fix moe quant (#3478)

Author: Jintao-Huang

swift/llm/export/quant.py

@@ -180,7 +180,8 @@ def get_block_name_to_quantize(self, model: nn.Module) -> Optional[str]:
 
         module_lists = []
         for n, m in model.named_modules():
-            if isinstance(m, nn.ModuleList) and len(m) >= 10:
+            if (isinstance(m, (nn.ModuleList, nn.Sequential)) and len(m) >= 10
+                    and 'mlp' not in m[0].__class__.__name__.lower()):  # fix moe
                 module_lists.append((n, m))
         if module_lists:
             module_list = max(module_lists, key=lambda x: len(x[1]))

swift/llm/utils.py

@@ -90,7 +90,8 @@ def find_module_list(model) -> Optional[nn.ModuleList]:
     for m in model.modules():
         if hasattr(m, 'gradient_checkpointing') or m.__class__.__name__ == 'CheckpointWrapper':
             return
-        if isinstance(m, (nn.ModuleList, nn.Sequential)) and len(m) >= 10:
+        if (isinstance(m, (nn.ModuleList, nn.Sequential)) and len(m) >= 10
+                and 'mlp' not in m[0].__class__.__name__.lower()):  # fix moe
             module_lists.append(m)
     if module_lists:
         return max(module_lists, key=lambda x: len(x))