fix: load already-quantized 4-bit model directly (no BnB)

train.py

@@ -33,23 +33,16 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load model with 4-bit quantization (fit in single GPU)
-    print(f"\n[INFO] Loading {config['base_model']} with BnB 4-bit...")
-    from transformers import BitsAndBytesConfig
-    bnb_config = BitsAndBytesConfig(
-        load_in_4bit=True,
-        bnb_4bit_quant_type="nf4",
-        bnb_4bit_compute_dtype=torch.float16,
-    )
+    # Load model (already 4-bit on disk, just move to GPU)
+    print(f"\n[INFO] Loading {config['base_model']} to GPU 0...")
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
         device_map="cuda:0",
-        quantization_config=bnb_config,
         torch_dtype=torch.float16,
         trust_remote_code=True,
         low_cpu_mem_usage=True,
     )
-    print("✓ Success: Model loaded to GPU 0 (4-bit)")
+    print("✓ Success: Model loaded to GPU 0")
     print(f"   GPU 0: {torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
     print(f"   Free VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9 - torch.cuda.memory_allocated(0) / 1e9:.2f} GB")