From ee462b5899c02db4e9d250f43c4c54d81253c4c6 Mon Sep 17 00:00:00 2001 From: Xiaoyu Zhang <1182563586@qq.com> Date: Tue, 1 Sep 2026 22:55:38 +0800 Subject: [PATCH] [Kernel] Add tuned LFM2.5 Triton MoE configs on B300 (#37158) Co-authored-by: Song Bian Co-authored-by: Cursor --- ...=1792,device_name=NVIDIA_B300_SXM6_AC.json | 74 +++++++++++++++++++ ...,device_name=NVIDIA_B300_SXM6_AC_down.json | 74 +++++++++++++++++++ 2 files changed, 148 insertions(+) create mode 100644 python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC.json create mode 100644 python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC_down.json diff --git a/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC.json b/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC.json new file mode 100644 index 000000000..92ab04b79 --- /dev/null +++ b/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC.json @@ -0,0 +1,74 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 6, + "USE_TMA": true + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4, + "USE_TMA": true + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4, + "USE_TMA": true + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4, + "USE_TMA": true + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4, + "USE_TMA": true + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4, + "USE_TMA": true + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 4, + "USE_TMA": true + }, + "16384": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3, + "USE_TMA": true + } +} diff --git a/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC_down.json b/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC_down.json new file mode 100644 index 000000000..0385bfe08 --- /dev/null +++ b/python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_7_1/E=32,N=1792,device_name=NVIDIA_B300_SXM6_AC_down.json @@ -0,0 +1,74 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 6, + "USE_TMA": true + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3, + "USE_TMA": true + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4, + "USE_TMA": true + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4, + "USE_TMA": true + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4, + "USE_TMA": true + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4, + "USE_TMA": true + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4, + "USE_TMA": true + }, + "16384": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 3, + "USE_TMA": true + } +}