update

examples/tutorials/dpo/ultrafeedback-dpo/step_2_train_dpo_model_single_gpu.py

@@ -168,7 +168,7 @@ def main():
         learning_rate=2e-5,
         warmup_steps=100,
         lr_scheduler_type="cosine",
-        fp16=True if torch.cuda.is_available() else False,
+        fp16=False,
         gradient_checkpointing=False,  # 如果内存紧张，可以设为 True
         optim="adamw_torch",
         report_to="none",