fix: load peft base model with explicit cache

Load the LoRA base model and adapter separately so both Hugging Face downloads use the Space writable cache instead of falling back to /.cache.

Made-with: Cursor

environment/model_inference.py

@@ -65,8 +65,8 @@ def _load_model():
         if _MODEL is not None and _TOKENIZER is not None:
             return _MODEL, _TOKENIZER
 
-        from peft import AutoPeftModelForCausalLM
-        from transformers import AutoTokenizer, BitsAndBytesConfig
+        from peft import PeftConfig, PeftModel
+        from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
 
         quantization_config = None
         if torch.cuda.is_available():
@@ -77,15 +77,21 @@ def _load_model():
                 bnb_4bit_use_double_quant=True,
             )
 
-        _TOKENIZER = AutoTokenizer.from_pretrained(MODEL_ID, cache_dir=str(_CACHE_DIR))
-        _MODEL = AutoPeftModelForCausalLM.from_pretrained(
-            MODEL_ID,
+        peft_config = PeftConfig.from_pretrained(MODEL_ID, cache_dir=str(_CACHE_DIR))
+        base_model = AutoModelForCausalLM.from_pretrained(
+            peft_config.base_model_name_or_path,
             device_map="auto" if torch.cuda.is_available() else "cpu",
             torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
             quantization_config=quantization_config,
             low_cpu_mem_usage=True,
             cache_dir=str(_CACHE_DIR),
         )
+        _TOKENIZER = AutoTokenizer.from_pretrained(MODEL_ID, cache_dir=str(_CACHE_DIR))
+        _MODEL = PeftModel.from_pretrained(
+            base_model,
+            MODEL_ID,
+            cache_dir=str(_CACHE_DIR),
+        )
         _MODEL.eval()
         return _MODEL, _TOKENIZER