enhancementhelp wanted
仓库指标
- 星标
- (1,513 个星标)
- PR 合并指标
- (平均合并 9天 17小时) (30 天内合并 200 个 PR)
描述
Currently I have already supported some weight checking logic. It would be great if we make it a test, and/or check every time. I do not have time recently thus write this as an issue.
I guess there may be a few things to do:
- Try to enable those flags and see whether there are any issues (or it works well and only enable that flag), especially if we want to enable it by default for all cases
- May implement some logic like bf16 training fp8 rollout checker etc
- May support things like "compute checksum in miles weight, compute checksum in sglang weight (after updates), check whether they are identical"