修复CompressedTensorsLinearMethod中的w4a16的冲突问题

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -723,6 +723,16 @@ class CompressedTensorsLinearMethod(LinearMethodBase):
         self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
         
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        weights_scheme = (
+            self.quantization_config
+            .target_scheme_map.get('Linear', {})
+            .get('weights')
+        )
+        if weights_scheme is not None:
+            num_bits = weights_scheme.num_bits
+            if num_bits == 4:
+                return layer.scheme.process_weights_after_loading(layer)   
+
         n=layer.weight.shape[0]
         k=layer.weight.shape[1]
         

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py

@@ -197,5 +197,5 @@ class CompressedTensorsWNA16(CompressedTensorsScheme):
         self.kernel.process_weights_after_loading(layer)
 
     def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
-                      bias: Optional[torch.Tensor]) -> torch.Tensor:
+                      bias: Optional[torch.Tensor], **kw) -> torch.Tensor:
         return self.kernel.apply_weights(layer, x, bias)