Patch loading SparseEncoder from Hub (#2)

- Patch loading SparseEncoder from Hub (ebcd7f425e4d05f421138d56ce1a06a0dfc7d5ef)
- Move adapters back to LoRA to avoid inconvenient auto-PEFT trigger (0cf3c13ddf40fb955f6bad5368dcf419f1fb07bb)
- Use snapshot_download to avoid PEFT Windows issue (c7dfa2772eda14c97bfd94b9330d83637a00d735)

splade.py

@@ -6,16 +6,24 @@ This file supports two loading paths:
 1. Sentence Transformers: `SparseEncoder("naver/splade-code-8B", trust_remote_code=True)` via AutoModelForMaskedLM -> Qwen3ForCausalLM
 2. Transformers: `AutoModelForCausalLM.from_pretrained("naver/splade-code-8B", trust_remote_code=True)` -> Splade
 
-The checkpoint is distributed as a LoRA adapter on top of Qwen/Qwen3-8B; `Qwen3ForCausalLM.from_pretrained`
-loads the base model and applies the adapter.
+The checkpoint is distributed as a LoRA adapter on top of Qwen/Qwen3-8B in the `lora/` subfolder;
+`Qwen3ForCausalLM.from_pretrained` loads the base model and applies the adapter.
 """
 
+import os
+
 import torch
 from transformers import Qwen3ForCausalLM as TransformersQwen3ForCausalLM
 from transformers import PretrainedConfig, PreTrainedModel, AutoConfig
 from transformers.utils import is_flash_attn_2_available
 from .utils import prepare_tokenizer, splade_max, similarity, encode
 
+# The adapter lives in this subfolder rather than at the repo root so that
+# `find_adapter_config_file` doesn't trigger transformers' auto-PEFT path,
+# which would otherwise redirect hub loads to `Qwen/Qwen3-8B` and lose the
+# `auto_map` routing to the classes in this file.
+ADAPTER_SUBFOLDER = "lora"
+
 
 class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
     def tie_weights(self, *args, **kwargs):
@@ -40,27 +48,33 @@ 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
 
-        try:
-            peft_config = PeftConfig.from_pretrained(
-                pretrained_model_name_or_path, token=kwargs.get("token")
+        token = kwargs.get("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):
-            config = AutoConfig.from_pretrained(
-                pretrained_model_name_or_path, token=kwargs.get("token")
-            )
-
-        # We apply the adapter manually below, so drop any auto-PEFT hints to avoid double loading
-        kwargs.pop("adapter_kwargs", None)
+            config = AutoConfig.from_pretrained(pretrained_model_name_or_path, token=token)
 
         base_model = super().from_pretrained(
             peft_config.base_model_name_or_path,
@@ -69,9 +83,7 @@ class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
             **kwargs,
         )
 
-        return PeftModel.from_pretrained(
-            base_model, pretrained_model_name_or_path, token=kwargs.get("token")
-        )
+        return PeftModel.from_pretrained(base_model, adapter_path, token=token)
 
 
 class SpladeConfig(PretrainedConfig):
@@ -128,7 +140,7 @@ class Splade(PreTrainedModel):
         )
 
     def save_pretrained(self, save_directory, *args, **kwargs):
-        self.model.save_pretrained(save_directory)
+        self.model.save_pretrained(os.path.join(save_directory, ADAPTER_SUBFOLDER))
         self.config.save_pretrained(save_directory)
 
     @classmethod