fix: load QLoRA model to CPU first, FSDP shards later

train.py

@@ -50,11 +50,11 @@ def train(config_path):
         model = AutoModelForCausalLM.from_pretrained(
             config["base_model"],
             quantization_config=bnb_config,
-            device_map="auto",
+            device_map="cpu",  # Load to CPU, FSDP shards later
             trust_remote_code=True,
             low_cpu_mem_usage=True,
         )
-        print("✓ Success: QLoRA 4-bit")
+        print("✓ Success: QLoRA 4-bit loaded to CPU")
     except Exception as e:
         errors.append(("QLoRA 4-bit", e))
         print(f"✗ Failed: {e}")