Why is the result variable when using vllm for inference, even though I’ve set the temperature to 0.01? The result should be fixed, right?

[hellen9527]

I directly used vllm to load the model saved after training with GRPO, and during inference, I used one data point from the training set. The prompt format is the same, and the temperature is set to 0.01, but I noticed that the result is different each time. After about 10 runs, it starts repeating. Is this an issue with the GRPO training mechanism, or could it be a problem with the saved model?