diff --git a/train.py b/train.py
index cb3ceee..289752e 100644
--- a/train.py
+++ b/train.py
@@ -33,26 +33,18 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load BnB 4-bit model (already quantized)
-    print(f"\n[INFO] Loading {config['base_model']} (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 BnB 4-bit model to single GPU
+    print(f"\n[INFO] Loading {config['base_model']} (BnB 4-bit) to GPU 0...")
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        quantization_config=bnb_config,
-        device_map="auto",
+        device_map="cuda:0",
         torch_dtype=torch.float16,
         trust_remote_code=True,
         low_cpu_mem_usage=True,
     )
-    print("✓ Success: Model loaded (BnB 4-bit)")
+    print("✓ Success: Model loaded to GPU 0 (BnB 4-bit)")
     print(f"   GPU 0: {torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
-    if torch.cuda.device_count() > 1:
-        print(f"   GPU 1: {torch.cuda.memory_allocated(1) / 1e9:.2f} GB")
+    print(f"   Free VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9 - torch.cuda.memory_allocated(0) / 1e9:.2f} GB")
     
     # Add LoRA
     lora_config = LoraConfig(