[Bugfix] Disable gptq_bitblas for <SM80 to fix GPTQ on V100/T4 (vllm-project#17541)

mgoin · Mu Huai · commit 5eae0e5134bf · 2025-05-12T19:19:25.000+08:00
Signed-off-by: mgoin &lt;mgoin64@gmail.com&gt;
Signed-off-by: Mu Huai &lt;tianbowen.tbw@antgroup.com&gt;
diff --git a/vllm/model_executor/layers/quantization/gptq_bitblas.py b/vllm/model_executor/layers/quantization/gptq_bitblas.py
@@ -134,7 +134,7 @@ def get_supported_act_dtypes(cls) -> List[torch.dtype]:
 
     @classmethod
     def get_min_capability(cls) -> int:
-        return 70
+        return 80
 
     @classmethod
     def get_config_filenames(cls) -> List[str]:
