Move adapters back to LoRA to avoid inconvenient auto-PEFT trigger

modules.json

@@ -3,7 +3,7 @@
     "idx": 0,
     "name": "0",
     "path": "",
-    "type": "splade.SpladeCodeMLMTransformer"
+    "type": "sentence_transformers.sparse_encoder.models.MLMTransformer"
   },
   {
     "idx": 1,

splade.py

@@ -3,19 +3,27 @@ Compared to standard Qwen3, we're using bidirectional attention and not causal a
 with `is_causal=False` in the config.
 
 This file supports two loading paths:
-1. Sentence Transformers: `SparseEncoder("naver/splade-code-8B", trust_remote_code=True)` via SpladeCodeMLMTransformer -> AutoModelForMaskedLM -> Qwen3ForCausalLM
+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):
@@ -42,9 +50,10 @@ class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         from peft import PeftConfig, PeftModel
 
+        token = kwargs.get("token")
         try:
             peft_config = PeftConfig.from_pretrained(
-                pretrained_model_name_or_path, token=kwargs.get("token")
+                pretrained_model_name_or_path, subfolder=ADAPTER_SUBFOLDER, token=token
             )
         except Exception:
             peft_config = None
@@ -55,12 +64,7 @@ class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
         # 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,
@@ -70,7 +74,7 @@ class Qwen3ForCausalLM(TransformersQwen3ForCausalLM):
         )
 
         return PeftModel.from_pretrained(
-            base_model, pretrained_model_name_or_path, token=kwargs.get("token")
+            base_model, pretrained_model_name_or_path, subfolder=ADAPTER_SUBFOLDER, token=token
         )
 
 
@@ -128,7 +132,7 @@ 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
@@ -166,19 +170,3 @@ class Splade(PreTrainedModel):
 
 
 __all__ = ["Qwen3ForCausalLM", "Splade"]
-
-
-# Override ST's `_load_config` to return our `Qwen3Config` (with `auto_map`)
-# instead of a `PeftConfig`, so hub-path loads route to `splade.Qwen3ForCausalLM`
-# instead of failing in `AutoModelForMaskedLM`. The LoRA is still applied by
-# transformers' built-in PEFT path.
-try:
-    from sentence_transformers.sparse_encoder.models import MLMTransformer
-
-    class SpladeCodeMLMTransformer(MLMTransformer):
-        def _load_config(self, model_name_or_path, backend, config_kwargs):
-            return AutoConfig.from_pretrained(model_name_or_path, **config_kwargs), False
-
-    __all__.append("SpladeCodeMLMTransformer")
-except ImportError:
-    pass