minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Intel GPU] Enable sgl-kernel-xpu fused_experts MoE kernel path for GPT-OSS bf16 models. (#22417)

Browse Source
This commit is contained in:
Chandrakant Khandelwal
2026-04-13 13:45:48 +08:00
committed by GitHub
parent d524f110ac
commit 1e9eecfa36
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/quantization/unquant.py
+2
View File
@@ -593,6 +593,8 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, MultiPlatformOp):
b1=getattr(layer, "w13_weight_bias", None), b1=getattr(layer, "w13_weight_bias", None),
b2=getattr(layer, "w2_weight_bias", None), b2=getattr(layer, "w2_weight_bias", None),
activation=moe_runner_config.activation, 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) return StandardCombineInput(hidden_states=output)
else: else: