Update src/llm/local_model_client.py

src/llm/local_model_client.py

@@ -1,4 +1,5 @@
 """Local Be.FM model client - OPTIMIZED for MPS performance"""
+"""Local Be.FM model client - OPTIMIZED for MPS performance"""
 
 import os
 import gc
@@ -99,10 +100,11 @@ class LocalModelClient:
             print(f"  [3/5] Loading base model ({self.base_model_id})...")
             if self.device == "cuda":
                 # Use 8-bit quantization to fit in 15GB GPU (saves ~50% memory)
+                # Don't use device_map="auto" - it causes meta tensor issues
                 self._model = AutoModelForCausalLM.from_pretrained(
                     self.base_model_id,
                     load_in_8bit=True,
-                    device_map="auto",
+                    device_map={"": 0},  # Load everything on GPU 0, avoiding meta tensors
                     token=hf_token,
                 )
             else: