Use snapshot_download to avoid PEFT Windows issue

splade.py

@@ -48,19 +48,29 @@ class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        from huggingface_hub import snapshot_download
         from peft import PeftConfig, PeftModel
 
         token = kwargs.get("token")
-        try:
-            peft_config = PeftConfig.from_pretrained(
-                pretrained_model_name_or_path, subfolder=ADAPTER_SUBFOLDER, token=token
+
+        # Resolve the adapter to a local path before handing it to PEFT.
+        # PEFT's `subfolder=` kwarg uses `os.path.join` on Windows, producing
+        # backslashed hub paths that break the safetensors-vs-bin fallback.
+        if os.path.isdir(pretrained_model_name_or_path):
+            adapter_path = os.path.join(pretrained_model_name_or_path, ADAPTER_SUBFOLDER)
+        else:
+            local_repo = snapshot_download(
+                pretrained_model_name_or_path,
+                allow_patterns=[f"{ADAPTER_SUBFOLDER}/*"],
+                token=token,
             )
-        except Exception:
-            peft_config = None
+            adapter_path = os.path.join(local_repo, ADAPTER_SUBFOLDER)
 
-        if peft_config is None:
+        if not os.path.isfile(os.path.join(adapter_path, "adapter_config.json")):
             return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
+        peft_config = PeftConfig.from_pretrained(adapter_path, token=token)
+
         # Use provided splade config (has is_causal=False) or load it from the adapter repo
         config = kwargs.pop("config", None)
         if config is None or not isinstance(config, PretrainedConfig):
@@ -73,9 +83,7 @@ class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
             **kwargs,
         )
 
-        return PeftModel.from_pretrained(
-            base_model, pretrained_model_name_or_path, subfolder=ADAPTER_SUBFOLDER, token=token
-        )
+        return PeftModel.from_pretrained(base_model, adapter_path, token=token)
 
 
 class SpladeConfig(PretrainedConfig):