From db1f9e6165251bad59156f414c6a87b7cb3e9ea2 Mon Sep 17 00:00:00 2001 From: Xiake Sun Date: Fri, 19 Dec 2025 18:21:52 +0800 Subject: [PATCH 1/2] Add tuned config for triton_fused_moe for Qwen3-Omni on MI308X Signed-off-by: Xiake Sun --- .../E=128,N=96,device_name=AMD_Instinct_MI308X.json | 12 ++++++++++++ 1 file changed, 12 insertions(+) create mode 100755 python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json new file mode 100755 index 000000000000..13ede06d9f18 --- /dev/null +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json @@ -0,0 +1,12 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 2, + "matrix_instr_nonkdim": 16 + } +} From c590a66aed8ef168f56fc54b110d1843fb4687d5 Mon Sep 17 00:00:00 2001 From: Xiake Sun Date: Fri, 19 Dec 2025 18:26:38 +0800 Subject: [PATCH 2/2] Fix format Signed-off-by: Xiake Sun --- .../E=128,N=96,device_name=AMD_Instinct_MI308X.json | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json index 13ede06d9f18..b7cc38452340 100755 --- a/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=96,device_name=AMD_Instinct_MI308X.json @@ -7,6 +7,6 @@ "num_warps": 2, "num_stages": 2, "waves_per_eu": 2, - "matrix_instr_nonkdim": 16 + "matrix_instr_nonkdim": 16 } }