct_cuda_time": 0.6088654970948281, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 107.079, + "pct_cuda_time": 0.5738742743769549, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.03499122271787315, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 1044.649, + "cuda_time_us": 547.362, + "pct_cuda_time": 2.933506762031013, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.023, + "cuda_time_us": 10.014, + "pct_cuda_time": 0.05366857164907056, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.014, + "pct_cuda_time": 0.05366857164907056, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 730.841, + "cuda_time_us": 219.50900000000001, + "pct_cuda_time": 1.1764264523782537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 83.234, + "cuda_time_us": 51.706, + "pct_cuda_time": 0.2771107615025807, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x128x128_MI16x16x1_SN_LDSB1_AFC0_AFEM8_AFEM8_ASEM32_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_4_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 51.706, + "pct_cuda_time": 0.2771107615025807, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 6144]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 6144]) <- linear(bfloat16[256, 4096], bfloat16[6144, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 75.873, + "cuda_time_us": 27.275000000000002, + "pct_cuda_time": 0.14617638223770718, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.026185803982160853, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False) <- clone(bfloat16[256, 4096], 0) <- contiguous(bfloat16[256, 4096], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.024685184842782006, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False) <- clone(bfloat16[256, 1024], 0) <- contiguous(bfloat16[256, 1024], 0)" + }, + "children": [] + }, + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 7.049, + "pct_cuda_time": 0.03777808683386244, + "trace": "_C::rotary_embedding(int64[256], bfloat16[256, 4096], bfloat16[256, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 5.928, + "pct_cuda_time": 0.031770250922277844, + "trace": "copy_(bfloat16[256, 4096], bfloat16[256, 4096], False)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#12}::operator()() const::{lambda(c10::BFloat16)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[256, 1024], bfloat16[256, 1024], False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 436.081, + "cuda_time_us": 100.74900000000001, + "pct_cuda_time": 0.5399495631188548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.968, + "pct_cuda_time": 0.03198462508504625, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[256], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "attn_fwd", + "cpu_time_us": 0, + "cuda_time_us": 94.781, + "pct_cuda_time": 0.5079649380338084, + "trace": "_attention(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], int32[2], int32[2], 256, 256, True, 0.08838834764831845, None, None, None) <- vllm::unified_attention_with_output(bfloat16[256, 32, 128], bfloat16[256, 8, 128], bfloat16[256, 8, 128], bfloat16[256, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 100.39, + "cuda_time_us": 39.779, + "pct_cuda_time": 0.21318974551911107, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT64x64x128_MI16x16x1_SN_LDSB1_AFC0_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA512_LBSPPB256_LBSPPM0_LPA16_LPB16_LPMn1_LRVW8_LWPMn1_MIAV0_MIWT4_2_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB1_WSGRA0_WSGRB0_WS64_WG16_8_2", + "cpu_time_us": 0, + "cuda_time_us": 39.779, + "pct_cuda_time": 0.21318974551911107, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 4096]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 4096]) <- linear(bfloat16[256, 4096], bfloat16[4096, 4096], None)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 38.147, + "cuda_time_us": 10.054, + "pct_cuda_time": 0.05388294581183898, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.054, + "pct_cuda_time": 0.05388294581183898, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 219.478, + "cuda_time_us": 307.78499999999997, + "pct_cuda_time": 1.6495287921918498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 85.207, + "cuda_time_us": 179.016, + "pct_cuda_time": 0.9594101280537264, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT96x256x64_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA256_LBSPPB512_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT6_4_MO40_NTn1_NTA4_NTB0_NTC0_NTD4_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW2_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA2_VWB4_WSGRA0_WSGRB0_WS64_WG16_16_1", + "cpu_time_us": 0, + "cuda_time_us": 179.016, + "pct_cuda_time": 0.9594101280537264, + "trace": "mm(bfloat16[256, 4096], bfloat16[4096, 28672]) <- matmul(bfloat16[256, 4096], bfloat16[4096, 28672]) <- linear(bfloat16[256, 4096], bfloat16[28672, 4096], None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 26.329, + "cuda_time_us": 13.579, + "pct_cuda_time": 0.07277466890580479, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 13.579, + "pct_cuda_time": 0.07277466890580479, + "trace": "_C::silu_and_mul(bfloat16[256, 14336], bfloat16[256, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 90.956, + "cuda_time_us": 115.19, + "pct_cuda_time": 0.6173439952323185, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Cijk_Alik_Bljk_BBS_BH_Bias_HAS_SAV_UserArgs_MT128x96x128_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_DTVA0_DTVB0_EPS0_FDSI0_GRPM1_GRVWA8_GRVWB8_GSUAMB_GLS0_ISA942_IU1_K1_LBSPPA1024_LBSPPB256_LBSPPM0_LPA16_LPB16_LPM0_LRVW8_LWPMn1_MIAV0_MIWT4_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS16_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW4_SK0_SKXCCM0_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA4_VWB1_WSGRA0_WSGRB0_WS64_WG32_8_1", + "cpu_time_us": 0, + "cuda_time_us": 108.301, + "pct_cuda_time": 0.5804234050495298, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + }, + { + "entry": { + "name": "Cijk_SB_BiasS_HAS_ScaleAlphaVec_PostGSU4_VW4", + "cpu_time_us": 0, + "cuda_time_us": 6.889, + "pct_cuda_time": 0.03692059018278881, + "trace": "mm(bfloat16[256, 14336], bfloat16[14336, 4096]) <- matmul(bfloat16[256, 14336], bfloat16[14336, 4096]) <- linear(bfloat16[256, 14336], bfloat16[4096, 14336], None)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.943, + "cuda_time_us": 9.974, + "pct_cuda_time": 0.05345419748630216, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 9.974, + "pct_cuda_time": 0.05345419748630216, + "trace": "_C::fused_add_rms_norm(bfloat16[256, 4096], bfloat16[256, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 142.503, + "cuda_time_us": 241.549, + "pct_cuda_time": 1.2945466160636456, + "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": 4.566, + "pct_cuda_time": 0.0244708106800136, + "trace": "index_select(bfloat16[256, 4096], 0, int64[1])" + }, + "children": [] + }, + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 236.983, + "pct_cuda_time": 1.270075805383632, + "trace": "_rocm_C::wvSplitK(bfloat16[128256, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 804.01, + "cuda_time_us": 226.40500000000003, + "pct_cuda_time": 1.213384558039527, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.02575705565662404, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.02597142981939245, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.026405537498998467, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.026828926470466075, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.927, + "pct_cuda_time": 0.026405537498998467, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.022321709698260327, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 2.443, + "pct_cuda_time": 0.013092901991080427, + "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 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)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", + "cpu_time_us": 0, + "cuda_time_us": 14.141, + "pct_cuda_time": 0.07578662589270091, + "trace": "div_(float32[1, 128256], bfloat16[1, 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": 70.985, + "pct_cuda_time": 0.3804337486028834, + "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 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": 40.42, + "pct_cuda_time": 0.2166250914774748, + "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.025542681493855638, + "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 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": 8.251, + "pct_cuda_time": 0.04422003042505305, + "trace": "index(float32[1, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 46.949, + "pct_cuda_time": 0.2516163141953479, + "trace": "argmax(float32[1, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Host)", + "cpu_time_us": 0, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.02683428582453528, + "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + }, + "decode_1": { + "metadata": { + "num_running_seqs": 1 + }, + "summary_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cuda_time_us": 5292.264999999999, + "pct_cuda_time": 92.2104376570522, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cuda_time_us": 3.844, + "pct_cuda_time": 0.0669764122457414, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 3.844, + "pct_cuda_time": 0.0669764122457414, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cuda_time_us": 5284.095, + "pct_cuda_time": 92.06808664559338, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 304.3069999999999, + "pct_cuda_time": 5.302130874418529, + "invocations": 64 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", + "cuda_time_us": 5.207, + "pct_cuda_time": 0.09072481232142963, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 299.0999999999999, + "pct_cuda_time": 5.211406062097099, + "invocations": 63 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cuda_time_us": 1559.3070000000002, + "pct_cuda_time": 27.16877951344181, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cuda_time_us": 564.5269999999999, + "pct_cuda_time": 9.836106419316247, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cuda_time_us": 564.5269999999999, + "pct_cuda_time": 9.836106419316247, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cuda_time_us": 125.679, + "pct_cuda_time": 2.1897836926723557, + "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": 125.679, + "pct_cuda_time": 2.1897836926723557, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cuda_time_us": 423.3379999999999, + "pct_cuda_time": 7.376082311989508, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cuda_time_us": 89.26100000000002, + "pct_cuda_time": 1.5552501387791688, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cuda_time_us": 204.47299999999998, + "pct_cuda_time": 3.5626607547147455, + "invocations": 32 + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cuda_time_us": 129.60399999999998, + "pct_cuda_time": 2.2581714184955954, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cuda_time_us": 445.7629999999999, + "pct_cuda_time": 7.76680708946369, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cuda_time_us": 445.7629999999999, + "pct_cuda_time": 7.76680708946369, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cuda_time_us": 3420.481, + "pct_cuda_time": 59.59717625773304, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cuda_time_us": 2051.6830000000004, + "pct_cuda_time": 35.74775400769498, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cuda_time_us": 2051.6830000000004, + "pct_cuda_time": 35.74775400769498, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cuda_time_us": 272.49499999999995, + "pct_cuda_time": 4.747850534574219, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cuda_time_us": 272.49499999999995, + "pct_cuda_time": 4.747850534574219, + "invocations": 32 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cuda_time_us": 1096.3029999999999, + "pct_cuda_time": 19.101571715463848, + "invocations": 32 + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cuda_time_us": 1096.3029999999999, + "pct_cuda_time": 19.101571715463848, + "invocations": 32 + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cuda_time_us": 4.326, + "pct_cuda_time": 0.07537459921307942, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cuda_time_us": 4.326, + "pct_cuda_time": 0.07537459921307942, + "invocations": 1 + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cuda_time_us": 261.408, + "pct_cuda_time": 4.554674810701033, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cuda_time_us": 3.123, + "pct_cuda_time": 0.05441397904356151, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cuda_time_us": 258.285, + "pct_cuda_time": 4.500260831657472, + "invocations": 1 + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cuda_time_us": 185.66100000000003, + "pct_cuda_time": 3.2348875322467734, + "invocations": 1 + }, + "children": [ + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cuda_time_us": 19.099999999999998, + "pct_cuda_time": 0.3327912262990793, + "invocations": 7 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 3.164, + "pct_cuda_time": 0.05512834764451765, + "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)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", + "cuda_time_us": 9.333, + "pct_cuda_time": 0.1626146866517962, + "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": 63.494, + "pct_cuda_time": 1.1062956085148556, + "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": 35.973, + "pct_cuda_time": 0.6267800410291507, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06627946726919884, + "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": 5.567, + "pct_cuda_time": 0.0969973171103128, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", + "cuda_time_us": 40.78, + "pct_cuda_time": 0.7105354035851547, + "invocations": 1 + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Host)", + "cuda_time_us": 4.446, + "pct_cuda_time": 0.07746543414270715, + "invocations": 1 + }, + "children": [] + } + ] + } + ], + "model_stats": [ + { + "entry": { + "name": "LlamaForCausalLM", + "cpu_time_us": 17010.776, + "cuda_time_us": 5292.264999999999, + "pct_cuda_time": 92.2104376570522, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "VocabParallelEmbedding(weight=bfloat16[128256, 4096])", + "cpu_time_us": 100.59, + "cuda_time_us": 3.844, + "pct_cuda_time": 0.0669764122457414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 3.844, + "pct_cuda_time": 0.0669764122457414, + "trace": "index_select(bfloat16[128256, 4096], 0, int64[1]) <- embedding(bfloat16[128256, 4096], int64[1], -1, False, False)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 952.202, + "cuda_time_us": 185.667, + "pct_cuda_time": 3.2349920739932543, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 119.438, + "cuda_time_us": 5.207, + "pct_cuda_time": 0.09072481232142963, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::rms_norm_kernel(c10::BFloat16*, c10::BFloat16 const*, c10::BFloat16 const*, float, int, unsigned long, unsigned long)", + "cpu_time_us": 0, + "cuda_time_us": 5.207, + "pct_cuda_time": 0.09072481232142963, + "trace": "_C::rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 587.837, + "cuda_time_us": 54.797, + "pct_cuda_time": 0.9547623469900863, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 145.587, + "cuda_time_us": 19.148, + "pct_cuda_time": 0.3336275602709304, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.148, + "pct_cuda_time": 0.3336275602709304, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 51.567, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.0788593240957923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.0788593240957923, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 248.961, + "cuda_time_us": 17.103, + "pct_cuda_time": 0.29799624834519134, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.15005225344961629, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 73.109, + "cuda_time_us": 14.02, + "pct_cuda_time": 0.24427921427817237, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.02, + "pct_cuda_time": 0.24427921427817237, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 40.33, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.09211870227451478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.09211870227451478, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 173.76, + "cuda_time_us": 120.376, + "pct_cuda_time": 2.097386212407224, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 59.489, + "cuda_time_us": 73.268, + "pct_cuda_time": 1.2765941135330336, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 73.268, + "pct_cuda_time": 1.2765941135330336, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 29.785, + "cuda_time_us": 9.253, + "pct_cuda_time": 0.16122079669871103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 9.253, + "pct_cuda_time": 0.16122079669871103, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 57.997, + "cuda_time_us": 37.855, + "pct_cuda_time": 0.6595713021754789, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 37.855, + "pct_cuda_time": 0.6595713021754789, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 558.274, + "cuda_time_us": 181.379, + "pct_cuda_time": 3.16027957250789, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.334, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.0872226638143032, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.006, + "pct_cuda_time": 0.0872226638143032, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 328.941, + "cuda_time_us": 52.191, + "pct_cuda_time": 0.9093563817683377, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.749, + "cuda_time_us": 19.468, + "pct_cuda_time": 0.339203120083271, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.468, + "pct_cuda_time": 0.339203120083271, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 25.168, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 150.595, + "cuda_time_us": 14.178, + "pct_cuda_time": 0.2470321469355155, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.163, + "pct_cuda_time": 0.05511092402010408, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.769, + "pct_cuda_time": 0.11794051365541718, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.246, + "pct_cuda_time": 0.0739807092599943, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 68.041, + "cuda_time_us": 13.979, + "pct_cuda_time": 0.2435648456772162, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.979, + "pct_cuda_time": 0.2435648456772162, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.285, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.08724008743871675, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.007, + "pct_cuda_time": 0.08724008743871675, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 146.809, + "cuda_time_us": 119.17499999999998, + "pct_cuda_time": 2.0764604394865325, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 49.975, + "cuda_time_us": 73.509, + "pct_cuda_time": 1.2807932070167025, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 73.509, + "pct_cuda_time": 1.2807932070167025, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 28.553, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.15144614340270143, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.692, + "pct_cuda_time": 0.15144614340270143, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 52.027, + "cuda_time_us": 36.974, + "pct_cuda_time": 0.6442210890671286, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 36.974, + "pct_cuda_time": 0.6442210890671286, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 524.864, + "cuda_time_us": 182.662, + "pct_cuda_time": 3.1826340826304937, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.558, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.08652571883776063, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.966, + "pct_cuda_time": 0.08652571883776063, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 308.4, + "cuda_time_us": 53.033, + "pct_cuda_time": 0.9240270735245588, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.335, + "cuda_time_us": 20.309, + "pct_cuda_time": 0.35385638821507864, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 20.309, + "pct_cuda_time": 0.35385638821507864, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 24.967, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.0788593240957923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.0788593240957923, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 142.643, + "cuda_time_us": 14.218, + "pct_cuda_time": 0.24772909191205808, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.404, + "pct_cuda_time": 0.059310017503773085, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.729, + "pct_cuda_time": 0.1172435686788746, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.07117550572941042, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 59.359, + "cuda_time_us": 13.98, + "pct_cuda_time": 0.2435822693016298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.98, + "pct_cuda_time": 0.2435822693016298, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.431, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.09002786734488706, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.167, + "pct_cuda_time": 0.09002786734488706, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 137.205, + "cuda_time_us": 119.496, + "pct_cuda_time": 2.082053422923287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.912, + "cuda_time_us": 73.669, + "pct_cuda_time": 1.283580986922873, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 73.669, + "pct_cuda_time": 1.283580986922873, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.927, + "cuda_time_us": 8.812, + "pct_cuda_time": 0.15353697833232913, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.812, + "pct_cuda_time": 0.15353697833232913, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.876, + "cuda_time_us": 37.015, + "pct_cuda_time": 0.6449354576680849, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 37.015, + "pct_cuda_time": 0.6449354576680849, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 484.804, + "cuda_time_us": 181.019, + "pct_cuda_time": 3.1540070677190073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.147, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.09351259222759993, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.367, + "pct_cuda_time": 0.09351259222759993, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 291.506, + "cuda_time_us": 51.351, + "pct_cuda_time": 0.8947205372609436, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 47.691, + "cuda_time_us": 18.787, + "pct_cuda_time": 0.32733763185763365, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 18.787, + "pct_cuda_time": 0.32733763185763365, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.704, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.07256939568249557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.165, + "pct_cuda_time": 0.07256939568249557, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 137.035, + "cuda_time_us": 14.578999999999999, + "pct_cuda_time": 0.2540190203253548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.284, + "pct_cuda_time": 0.057219182574145364, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 7.25, + "pct_cuda_time": 0.12632127699834164, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.894, + "cuda_time_us": 13.82, + "pct_cuda_time": 0.24079448939545947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.82, + "pct_cuda_time": 0.24079448939545947, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.328, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 126.298, + "cuda_time_us": 119.41499999999999, + "pct_cuda_time": 2.080642109345788, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 46.61, + "cuda_time_us": 73.429, + "pct_cuda_time": 1.2793993170636175, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 73.429, + "pct_cuda_time": 1.2793993170636175, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.947, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.15005225344961629, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.15005225344961629, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 47.0, + "cuda_time_us": 37.374, + "pct_cuda_time": 0.6511905388325545, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 37.374, + "pct_cuda_time": 0.6511905388325545, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 501.038, + "cuda_time_us": 161.912, + "pct_cuda_time": 2.821093876049033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.376, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07816237911924973, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.486, + "pct_cuda_time": 0.07816237911924973, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 301.811, + "cuda_time_us": 49.989000000000004, + "pct_cuda_time": 0.8709895608096689, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.655, + "cuda_time_us": 19.308, + "pct_cuda_time": 0.33641534017710073, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 19.308, + "pct_cuda_time": 0.33641534017710073, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.873, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 138.146, + "cuda_time_us": 14.498999999999999, + "pct_cuda_time": 0.25262513037226964, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.048838419231220905, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 7.49, + "pct_cuda_time": 0.13050294685759706, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.206, + "pct_cuda_time": 0.07328376428345171, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 61.572, + "cuda_time_us": 12.137, + "pct_cuda_time": 0.21147052950743067, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 12.137, + "pct_cuda_time": 0.21147052950743067, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.36, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.0844348839081329, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.846, + "pct_cuda_time": 0.0844348839081329, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 130.605, + "cuda_time_us": 102.59100000000001, + "pct_cuda_time": 1.7875070522119816, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 48.502, + "cuda_time_us": 63.214, + "pct_cuda_time": 1.1014169936790574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.214, + "pct_cuda_time": 1.1014169936790574, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.366, + "cuda_time_us": 8.572, + "pct_cuda_time": 0.14935530847307368, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.572, + "pct_cuda_time": 0.14935530847307368, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 49.454, + "cuda_time_us": 30.805, + "pct_cuda_time": 0.5367347500598502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 30.805, + "pct_cuda_time": 0.5367347500598502, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 515.599, + "cuda_time_us": 157.022, + "pct_cuda_time": 2.735892352666703, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.627, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 297.354, + "cuda_time_us": 46.022, + "pct_cuda_time": 0.8018700427610591, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 48.973, + "cuda_time_us": 16.544, + "pct_cuda_time": 0.28825644229800884, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.544, + "pct_cuda_time": 0.28825644229800884, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 23.586, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.06978161577632526, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.06978161577632526, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 141.591, + "cuda_time_us": 13.016000000000002, + "pct_cuda_time": 0.22678589536695376, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.883, + "pct_cuda_time": 0.05023230918430605, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.048, + "pct_cuda_time": 0.10537808045323727, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.07117550572941042, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.753, + "cuda_time_us": 12.457, + "pct_cuda_time": 0.21704608931977126, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 12.457, + "pct_cuda_time": 0.21704608931977126, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 30.135, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 141.712, + "cuda_time_us": 101.388, + "pct_cuda_time": 1.7665464320424638, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 53.409, + "cuda_time_us": 61.211, + "pct_cuda_time": 1.066517473978688, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.211, + "pct_cuda_time": 1.066517473978688, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 18.357, + "cuda_time_us": 8.451, + "pct_cuda_time": 0.14724704991903242, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.451, + "pct_cuda_time": 0.14724704991903242, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.805, + "cuda_time_us": 31.726, + "pct_cuda_time": 0.5527819081447429, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.726, + "pct_cuda_time": 0.5527819081447429, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 566.066, + "cuda_time_us": 165.565, + "pct_cuda_time": 2.884742376031783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.747, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 369.712, + "cuda_time_us": 50.348, + "pct_cuda_time": 0.8772446419741385, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 89.253, + "cuda_time_us": 18.336, + "pct_cuda_time": 0.3194795772471161, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 18.336, + "pct_cuda_time": 0.3194795772471161, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 24.236, + "cuda_time_us": 4.205, + "pct_cuda_time": 0.07326634065903814, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.205, + "pct_cuda_time": 0.07326634065903814, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 168.952, + "cuda_time_us": 13.216000000000001, + "pct_cuda_time": 0.23027062024966663, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.843, + "pct_cuda_time": 0.04953536420776348, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.408, + "pct_cuda_time": 0.11165058524212045, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 57.085, + "cuda_time_us": 14.591, + "pct_cuda_time": 0.25422810381831756, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.591, + "pct_cuda_time": 0.25422810381831756, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.049, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08304099395504776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08304099395504776, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 129.143, + "cuda_time_us": 105.76500000000001, + "pct_cuda_time": 1.8428096361006352, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.43, + "cuda_time_us": 63.264, + "pct_cuda_time": 1.102288174899736, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.264, + "pct_cuda_time": 1.102288174899736, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.847, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14306538005977698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14306538005977698, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 49.184, + "cuda_time_us": 34.29, + "pct_cuda_time": 0.597456081141122, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.29, + "pct_cuda_time": 0.597456081141122, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 487.197, + "cuda_time_us": 159.497, + "pct_cuda_time": 2.7790158230902753, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.835, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 293.828, + "cuda_time_us": 48.566, + "pct_cuda_time": 0.8461957432691668, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 47.891, + "cuda_time_us": 17.425, + "pct_cuda_time": 0.303606655406359, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.425, + "pct_cuda_time": 0.303606655406359, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 23.715, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 135.102, + "cuda_time_us": 12.735999999999999, + "pct_cuda_time": 0.2219072805311557, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.048838419231220905, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.048, + "pct_cuda_time": 0.10537808045323727, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 54.371, + "cuda_time_us": 14.52, + "pct_cuda_time": 0.25299102648495453, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.52, + "pct_cuda_time": 0.25299102648495453, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.038, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 126.649, + "cuda_time_us": 101.43900000000001, + "pct_cuda_time": 1.7674350368875553, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 48.122, + "cuda_time_us": 61.481, + "pct_cuda_time": 1.0712218525703505, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.481, + "pct_cuda_time": 1.0712218525703505, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.396, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.14097454513014926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.14097454513014926, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.689, + "cuda_time_us": 31.867, + "pct_cuda_time": 0.5552386391870555, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.867, + "pct_cuda_time": 0.5552386391870555, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 472.526, + "cuda_time_us": 164.274, + "pct_cuda_time": 2.862248476913872, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 22.944, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 281.971, + "cuda_time_us": 49.408, + "pct_cuda_time": 0.860866435025388, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 45.208, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.3063944353125293, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.585, + "pct_cuda_time": 0.3063944353125293, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.832, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 132.999, + "cuda_time_us": 13.136999999999999, + "pct_cuda_time": 0.22889415392099502, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.04814147425467833, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.329, + "pct_cuda_time": 0.11027411891344885, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.7, + "cuda_time_us": 14.801, + "pct_cuda_time": 0.2578870649451661, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.801, + "pct_cuda_time": 0.2578870649451661, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.738, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 125.337, + "cuda_time_us": 105.614, + "pct_cuda_time": 1.8401786688141866, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.761, + "cuda_time_us": 63.173, + "pct_cuda_time": 1.1007026250781016, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.173, + "pct_cuda_time": 1.1007026250781016, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.346, + "cuda_time_us": 8.452, + "pct_cuda_time": 0.147264473543446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.452, + "pct_cuda_time": 0.147264473543446, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.769, + "cuda_time_us": 33.989, + "pct_cuda_time": 0.592211570192639, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.989, + "pct_cuda_time": 0.592211570192639, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 502.76, + "cuda_time_us": 163.022, + "pct_cuda_time": 2.8404340991480894, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.285, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 295.06, + "cuda_time_us": 46.123000000000005, + "pct_cuda_time": 0.8036298288268292, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 52.628, + "cuda_time_us": 16.634, + "pct_cuda_time": 0.2898245684952296, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.634, + "pct_cuda_time": 0.2898245684952296, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.612, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.0641886323395711, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.684, + "pct_cuda_time": 0.0641886323395711, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 135.111, + "cuda_time_us": 12.736, + "pct_cuda_time": 0.22190728053115574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04605063932505061, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.10677197040632241, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 57.947, + "cuda_time_us": 13.069, + "pct_cuda_time": 0.22770934746087265, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.069, + "pct_cuda_time": 0.22770934746087265, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 37.256, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 129.774, + "cuda_time_us": 107.608, + "pct_cuda_time": 1.874921375894834, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 50.255, + "cuda_time_us": 63.774, + "pct_cuda_time": 1.1111742233506536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.774, + "pct_cuda_time": 1.1111742233506536, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.647, + "cuda_time_us": 9.533, + "pct_cuda_time": 0.16609941153450905, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 9.533, + "pct_cuda_time": 0.16609941153450905, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 47.28, + "cuda_time_us": 34.301, + "pct_cuda_time": 0.5976477410096712, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.301, + "pct_cuda_time": 0.5976477410096712, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 523.812, + "cuda_time_us": 163.43200000000002, + "pct_cuda_time": 2.847577785157651, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.336, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 321.43, + "cuda_time_us": 46.582, + "pct_cuda_time": 0.8116272724326551, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.409, + "cuda_time_us": 17.264, + "pct_cuda_time": 0.3008014518757751, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.264, + "pct_cuda_time": 0.3008014518757751, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.003, + "cuda_time_us": 3.844, + "pct_cuda_time": 0.0669764122457414, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.844, + "pct_cuda_time": 0.0669764122457414, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 167.27, + "cuda_time_us": 12.776, + "pct_cuda_time": 0.22260422550769826, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.04465674937196546, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.088, + "pct_cuda_time": 0.10607502542977984, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.125, + "pct_cuda_time": 0.071872450705953, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 58.697, + "cuda_time_us": 12.698, + "pct_cuda_time": 0.22124518280344027, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 12.698, + "pct_cuda_time": 0.22124518280344027, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.209, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 135.462, + "cuda_time_us": 107.47800000000001, + "pct_cuda_time": 1.8726563047210705, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.851, + "cuda_time_us": 63.774, + "pct_cuda_time": 1.1111742233506536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.774, + "pct_cuda_time": 1.1111742233506536, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.777, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 50.906, + "cuda_time_us": 35.573, + "pct_cuda_time": 0.6198105912637251, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 35.573, + "pct_cuda_time": 0.6198105912637251, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 526.496, + "cuda_time_us": 159.689, + "pct_cuda_time": 2.7823611589776793, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.436, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 331.625, + "cuda_time_us": 48.17699999999999, + "pct_cuda_time": 0.8394179533722901, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.539, + "cuda_time_us": 17.445, + "pct_cuda_time": 0.3039551278946303, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.445, + "pct_cuda_time": 0.3039551278946303, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.352, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 144.165, + "cuda_time_us": 12.616999999999999, + "pct_cuda_time": 0.21983386922594153, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.844, + "pct_cuda_time": 0.049552787832177043, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.728, + "pct_cuda_time": 0.09980252064089666, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 57.076, + "cuda_time_us": 14.27, + "pct_cuda_time": 0.24863512038156343, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.27, + "pct_cuda_time": 0.24863512038156343, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.159, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 127.089, + "cuda_time_us": 102.02000000000001, + "pct_cuda_time": 1.7775581626718364, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.371, + "cuda_time_us": 61.871, + "pct_cuda_time": 1.0780170660916406, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.871, + "pct_cuda_time": 1.0780170660916406, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.096, + "cuda_time_us": 8.132, + "pct_cuda_time": 0.14168891373110537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.132, + "pct_cuda_time": 0.14168891373110537, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 47.21, + "cuda_time_us": 32.017, + "pct_cuda_time": 0.5578521828490902, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.017, + "pct_cuda_time": 0.5578521828490902, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 476.361, + "cuda_time_us": 165.796, + "pct_cuda_time": 2.888767233271317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.755, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 284.374, + "cuda_time_us": 48.727000000000004, + "pct_cuda_time": 0.8490009467997507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.169, + "cuda_time_us": 17.234, + "pct_cuda_time": 0.3002787431433682, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.234, + "pct_cuda_time": 0.3002787431433682, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.702, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 135.322, + "cuda_time_us": 13.217, + "pct_cuda_time": 0.23028804387408017, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04605063932505061, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.529, + "pct_cuda_time": 0.11375884379616172, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 54.211, + "cuda_time_us": 14.311, + "pct_cuda_time": 0.2493494889825196, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.311, + "pct_cuda_time": 0.2493494889825196, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.248, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 122.643, + "cuda_time_us": 107.618, + "pct_cuda_time": 1.8750956121389692, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 45.659, + "cuda_time_us": 62.272, + "pct_cuda_time": 1.0850039394814799, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.272, + "pct_cuda_time": 1.0850039394814799, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.056, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14445927001286213, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.291, + "pct_cuda_time": 0.14445927001286213, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.378, + "cuda_time_us": 37.055, + "pct_cuda_time": 0.6456324026446274, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 37.055, + "pct_cuda_time": 0.6456324026446274, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 467.277, + "cuda_time_us": 166.17900000000003, + "pct_cuda_time": 2.895440481421713, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.045, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 280.098, + "cuda_time_us": 49.229, + "pct_cuda_time": 0.85774760625536, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.509, + "cuda_time_us": 17.635, + "pct_cuda_time": 0.30726561653320755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.635, + "pct_cuda_time": 0.30726561653320755, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.432, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 132.468, + "cuda_time_us": 12.897000000000002, + "pct_cuda_time": 0.2247124840617396, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.04744452927813575, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.089, + "pct_cuda_time": 0.1060924490541934, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.07117550572941042, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 52.839, + "cuda_time_us": 14.892, + "pct_cuda_time": 0.25947261476680045, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.892, + "pct_cuda_time": 0.25947261476680045, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.817, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 122.874, + "cuda_time_us": 107.81800000000001, + "pct_cuda_time": 1.8785803370216827, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 46.289, + "cuda_time_us": 62.322, + "pct_cuda_time": 1.0858751207021582, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.322, + "pct_cuda_time": 1.0858751207021582, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.025, + "cuda_time_us": 9.253, + "pct_cuda_time": 0.16122079669871103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 9.253, + "pct_cuda_time": 0.16122079669871103, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.628, + "cuda_time_us": 36.243, + "pct_cuda_time": 0.6314844196208133, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 36.243, + "pct_cuda_time": 0.6314844196208133, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 554.357, + "cuda_time_us": 161.50900000000001, + "pct_cuda_time": 2.8140721554103667, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 22.614, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.07675106554175103, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.405, + "pct_cuda_time": 0.07675106554175103, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 363.232, + "cuda_time_us": 47.735, + "pct_cuda_time": 0.8317167113814947, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 44.887, + "cuda_time_us": 16.724, + "pct_cuda_time": 0.2913926946924504, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.724, + "pct_cuda_time": 0.2913926946924504, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.381, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 208.181, + "cuda_time_us": 12.495, + "pct_cuda_time": 0.2177081870474867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.04465674937196546, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.887, + "pct_cuda_time": 0.10257287692265339, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 60.3, + "cuda_time_us": 14.711, + "pct_cuda_time": 0.25631893874794537, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.711, + "pct_cuda_time": 0.25631893874794537, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.071, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 125.097, + "cuda_time_us": 104.72300000000001, + "pct_cuda_time": 1.8246542194617008, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 46.259, + "cuda_time_us": 62.382, + "pct_cuda_time": 1.086920538166972, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.382, + "pct_cuda_time": 1.086920538166972, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.867, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.1674933014875942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 9.613, + "pct_cuda_time": 0.1674933014875942, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.449, + "cuda_time_us": 32.728, + "pct_cuda_time": 0.5702403798071344, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.728, + "pct_cuda_time": 0.5702403798071344, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 480.728, + "cuda_time_us": 159.458, + "pct_cuda_time": 2.778336301738146, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.555, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 288.711, + "cuda_time_us": 48.397, + "pct_cuda_time": 0.8432511507432744, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.179, + "cuda_time_us": 17.726, + "pct_cuda_time": 0.30885116635484183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.726, + "pct_cuda_time": 0.30885116635484183, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.654, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 135.082, + "cuda_time_us": 12.736, + "pct_cuda_time": 0.22190728053115574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04674758430159318, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.088, + "pct_cuda_time": 0.10607502542977984, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 56.945, + "cuda_time_us": 14.13, + "pct_cuda_time": 0.24619581296366444, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.13, + "pct_cuda_time": 0.24619581296366444, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.029, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08025321404887745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08025321404887745, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 125.868, + "cuda_time_us": 101.729, + "pct_cuda_time": 1.7724878879674888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 46.329, + "cuda_time_us": 62.142, + "pct_cuda_time": 1.0827388683077166, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.142, + "pct_cuda_time": 1.0827388683077166, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 18.528, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14236843508323438, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14236843508323438, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.199, + "cuda_time_us": 31.416, + "pct_cuda_time": 0.5473805845765379, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.416, + "pct_cuda_time": 0.5473805845765379, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 471.955, + "cuda_time_us": 168.769, + "pct_cuda_time": 2.940567668652844, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 22.935, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 284.725, + "cuda_time_us": 48.865, + "pct_cuda_time": 0.8514054069688226, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 45.197, + "cuda_time_us": 17.725, + "pct_cuda_time": 0.30883374273042835, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.725, + "pct_cuda_time": 0.30883374273042835, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 24.697, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06767335722228397, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.884, + "pct_cuda_time": 0.06767335722228397, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 134.511, + "cuda_time_us": 13.136, + "pct_cuda_time": 0.22887673029658145, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.803, + "pct_cuda_time": 0.048838419231220905, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.248, + "pct_cuda_time": 0.10886280533595014, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.07117550572941042, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.99, + "cuda_time_us": 14.12, + "pct_cuda_time": 0.24602157671952876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.12, + "pct_cuda_time": 0.24602157671952876, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.406, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 124.005, + "cuda_time_us": 110.452, + "pct_cuda_time": 1.9244741637270106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 46.209, + "cuda_time_us": 64.545, + "pct_cuda_time": 1.1246078377735118, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.545, + "pct_cuda_time": 1.1246078377735118, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.586, + "cuda_time_us": 8.452, + "pct_cuda_time": 0.147264473543446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.452, + "pct_cuda_time": 0.147264473543446, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.239, + "cuda_time_us": 37.455, + "pct_cuda_time": 0.6526018524100532, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 37.455, + "pct_cuda_time": 0.6526018524100532, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 492.114, + "cuda_time_us": 162.19099999999997, + "pct_cuda_time": 2.825955067260417, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.717, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08025321404887745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08025321404887745, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 281.1, + "cuda_time_us": 48.756, + "pct_cuda_time": 0.8495062319077441, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 45.678, + "cuda_time_us": 17.274, + "pct_cuda_time": 0.3009756881199108, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.274, + "pct_cuda_time": 0.3009756881199108, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.833, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 133.759, + "cuda_time_us": 12.936, + "pct_cuda_time": 0.22539200541386856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.04814147425467833, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.208, + "pct_cuda_time": 0.10816586035940756, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.881, + "cuda_time_us": 14.701, + "pct_cuda_time": 0.25614470250380966, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.701, + "pct_cuda_time": 0.25614470250380966, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 35.113, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08793703241525933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08793703241525933, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 135.362, + "cuda_time_us": 103.78199999999998, + "pct_cuda_time": 1.8082585888885363, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 54.892, + "cuda_time_us": 62.022, + "pct_cuda_time": 1.0806480333780888, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.022, + "pct_cuda_time": 1.0806480333780888, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.997, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.179, + "cuda_time_us": 33.629, + "pct_cuda_time": 0.585939065403756, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.629, + "pct_cuda_time": 0.585939065403756, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 469.631, + "cuda_time_us": 164.523, + "pct_cuda_time": 2.8665869593928495, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.174, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.0788593240957923, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.526, + "pct_cuda_time": 0.0788593240957923, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 280.92, + "cuda_time_us": 48.19500000000001, + "pct_cuda_time": 0.8397315786117345, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 44.837, + "cuda_time_us": 17.344, + "pct_cuda_time": 0.3021953418288603, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.344, + "pct_cuda_time": 0.3021953418288603, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.762, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 133.509, + "cuda_time_us": 12.696000000000002, + "pct_cuda_time": 0.22121033555461317, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.723, + "pct_cuda_time": 0.04744452927813575, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.048, + "pct_cuda_time": 0.10537808045323727, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06838772582324011, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.36, + "cuda_time_us": 14.31, + "pct_cuda_time": 0.249332065358106, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.31, + "pct_cuda_time": 0.249332065358106, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.747, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 122.614, + "cuda_time_us": 106.916, + "pct_cuda_time": 1.8628642278006473, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.261, + "cuda_time_us": 61.801, + "pct_cuda_time": 1.076797412382691, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.801, + "pct_cuda_time": 1.076797412382691, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 16.675, + "cuda_time_us": 8.372, + "pct_cuda_time": 0.14587058359036084, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.372, + "pct_cuda_time": 0.14587058359036084, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 44.476, + "cuda_time_us": 36.743, + "pct_cuda_time": 0.6401962318275954, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 36.743, + "pct_cuda_time": 0.6401962318275954, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 462.14, + "cuda_time_us": 162.62199999999999, + "pct_cuda_time": 2.8334646493826634, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.064, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 276.382, + "cuda_time_us": 48.626999999999995, + "pct_cuda_time": 0.8472585843583942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 45.498, + "cuda_time_us": 17.595, + "pct_cuda_time": 0.3065686715566649, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.595, + "pct_cuda_time": 0.3065686715566649, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 20.851, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.06490300094052724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.06490300094052724, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 130.995, + "cuda_time_us": 13.056, + "pct_cuda_time": 0.2274828403434963, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04605063932505061, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.368, + "pct_cuda_time": 0.11095364026557787, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 52.779, + "cuda_time_us": 14.251, + "pct_cuda_time": 0.2483040715177057, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.251, + "pct_cuda_time": 0.2483040715177057, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.357, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 122.192, + "cuda_time_us": 104.743, + "pct_cuda_time": 1.8250026919499718, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 45.488, + "cuda_time_us": 61.971, + "pct_cuda_time": 1.079759428532997, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.971, + "pct_cuda_time": 1.079759428532997, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.226, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.15005225344961629, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.612, + "pct_cuda_time": 0.15005225344961629, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.568, + "cuda_time_us": 34.16, + "pct_cuda_time": 0.5951910099673585, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.16, + "pct_cuda_time": 0.5951910099673585, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 475.079, + "cuda_time_us": 167.35999999999999, + "pct_cuda_time": 2.9160177818541313, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.085, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 283.984, + "cuda_time_us": 47.79599999999999, + "pct_cuda_time": 0.8327795524707221, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.85, + "cuda_time_us": 16.945, + "pct_cuda_time": 0.2952433156878481, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.945, + "pct_cuda_time": 0.2952433156878481, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.312, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06627946726919884, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06627946726919884, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 132.928, + "cuda_time_us": 12.857, + "pct_cuda_time": 0.224015539085197, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.04465674937196546, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.369, + "pct_cuda_time": 0.11097106388999142, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06838772582324011, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 56.615, + "cuda_time_us": 14.19, + "pct_cuda_time": 0.2472412304284783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.19, + "pct_cuda_time": 0.2472412304284783, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.428, + "cuda_time_us": 5.046, + "pct_cuda_time": 0.08791960879084577, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.046, + "pct_cuda_time": 0.08791960879084577, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 126.569, + "cuda_time_us": 109.952, + "pct_cuda_time": 1.9157623515202287, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.611, + "cuda_time_us": 64.185, + "pct_cuda_time": 1.1183353329846286, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 64.185, + "pct_cuda_time": 1.1183353329846286, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.576, + "cuda_time_us": 8.492, + "pct_cuda_time": 0.1479614185199886, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.492, + "pct_cuda_time": 0.1479614185199886, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.608, + "cuda_time_us": 37.275, + "pct_cuda_time": 0.6494656000156115, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 37.275, + "pct_cuda_time": 0.6494656000156115, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 481.148, + "cuda_time_us": 157.985, + "pct_cuda_time": 2.752671302976966, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.515, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 286.778, + "cuda_time_us": 47.095000000000006, + "pct_cuda_time": 0.8205655917568137, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 47.871, + "cuda_time_us": 16.754, + "pct_cuda_time": 0.29191540342485733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.754, + "pct_cuda_time": 0.29191540342485733, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 23.095, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 133.069, + "cuda_time_us": 12.576, + "pct_cuda_time": 0.21911950062498542, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.883, + "pct_cuda_time": 0.05023230918430605, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.688, + "pct_cuda_time": 0.09910557566435409, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.06978161577632526, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.343, + "cuda_time_us": 13.88, + "pct_cuda_time": 0.24183990686027337, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.88, + "pct_cuda_time": 0.24183990686027337, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 27.331, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.09211870227451478, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.287, + "pct_cuda_time": 0.09211870227451478, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 127.02, + "cuda_time_us": 100.958, + "pct_cuda_time": 1.7590542735446308, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 48.262, + "cuda_time_us": 61.301, + "pct_cuda_time": 1.068085600175909, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.301, + "pct_cuda_time": 1.068085600175909, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.937, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.969, + "cuda_time_us": 31.526, + "pct_cuda_time": 0.5492971832620301, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 31.526, + "pct_cuda_time": 0.5492971832620301, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 602.8, + "cuda_time_us": 162.351, + "pct_cuda_time": 2.8287428471665876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.035, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.07953884544792132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.565, + "pct_cuda_time": 0.07953884544792132, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 308.791, + "cuda_time_us": 47.63600000000001, + "pct_cuda_time": 0.8299917725645521, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 45.849, + "cuda_time_us": 16.734, + "pct_cuda_time": 0.2915669309365861, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.734, + "pct_cuda_time": 0.2915669309365861, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.614, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.805, + "pct_cuda_time": 0.06629689089361239, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 158.687, + "cuda_time_us": 12.736, + "pct_cuda_time": 0.22190728053115574, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.603, + "pct_cuda_time": 0.045353694348508035, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.048, + "pct_cuda_time": 0.10537808045323727, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.07117550572941042, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.162, + "cuda_time_us": 14.361, + "pct_cuda_time": 0.2502206702031978, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.361, + "pct_cuda_time": 0.2502206702031978, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 109.714, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08793703241525933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08793703241525933, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 143.344, + "cuda_time_us": 105.10300000000001, + "pct_cuda_time": 1.8312751967388554, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.913, + "cuda_time_us": 63.804, + "pct_cuda_time": 1.1116969320830605, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.804, + "pct_cuda_time": 1.1116969320830605, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 18.618, + "cuda_time_us": 8.051, + "pct_cuda_time": 0.14027760015360669, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.051, + "pct_cuda_time": 0.14027760015360669, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.88, + "cuda_time_us": 33.248, + "pct_cuda_time": 0.5793006645021879, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.248, + "pct_cuda_time": 0.5793006645021879, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 492.195, + "cuda_time_us": 161.511, + "pct_cuda_time": 2.8141070026591932, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 29.384, + "cuda_time_us": 4.767, + "pct_cuda_time": 0.08305841757946132, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.767, + "pct_cuda_time": 0.08305841757946132, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 297.023, + "cuda_time_us": 47.49499999999999, + "pct_cuda_time": 0.8275350415222392, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 51.817, + "cuda_time_us": 17.826, + "pct_cuda_time": 0.3105935287961983, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.826, + "pct_cuda_time": 0.3105935287961983, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.924, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 138.978, + "cuda_time_us": 12.815999999999999, + "pct_cuda_time": 0.22330117048424084, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04605063932505061, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.168, + "pct_cuda_time": 0.10746891538286499, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.06978161577632526, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.322, + "cuda_time_us": 13.008, + "pct_cuda_time": 0.22664650637164518, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.008, + "pct_cuda_time": 0.22664650637164518, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.238, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 123.013, + "cuda_time_us": 104.68299999999999, + "pct_cuda_time": 1.8239572744851578, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 45.788, + "cuda_time_us": 61.481, + "pct_cuda_time": 1.0712218525703505, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 61.481, + "pct_cuda_time": 1.0712218525703505, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.366, + "cuda_time_us": 8.772, + "pct_cuda_time": 0.15284003335578658, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.772, + "pct_cuda_time": 0.15284003335578658, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 45.387, + "cuda_time_us": 34.43, + "pct_cuda_time": 0.599895388559021, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.43, + "pct_cuda_time": 0.599895388559021, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 467.427, + "cuda_time_us": 167.881, + "pct_cuda_time": 2.925095490173599, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.655, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 279.127, + "cuda_time_us": 48.678, + "pct_cuda_time": 0.848147189203486, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 44.917, + "cuda_time_us": 17.896, + "pct_cuda_time": 0.31181318250514783, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.896, + "pct_cuda_time": 0.31181318250514783, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.442, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06559994591706982, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06559994591706982, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 132.578, + "cuda_time_us": 12.897, + "pct_cuda_time": 0.22471248406173958, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.04465674937196546, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.369, + "pct_cuda_time": 0.11097106388999142, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.28, + "cuda_time_us": 14.12, + "pct_cuda_time": 0.24602157671952876, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.12, + "pct_cuda_time": 0.24602157671952876, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.287, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 123.975, + "cuda_time_us": 109.871, + "pct_cuda_time": 1.9143510379427298, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 47.591, + "cuda_time_us": 65.326, + "pct_cuda_time": 1.1382156884405052, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 65.326, + "pct_cuda_time": 1.1382156884405052, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.566, + "cuda_time_us": 8.452, + "pct_cuda_time": 0.147264473543446, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.452, + "pct_cuda_time": 0.147264473543446, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 43.966, + "cuda_time_us": 36.093, + "pct_cuda_time": 0.6288708759587786, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 36.093, + "pct_cuda_time": 0.6288708759587786, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 509.521, + "cuda_time_us": 162.221, + "pct_cuda_time": 2.8264777759928243, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.084, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08304099395504776, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.766, + "pct_cuda_time": 0.08304099395504776, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 300.218, + "cuda_time_us": 48.997, + "pct_cuda_time": 0.8537053253914131, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 44.326, + "cuda_time_us": 17.715, + "pct_cuda_time": 0.3086595064862927, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.715, + "pct_cuda_time": 0.3086595064862927, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 28.953, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06559994591706982, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.765, + "pct_cuda_time": 0.06559994591706982, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 133.48, + "cuda_time_us": 13.657, + "pct_cuda_time": 0.2379544386160485, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04605063932505061, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.969, + "pct_cuda_time": 0.12142523853813006, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.045, + "pct_cuda_time": 0.07047856075286783, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 56.494, + "cuda_time_us": 13.86, + "pct_cuda_time": 0.24149143437200205, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 13.86, + "pct_cuda_time": 0.24149143437200205, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 28.312, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.726, + "pct_cuda_time": 0.08234404897850517, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 141.731, + "cuda_time_us": 103.732, + "pct_cuda_time": 1.8073874076678584, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 61.702, + "cuda_time_us": 62.693, + "pct_cuda_time": 1.0923392853595904, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.693, + "pct_cuda_time": 1.0923392853595904, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.706, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14236843508323438, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14236843508323438, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.81, + "cuda_time_us": 32.868, + "pct_cuda_time": 0.5726796872250335, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.868, + "pct_cuda_time": 0.5726796872250335, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 504.664, + "cuda_time_us": 165.92700000000002, + "pct_cuda_time": 2.8910497280694942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.677, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 298.125, + "cuda_time_us": 48.908, + "pct_cuda_time": 0.8521546228186059, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.279, + "cuda_time_us": 17.265, + "pct_cuda_time": 0.3008188755001887, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.265, + "pct_cuda_time": 0.3008188755001887, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 23.184, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 135.693, + "cuda_time_us": 13.017, + "pct_cuda_time": 0.2268033189913673, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.763, + "pct_cuda_time": 0.04814147425467833, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.249, + "pct_cuda_time": 0.10888022896036369, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.06978161577632526, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 65.328, + "cuda_time_us": 14.661, + "pct_cuda_time": 0.2554477575272671, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.661, + "pct_cuda_time": 0.2554477575272671, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.98, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.806, + "pct_cuda_time": 0.08373793893159033, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 132.878, + "cuda_time_us": 107.56700000000001, + "pct_cuda_time": 1.874207007293878, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 54.261, + "cuda_time_us": 62.903, + "pct_cuda_time": 1.095998246486439, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.903, + "pct_cuda_time": 1.095998246486439, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.546, + "cuda_time_us": 9.693, + "pct_cuda_time": 0.16888719144067935, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 9.693, + "pct_cuda_time": 0.16888719144067935, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.119, + "cuda_time_us": 34.971, + "pct_cuda_time": 0.6093215693667592, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.971, + "pct_cuda_time": 0.6093215693667592, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 491.043, + "cuda_time_us": 161.49, + "pct_cuda_time": 2.8137411065465088, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.276, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 291.174, + "cuda_time_us": 47.595, + "pct_cuda_time": 0.8292774039635957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 44.947, + "cuda_time_us": 16.864, + "pct_cuda_time": 0.29383200211034943, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 16.864, + "pct_cuda_time": 0.29383200211034943, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.883, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.06490300094052724, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.725, + "pct_cuda_time": 0.06490300094052724, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 141.191, + "cuda_time_us": 12.776, + "pct_cuda_time": 0.22260422550769826, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.683, + "pct_cuda_time": 0.04674758430159318, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.128, + "pct_cuda_time": 0.10677197040632241, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.965, + "pct_cuda_time": 0.06908467079978269, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.202, + "cuda_time_us": 14.23, + "pct_cuda_time": 0.24793817540502086, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.23, + "pct_cuda_time": 0.24793817540502086, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.768, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.686, + "pct_cuda_time": 0.0816471040019626, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 134.52, + "cuda_time_us": 104.523, + "pct_cuda_time": 1.8211694945789878, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 54.161, + "cuda_time_us": 62.162, + "pct_cuda_time": 1.0830873407959878, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.162, + "pct_cuda_time": 1.0830873407959878, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.747, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.14097454513014926, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.091, + "pct_cuda_time": 0.14097454513014926, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.91, + "cuda_time_us": 34.27, + "pct_cuda_time": 0.5971076086528507, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 34.27, + "pct_cuda_time": 0.5971076086528507, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 476.221, + "cuda_time_us": 165.978, + "pct_cuda_time": 2.891938332914586, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.705, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08025321404887745, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.606, + "pct_cuda_time": 0.08025321404887745, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 280.058, + "cuda_time_us": 48.577000000000005, + "pct_cuda_time": 0.8463874031377161, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 45.288, + "cuda_time_us": 17.315, + "pct_cuda_time": 0.30169005672086696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.315, + "pct_cuda_time": 0.30169005672086696, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.103, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.885, + "pct_cuda_time": 0.06769078084669755, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 132.468, + "cuda_time_us": 13.017000000000001, + "pct_cuda_time": 0.22680331899136733, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.483, + "pct_cuda_time": 0.04326285941888031, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.448, + "pct_cuda_time": 0.11234753021866302, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.086, + "pct_cuda_time": 0.07119292935382399, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 52.649, + "cuda_time_us": 14.36, + "pct_cuda_time": 0.25020324657878423, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.36, + "pct_cuda_time": 0.25020324657878423, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.887, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 132.598, + "cuda_time_us": 108.149, + "pct_cuda_time": 1.8843475567025723, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 54.832, + "cuda_time_us": 63.624, + "pct_cuda_time": 1.108560679688619, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.624, + "pct_cuda_time": 1.108560679688619, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.136, + "cuda_time_us": 8.292, + "pct_cuda_time": 0.14447669363727567, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.292, + "pct_cuda_time": 0.14447669363727567, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 44.897, + "cuda_time_us": 36.233, + "pct_cuda_time": 0.6313101833766775, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 36.233, + "pct_cuda_time": 0.6313101833766775, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 501.038, + "cuda_time_us": 163.602, + "pct_cuda_time": 2.850539801307957, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 23.655, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 281.08, + "cuda_time_us": 49.909, + "pct_cuda_time": 0.8695956708565837, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 44.467, + "cuda_time_us": 18.436, + "pct_cuda_time": 0.32122193968847257, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 18.436, + "pct_cuda_time": 0.32122193968847257, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.002, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06420605596398468, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.685, + "pct_cuda_time": 0.06420605596398468, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 134.501, + "cuda_time_us": 13.578, + "pct_cuda_time": 0.23657797228737687, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.643, + "pct_cuda_time": 0.04605063932505061, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 7.01, + "pct_cuda_time": 0.12213960713908618, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 3.925, + "pct_cuda_time": 0.06838772582324011, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 53.319, + "cuda_time_us": 14.21, + "pct_cuda_time": 0.2475897029167496, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.21, + "pct_cuda_time": 0.2475897029167496, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.369, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.645, + "pct_cuda_time": 0.08093273540100646, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 147.89, + "cuda_time_us": 104.482, + "pct_cuda_time": 1.8204551259780315, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 68.192, + "cuda_time_us": 62.382, + "pct_cuda_time": 1.086920538166972, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 62.382, + "pct_cuda_time": 1.086920538166972, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 18.358, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.131, + "pct_cuda_time": 0.14167149010669183, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.529, + "cuda_time_us": 33.969, + "pct_cuda_time": 0.5918630977043678, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 33.969, + "pct_cuda_time": 0.5918630977043678, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 504.343, + "cuda_time_us": 161.89999999999998, + "pct_cuda_time": 2.8208847925560696, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 31.226, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.07884190047137875, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.525, + "pct_cuda_time": 0.07884190047137875, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 297.524, + "cuda_time_us": 48.675999999999995, + "pct_cuda_time": 0.8481123419546589, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 50.546, + "cuda_time_us": 17.435, + "pct_cuda_time": 0.3037808916504946, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 17.435, + "pct_cuda_time": 0.3037808916504946, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 22.143, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 142.073, + "cuda_time_us": 12.936, + "pct_cuda_time": 0.22539200541386856, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.563, + "pct_cuda_time": 0.04465674937196546, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 6.288, + "pct_cuda_time": 0.10955975031249271, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.085, + "pct_cuda_time": 0.07117550572941042, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 55.543, + "cuda_time_us": 14.46, + "pct_cuda_time": 0.2519456090201407, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 14.46, + "pct_cuda_time": 0.2519456090201407, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 25.598, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08793703241525933, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 5.047, + "pct_cuda_time": 0.08793703241525933, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 134.181, + "cuda_time_us": 103.652, + "pct_cuda_time": 1.8059935177147732, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 55.513, + "cuda_time_us": 63.074, + "pct_cuda_time": 1.0989776862611587, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.074, + "pct_cuda_time": 1.0989776862611587, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 17.396, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14236843508323438, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.171, + "pct_cuda_time": 0.14236843508323438, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 46.009, + "cuda_time_us": 32.407, + "pct_cuda_time": 0.5646473963703802, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 32.407, + "pct_cuda_time": 0.5646473963703802, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "LlamaDecoderLayer", + "cpu_time_us": 573.466, + "cuda_time_us": 149.702, + "pct_cuda_time": 2.6083514219594117, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.437, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.566, + "pct_cuda_time": 0.07955626907233487, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaAttention", + "cpu_time_us": 368.169, + "cuda_time_us": 42.827, + "pct_cuda_time": 0.7462015627597209, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "QKVParallelLinear(weight=bfloat16[6144, 4096])", + "cpu_time_us": 46.018, + "cuda_time_us": 15.832, + "pct_cuda_time": 0.275850821715551, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 15.832, + "pct_cuda_time": 0.275850821715551, + "trace": "_rocm_C::wvSplitK(bfloat16[6144, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Llama3RotaryEmbedding", + "cpu_time_us": 21.602, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::rotary_embedding_kernel(long const*, c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, int, long, long, int, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 3.845, + "pct_cuda_time": 0.06699383587015498, + "trace": "_C::rotary_embedding(int64[1], bfloat16[1, 4096], bfloat16[1, 1024], 128, bfloat16[131072, 128], True)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Attention", + "cpu_time_us": 214.731, + "cuda_time_us": 12.495000000000001, + "pct_cuda_time": 0.2177081870474867, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void vllm::reshape_and_cache_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, __hip_bfloat16*, long const*, int, int, int, int, int, int, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 2.843, + "pct_cuda_time": 0.04953536420776348, + "trace": "_C_cache_ops::reshape_and_cache(bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], int64[1], None, float32[], float32[]) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_QKV_mfma4_kernel<__hip_bfloat16, __hip_bfloat16, (vllm::Fp8KVCacheDataType)0, __hip_bfloat16, 16, 128, 256, false, 4>(__hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16 const*, int, float, int const*, int const*, int const*, int, float const*, int, int, int, float*, float*, __hip_bfloat16*, __hip_bfloat16*, int, float const*, float const*, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 5.647, + "pct_cuda_time": 0.09839120706339796, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void paged_attention_ll4mi_reduce_kernel<__hip_bfloat16, __hip_bfloat16, 128, 128, 256, 1>(__hip_bfloat16*, float const*, float const*, __hip_bfloat16 const*, int const*, int const*, int, float const*)", + "cpu_time_us": 0, + "cuda_time_us": 4.005, + "pct_cuda_time": 0.06978161577632526, + "trace": "_rocm_C::paged_attention(bfloat16[1, 32, 128], float32[1, 32, 2], float32[1, 32, 2], bfloat16[1, 32, 2, 128], bfloat16[1, 32, 128], bfloat16[79739, 8, 16, 16, 8], bfloat16[79739, 8, 128, 16], 8, 0.08838834764831845, int32[1, 17], int32[1], None, 16, 257, None, None, float32[], float32[], None) <- vllm::unified_attention_with_output(bfloat16[1, 32, 128], bfloat16[1, 8, 128], bfloat16[1, 8, 128], bfloat16[1, 32, 128], None, None)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 4096])", + "cpu_time_us": 58.908, + "cuda_time_us": 10.655, + "pct_cuda_time": 0.18564871812652828, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 10.655, + "pct_cuda_time": 0.18564871812652828, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 26.941, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.646, + "pct_cuda_time": 0.08095015902542002, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "LlamaMLP", + "cpu_time_us": 134.261, + "cuda_time_us": 97.66300000000001, + "pct_cuda_time": 1.7016434311019364, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "MergedColumnParallelLinear(weight=bfloat16[28672, 4096])", + "cpu_time_us": 54.301, + "cuda_time_us": 63.654, + "pct_cuda_time": 1.109083388421026, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 63.654, + "pct_cuda_time": 1.109083388421026, + "trace": "_rocm_C::wvSplitK(bfloat16[28672, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "SiluAndMul", + "cpu_time_us": 18.558, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14306538005977698, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "_ZN4vllm18act_and_mul_kernelIN3c108BFloat16ETnPFT_RKS3_EXadL_ZNS_11silu_kernelIS2_EES3_S5_EELb1EEEvPS3_PS4_i", + "cpu_time_us": 0, + "cuda_time_us": 8.211, + "pct_cuda_time": 0.14306538005977698, + "trace": "_C::silu_and_mul(bfloat16[1, 14336], bfloat16[1, 28672])" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "RowParallelLinear(weight=bfloat16[4096, 14336])", + "cpu_time_us": 47.101, + "cuda_time_us": 25.798, + "pct_cuda_time": 0.4494946626211334, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 25.798, + "pct_cuda_time": 0.4494946626211334, + "trace": "_rocm_C::wvSplitK(bfloat16[4096, 14336], bfloat16[1, 14336], 304)" + }, + "children": [] + } + ] + } + ] + } + ] + }, + { + "entry": { + "name": "RMSNorm(weight=bfloat16[4096])", + "cpu_time_us": 24.777, + "cuda_time_us": 4.326, + "pct_cuda_time": 0.07537459921307942, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "std::enable_if<(((8)>(0)))&&_typeConvert::exists, void>::type vllm::fused_add_rms_norm_kernel(c10::BFloat16*, c10::BFloat16*, c10::BFloat16 const*, float, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 4.326, + "pct_cuda_time": 0.07537459921307942, + "trace": "_C::fused_add_rms_norm(bfloat16[1, 4096], bfloat16[1, 4096], bfloat16[4096], 1e-05)" + }, + "children": [] + } + ] + } + ] + }, + { + "entry": { + "name": "LogitsProcessor", + "cpu_time_us": 120.139, + "cuda_time_us": 261.408, + "pct_cuda_time": 4.554674810701033, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex(at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, at::cuda::detail::TensorInfo, int, int, unsigned int, long)", + "cpu_time_us": 0, + "cuda_time_us": 3.123, + "pct_cuda_time": 0.05441397904356151, + "trace": "index_select(bfloat16[1, 4096], 0, int64[1])" + }, + "children": [] + }, + { + "entry": { + "name": "void wvSplitK_hf_sml_<__hip_bfloat16, 64, 2, 16, 8, 2, 1>(int, int, __hip_bfloat16 const*, __hip_bfloat16 const*, __hip_bfloat16*, int, int)", + "cpu_time_us": 0, + "cuda_time_us": 258.285, + "pct_cuda_time": 4.500260831657472, + "trace": "_rocm_C::wvSplitK(bfloat16[128256, 4096], bfloat16[1, 4096], 304)" + }, + "children": [] + } + ] + }, + { + "entry": { + "name": "Sampler", + "cpu_time_us": 875.007, + "cuda_time_us": 185.66100000000003, + "pct_cuda_time": 3.2348875322467734, + "trace": "" + }, + "children": [ + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 4.886, + "pct_cuda_time": 0.08513182888467548, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.442, + "pct_cuda_time": 0.042548490817924174, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.483, + "pct_cuda_time": 0.04326285941888031, + "trace": "copy_(int32[1], int32[1], True) <- _to_copy(int32[1], 3, 0, None, None, True, None) <- to(int32[1], 3, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.362, + "pct_cuda_time": 0.041154600864839026, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.323, + "pct_cuda_time": 0.04047507951271001, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.282, + "pct_cuda_time": 0.03976071091175388, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoD (Device -> Device)", + "cpu_time_us": 0, + "cuda_time_us": 2.322, + "pct_cuda_time": 0.04045765588829645, + "trace": "copy_(bfloat16[1], bfloat16[1], True) <- _to_copy(bfloat16[1], 15, 0, None, None, True, None) <- to(bfloat16[1], 15, 0, None, None, True, False, None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 3.164, + "pct_cuda_time": 0.05512834764451765, + "trace": "copy_(float32[1, 128256], bfloat16[1, 128256], False) <- _to_copy(bfloat16[1, 128256], 6, None, None, None, False, None) <- to(bfloat16[1, 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)#2}>(int, at::native::gpu_kernel_impl > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#2})", + "cpu_time_us": 0, + "cuda_time_us": 9.333, + "pct_cuda_time": 0.1626146866517962, + "trace": "div_(float32[1, 128256], bfloat16[1, 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": 63.494, + "pct_cuda_time": 1.1062956085148556, + "trace": "_softmax(float32[1, 128256], -1, False) <- softmax(float32[1, 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": 35.973, + "pct_cuda_time": 0.6267800410291507, + "trace": "_log_softmax(float32[1, 128256], -1, False) <- log_softmax(float32[1, 128256], -1, 6)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1} const&)::{lambda(int)#1})", + "cpu_time_us": 0, + "cuda_time_us": 3.804, + "pct_cuda_time": 0.06627946726919884, + "trace": "copy_(int64[1], int32[1], False) <- _to_copy(int32[1], 4, None, None, None, False, None) <- to(int32[1], 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": 5.567, + "pct_cuda_time": 0.0969973171103128, + "trace": "index(float32[1, 128256], None)" + }, + "children": [] + }, + { + "entry": { + "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp, unsigned int, long, 4, 4> >(at::native::ReduceOp, unsigned int, long, 4, 4>)", + "cpu_time_us": 0, + "cuda_time_us": 40.78, + "pct_cuda_time": 0.7105354035851547, + "trace": "argmax(float32[1, 128256], -1, False)" + }, + "children": [] + }, + { + "entry": { + "name": "Memcpy DtoH (Device -> Host)", + "cpu_time_us": 0, + "cuda_time_us": 4.446, + "pct_cuda_time": 0.07746543414270715, + "trace": "copy_(int64[1], int64[1], False) <- _to_copy(int64[1], 4, 0, None, None, False, None) <- to(int64[1], 4, 0, None, None, False, False, None)" + }, + "children": [] + } + ] + } + ] + } +} \ No newline at end of file