Patch loading SparseEncoder from Hub

modules.json

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

splade.py

@@ -3,7 +3,7 @@ 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 AutoModelForMaskedLM -> Qwen3ForCausalLM
+1. Sentence Transformers: `SparseEncoder("naver/splade-code-8B", trust_remote_code=True)` via SpladeCodeMLMTransformer -> 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`
@@ -166,3 +166,19 @@ 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