[Bug] Identical inputs produce different results with temperature=0

I deployed my fine-tuned model using lmdeploy and connected it to the OpenAI API through a gRPC proxy. Recently, I noticed that for the same request and the same image, the model sometimes returns different results, even though I have already confirmed that temperature is set to 0. Later, I also set both top_k and top_p to 1, but the outputs can still vary.

Why is this happening?
Is there any way to explicitly control parameters such as do_sample (to disable sampling entirely) so that the model becomes strictly deterministic?