bfloat16

app.py

@@ -25,7 +25,7 @@ def upload(hf_token, base_model_name_or_path, peft_model_path, output_dir):
 
       info_fn(f"Loading base model: {base_model_name_or_path}")
 
-      base_model = AutoModelForCausalLM.from_pretrained(base_model_name_or_path, torch_dtype=torch.float16, **device_arg)
+      base_model = AutoModelForCausalLM.from_pretrained(base_model_name_or_path, torch_dtype=torch.bfloat16, **device_arg)
 
       info_fn(f"Loading PEFT: {peft_model_path}")