When temperature set to 0 and top_k set to 1 , the answers before and after restarting sglang are inconsistent

[Flynn-Zh]

Starting Qwen2.5-32B-Instruct-GPTQ-Int4 with sglang 0.3.2 and 0.3.4.post2, the v1/chat/compltions temperature is 0 and top_k is 1, when i restart sglang, the answers are inconsistent, who can give me some advice？

[merrymercy]

In 0.3.4.post2, we improved the sampling kernel to make it more consistent for greedy decoding.
This might change some results, but the new implementation is better so we will use that going forward!
#1589