From 3adc70bb5e59db6f4760aa22e8384938a2bca2ee Mon Sep 17 00:00:00 2001 From: TobyMint <130973409+TobyMint@users.noreply.github.com> Date: Mon, 17 Aug 2026 10:41:03 +0800 Subject: [PATCH] [MoE] Add H20 fp8_w8a8 tuned configs for Qwen3.8 (triton 3.7.1) + fix Qwen3_5MoeForCausalLM tuning (#34795) --- .../kernels/fused_moe_triton/common_utils.py | 1 + ...dtype=fp8_w8a8,block_shape=[128, 128].json | 146 ++++++++++++++++++ 2 files changed, 147 insertions(+) create mode 100644 python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=512,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json diff --git a/benchmark/kernels/fused_moe_triton/common_utils.py b/benchmark/kernels/fused_moe_triton/common_utils.py index 000ef3bdd..d0b9d623a 100644 --- a/benchmark/kernels/fused_moe_triton/common_utils.py +++ b/benchmark/kernels/fused_moe_triton/common_utils.py @@ -82,6 +82,7 @@ def get_model_config( "Qwen3MoeForCausalLM", "Qwen3NextForCausalLM", "Qwen3VLMoeForConditionalGeneration", + "Qwen3_5MoeForCausalLM", "Qwen3_5MoeForConditionalGeneration", "InternS2PreviewForConditionalGeneration", "MellumForCausalLM", diff --git a/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=512,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json b/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=512,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 000000000..bc2771140 --- /dev/null +++ b/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=512,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +}