Update grpo_train.py

grpo_train.py

@@ -305,6 +305,7 @@ USE_4BIT = not torch.cuda.is_available() or torch.cuda.get_device_properties(0).
 model, tokenizer = FastLanguageModel.from_pretrained(
     model_name="unsloth/Llama-3.1-8B-Instruct",
     load_in_4bit=USE_4BIT,
+    dtype = torch.bfloat16,
     max_seq_length=2048,
     dtype=None,  # auto-detect bf16 on A100
 )