update

examples/tutorials/grpo/step_2_train_grpo_model_unsloth_ddp.py

@@ -93,7 +93,7 @@ else:
 import unsloth
 from datasets import load_dataset
 from trl import GRPOConfig, GRPOTrainer
-from unsloth import FastLanguageModel
+from unsloth import FastLanguageModel, is_bf16_supported
 
 
 def get_args():
@@ -395,8 +395,8 @@ def main():
         load_best_model_at_end=True,
         metric_for_best_model="reward",  # 使用奖励作为评估指标
         greater_is_better=True,
-        fp16=True,
-        bf16=False,
+        fp16=False if is_bf16_supported() else True,
+        bf16=True if is_bf16_supported else False,
         max_grad_norm=1.0,
         report_to="none",  # 可根据需要改为"wandb"等