diff --git a/train.py b/train.py
index 289752e..a547fb8 100644
--- a/train.py
+++ b/train.py
@@ -33,16 +33,46 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load BnB 4-bit model to single GPU
-    print(f"\n[INFO] Loading {config['base_model']} (BnB 4-bit) to GPU 0...")
+    # Load bf16 to CPU first
+    print(f"\n[INFO] Loading {config['base_model']} bf16 to CPU...")
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        device_map="cuda:0",
+        device_map="cpu",
+        torch_dtype=torch.bfloat16,
+        trust_remote_code=True,
+        low_cpu_mem_usage=True,
+    )
+    print("✓ Model loaded to CPU (~70GB bf16)")
+    
+    # Quantize with BnB on CPU
+    print("   Quantizing with BnB 4-bit on CPU...")
+    from transformers import BitsAndBytesConfig
+    bnb_config = BitsAndBytesConfig(
+        load_in_4bit=True,
+        bnb_4bit_quant_type="nf4",
+        bnb_4bit_compute_dtype=torch.float16,
+    )
+    
+    # Reload with quantization
+    del model
+    import gc
+    gc.collect()
+    torch.cuda.empty_cache()
+    
+    model = AutoModelForCausalLM.from_pretrained(
+        config["base_model"],
+        quantization_config=bnb_config,
+        device_map="cpu",
         torch_dtype=torch.float16,
         trust_remote_code=True,
         low_cpu_mem_usage=True,
     )
-    print("✓ Success: Model loaded to GPU 0 (BnB 4-bit)")
+    print("✓ Model quantized to 4-bit on CPU")
+    
+    # Move to GPU
+    print("   Moving to GPU 0...")
+    model = model.to("cuda:0")
+    print("✓ Success: Model loaded to GPU 0 (4-bit)")
     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")