fix: load NVFP4 as bf16 by overriding quantization config

training/scripts/train.py

@@ -32,17 +32,17 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load model with BitsAndBytes INT4 (PEFT supports this)
+    # Load model and convert to bf16 (remove NVFP4 quantization)
     print(f"Loading model: {config['base_model']}")
-    from transformers import BitsAndBytesConfig
-    
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
-        quantization_config=BitsAndBytesConfig(load_in_4bit=True),
+        torch_dtype=torch.bfloat16,
         device_map="cpu",  # Load to CPU first
         trust_remote_code=True,
+        # Override any quantization config (NVFP4 -> bf16)
+        _fast_init=False,
     )
-    print("Model loaded with INT4 quantization.")
+    print("Model loaded and converted to bf16.")
     
     # Add LoRA
     lora_config = LoraConfig(