fix: load model snapshots from writable cache

Download base and adapter snapshots into the Space cache before loading so Transformers does not probe remote repos through /.cache.

Made-with: Cursor

environment/model_inference.py

@@ -77,20 +77,35 @@ def _load_model():
                 bnb_4bit_use_double_quant=True,
             )
 
-        peft_config = PeftConfig.from_pretrained(MODEL_ID, cache_dir=str(_CACHE_DIR))
+        from huggingface_hub import snapshot_download
+
+        adapter_path = snapshot_download(
+            repo_id=MODEL_ID,
+            cache_dir=str(_CACHE_DIR),
+        )
+        peft_config = PeftConfig.from_pretrained(adapter_path)
+        base_path = snapshot_download(
+            repo_id=peft_config.base_model_name_or_path,
+            cache_dir=str(_CACHE_DIR),
+            allow_patterns=[
+                "*.json",
+                "*.safetensors",
+                "*.model",
+                "*.txt",
+                "*.jinja",
+            ],
+        )
         base_model = AutoModelForCausalLM.from_pretrained(
-            peft_config.base_model_name_or_path,
+            base_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))
+        _TOKENIZER = AutoTokenizer.from_pretrained(adapter_path)
         _MODEL = PeftModel.from_pretrained(
             base_model,
-            MODEL_ID,
-            cache_dir=str(_CACHE_DIR),
+            adapter_path,
         )
         _MODEL.eval()
         return _MODEL, _TOKENIZER