vllm-project/vllm-ascend

[Misc]: Discussion on accuracy variance

已关闭

#6,623 创建于 2026年2月9日

 (2 条评论) (0 个反应) (0 位负责人)C++ (2,048 个派生)github user discovery
help wantedstalewait-feedback

仓库指标

星标
 (2,637 个星标)
PR 合并指标
 (平均合并 4天 5小时) (30 天内合并 559 个 PR)

描述

Anything you want to discuss about vllm on ascend.

Because of batch variance, We cannot guarantee that the same input will yield the same output in a multi-batch inference case. And this accuracy variance is more explicit in certain datasets.

model dataset acc acc variance
deepseek-v3.1 GPQA 74~82 8
qwen3-235b GPQA 64~71 7
qwen3-480b GPQA 60~67 7

We need to test these case on GPU or use other inference engine, such as sglang to check if this is an acc bug.

If GPU also has the same acc variance like us, we believe that this acc variance is reasonable for this dataset. Otherwise, we need to solve this acc bug.

贡献者指南