diff --git a/training/scripts/train.py b/training/scripts/train.py
index 5850373..7420404 100755
--- a/training/scripts/train.py
+++ b/training/scripts/train.py
@@ -32,13 +32,14 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load model - let the model's own quantization config handle it
-    # (Ornith uses CompressedTensors, not BitsAndBytes)
-    # Load on CPU first, then DeepSpeed will distribute
+    # Load model in bf16 (fp8 not supported for training)
+    # Disable quantization config to train full precision
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        dtype=torch.bfloat16,
-        device_map="cpu",  # Load on CPU, DeepSpeed distributes
+        torch_dtype=torch.bfloat16,
+        device_map="cpu",
+        quantization_config=None,  # Override any fp8 quantization
+        trust_remote_code=True,
     )
     
     # Add LoRA