fix: load QLoRA directly to GPU (skip CPU)

training/scripts/train.py

@@ -47,7 +47,7 @@ def train(config_path):
         config["base_model"],
         quantization_config=bnb_config,
         dtype=torch.bfloat16,
-        device_map="cpu",  # Load to CPU first
+        device_map="auto",  # Load directly to GPU
         trust_remote_code=True,
     )
     print("Model loaded with QLoRA (4-bit).")