| 1 | { |
| 2 | "vllm": { |
| 3 | "llama": { |
| 4 | "fused_moe_kernel|GroupProblemShape|group_gemm_starts|bmm_|GemmUniversal": "moe_gemm", |
| 5 | "gemm|nvjet": "gemm", |
| 6 | "moe|sigmoid": "moe", |
| 7 | "CatArrayBatched|prepare_inputs": "prepare_next", |
| 8 | "ncclDevKernel|cross_device_reduce": "nccl_and_custom_ar", |
| 9 | "_norm_|Norm": "norm", |
| 10 | "act_and_mul_": "activation", |
| 11 | "Rotary": "rope", |
| 12 | "SoftMax": "softmax", |
| 13 | "flash|fmha": "attn", |
| 14 | "elementwise": "elementwise", |
| 15 | "fp8_quant|cvt_": "quantize", |
| 16 | "reduce_kernel": "reduce", |
| 17 | "triton": "triton_kernel", |
| 18 | "CUDA mem": "non-gpu-H_D_memops", |
| 19 | ".*": "misc" |
| 20 | }, |
| 21 | "ds": { |
| 22 | "block_fp8|gemm_fp8_blockwise": "block_fp8_gemm", |
| 23 | "fused_moe_kernel|_group_gemm|GroupProblemShape|GemmUniversal|bmm_": "moe_gemm", |
| 24 | "gemm|matmul|nvjet": "gemm", |
| 25 | "moe|sigmoid|expert": "moe", |
| 26 | "CatArrayBatched": "prepare_next", |
| 27 | "ncclDevKernel|cross_device_reduce": "nccl_and_custom_ar", |
| 28 | "Norm|_norm_": "norm", |
| 29 | "sbtopk": "topk", |
| 30 | "act_and_mul_": "activation", |
| 31 | "compute_position_kernel": "rope", |
| 32 | "elementwise": "elementwise", |
| 33 | "fp8_quant|quant_fp8|cvt_": "quantize", |
| 34 | "reduce": "reduce", |
| 35 | "SoftMax": "softmax", |
| 36 | "_fwd_|FlashAttn|_mla_|_attn_|fmha": "attn", |
| 37 | "triton": "triton_kernel", |
| 38 | "topk": "topk", |
| 39 | "CUDA mem": "non-gpu-H_D_memops", |
| 40 | ".*": "misc" |
| 41 | }, |
| 42 | "gpt-oss": { |
| 43 | "block_fp8|gemm_fp8_blockwise": "block_fp8_gemm", |
| 44 | "fused_moe_kernel|_group_gemm|GroupProblemShape|GemmUniversal|bmm_|matmul_ogs_|_topk_forward|_combined_routing|_sum_bitmatrix_rows|_compute_writeback_idx": "moe_gemm", |
| 45 | "gemm|matmul|nvjet": "gemm", |
| 46 | "moe|sigmoid|expert|splitKreduce": "moe", |
| 47 | "CatArrayBatched": "prepare_next", |
| 48 | "ncclDevKernel|cross_device_reduce": "nccl_and_custom_ar", |
| 49 | "Norm|_norm_": "norm", |
| 50 | "topk": "topk", |
| 51 | "act_and_mul_": "activation", |
| 52 | "compute_position_kernel": "rope", |
| 53 | "elementwise": "elementwise", |
| 54 | "fp8_quant|quant_fp8|cvt_|quantize": "quantize", |
| 55 | "reduce": "reduce", |
| 56 | "SoftMax": "softmax", |
| 57 | "_fwd_|FlashAttn|_mla_|_attn_|_flash_|flash::prepare_varlen|fmha": "attn", |
| 58 | "triton": "triton_kernel", |
| 59 | "CUDA mem": "non-gpu-H_D_memops", |
| 60 | ".*": "misc" |
| 61 | } |
| 62 | } |
| 63 | } |