From 1e9eecfa36ce899d27cb2d794a083049ba82d937 Mon Sep 17 00:00:00 2001 From: Chandrakant Khandelwal Date: Mon, 13 Apr 2026 11:15:48 +0530 Subject: [PATCH] [Intel GPU] Enable sgl-kernel-xpu fused_experts MoE kernel path for GPT-OSS bf16 models. (#22417) --- python/sglang/srt/layers/quantization/unquant.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/python/sglang/srt/layers/quantization/unquant.py b/python/sglang/srt/layers/quantization/unquant.py index 7339640c4..3cc28e2ae 100644 --- a/python/sglang/srt/layers/quantization/unquant.py +++ b/python/sglang/srt/layers/quantization/unquant.py @@ -593,6 +593,8 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp): b1=getattr(layer, "w13_weight_bias", None), b2=getattr(layer, "w2_weight_bias", None), activation=moe_runner_config.activation, + gemm1_alpha=moe_runner_config.gemm1_alpha, + gemm1_limit=moe_runner_config.gemm1_clamp_limit, ) return StandardCombineInput(hidden_states=output) else: