diff --git a/train.py b/train.py
index 4f459ce..c50c6d1 100644
--- a/train.py
+++ b/train.py
@@ -38,24 +38,18 @@ def train(config_path):
     errors = []
     
     # ------------------------------------------------------------------
-    # Strategy 1: QLoRA 4-bit with device_map="auto" (distributed across GPUs, no FSDP)
+    # Strategy 1: Load already-quantized 4-bit model (distributed across GPUs)
     # ------------------------------------------------------------------
-    print("\n[1/4] Trying: 4-bit QLoRA (distributed across GPUs, no FSDP)...")
+    print("\n[1/4] Trying: 4-bit model AS-IS (distributed across GPUs)...")
     try:
-        bnb_config = BitsAndBytesConfig(
-            load_in_4bit=True,
-            bnb_4bit_quant_type="nf4",
-            bnb_4bit_compute_dtype=torch.bfloat16,
-            bnb_4bit_use_double_quant=True,
-        )
         model = AutoModelForCausalLM.from_pretrained(
             config["base_model"],
-            quantization_config=bnb_config,
             device_map="auto",  # Distribute layers across GPUs automatically
+            torch_dtype=torch.float16,
             trust_remote_code=True,
             low_cpu_mem_usage=True,
         )
-        print("✓ Success: QLoRA 4-bit distributed across GPUs (no FSDP)")
+        print("✓ Success: 4-bit model distributed across GPUs (no FSDP)")
     except Exception as e:
         errors.append(("QLoRA 4-bit", e))
         print(f"✗ Failed: {e}")