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

[CPU] Fix weight missing issue in fused_input_proj_cpu for GPTQ INT4 for Qwen 3.5 (#35805)

Browse Source
This commit is contained in:
YanbingJiang
2026-08-31 10:14:41 +08:00
committed by GitHub
parent 4f761e8649
commit a9d5ca723a
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen3_5.py
+4 -2
View File
@@ -388,8 +388,10 @@ class Qwen3_5GatedDeltaNet(nn.Module):
self._fused_input_proj_cpu_enabled = LazyValue( self._fused_input_proj_cpu_enabled = LazyValue(
lambda: ( lambda: (
_is_cpu _is_cpu
and self.in_proj_qkvz.weight.dtype == torch.bfloat16 and self.in_proj_qkvz._parameters.get("weight") is not None
and self.in_proj_ba.weight.dtype == torch.bfloat16 and self.in_proj_ba._parameters.get("weight") is not None
and self.in_proj_qkvz._parameters["weight"].dtype == torch.bfloat16
and self.in_proj_ba._parameters["weight"].dtype == torch.bfloat16
and self.in_proj_qkvz.bias is None and self.in_proj_qkvz.bias is None
and self.in_proj_ba.bias is None and self.in_proj_ba.bias is None
and use_intel_amx_backend(self.in_proj_qkvz) and use_intel_amx_backend(self.in_proj_qkvz)