fix: convert FP8 to bf16 for PEFT compatibility

training/scripts/train.py

@@ -32,15 +32,15 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load bf16 model
+    # Load model and convert to bf16 (ignore FP8 quantization)
     print(f"Loading model: {config['base_model']}")
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        torch_dtype=torch.bfloat16,
+        torch_dtype=torch.bfloat16,  # Convert FP8 -> bf16
         device_map="cpu",  # Load to CPU first
         trust_remote_code=True,
     )
-    print("Model loaded.")
+    print("Model loaded and converted to bf16.")
     
     # Add LoRA
     lora_config = LoraConfig(