From a9d5ca723a30b80655954688eef6e0e66cacb025 Mon Sep 17 00:00:00 2001 From: YanbingJiang Date: Mon, 31 Aug 2026 10:14:41 +0800 Subject: [PATCH] [CPU] Fix weight missing issue in fused_input_proj_cpu for GPTQ INT4 for Qwen 3.5 (#35805) --- python/sglang/srt/models/qwen3_5.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/models/qwen3_5.py b/python/sglang/srt/models/qwen3_5.py index f1f8582a6..6be5e3a89 100644 --- a/python/sglang/srt/models/qwen3_5.py +++ b/python/sglang/srt/models/qwen3_5.py @@ -388,8 +388,10 @@ class Qwen3_5GatedDeltaNet(nn.Module): self._fused_input_proj_cpu_enabled = LazyValue( lambda: ( _is_cpu - and self.in_proj_qkvz.weight.dtype == torch.bfloat16 - and self.in_proj_ba.weight.dtype == torch.bfloat16 + and self.in_proj_qkvz._parameters.get("weight") is not None + 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_ba.bias is None and use_intel_amx_backend(self.in_proj_qkvz)