Update grpo_train.py

grpo_train.py

@@ -362,7 +362,7 @@ model, tokenizer = FastLanguageModel.from_pretrained(
     model_name="unsloth/Llama-3.1-8B-Instruct",
     load_in_4bit=USE_4BIT,
     max_seq_length=2048,
-    dtype=None,
+    dtype=torch.float16 if USE_4BIT else None,
 )
 
 model = FastLanguageModel.get_peft_model(