Skip to content

fix(klx): fix qwen3.5-9B mm train/train_rollout_prob_abs_diff mismatch - #301

Open
grootZ wants to merge 1 commit into
redai-studio:mainfrom
grootZ:mr_master
Open

fix(klx): fix qwen3.5-9B mm train/train_rollout_prob_abs_diff mismatch#301
grootZ wants to merge 1 commit into
redai-studio:mainfrom
grootZ:mr_master

Conversation

@grootZ

@grootZ grootZ commented Sep 7, 2026

Copy link
Copy Markdown
Contributor

XSGL_QUANT_ALL_GATHER默认配置取到1进而导致rollout prefill阶段做了低精量化通信,精度偏低,属于逻辑错误。目前暂时通过env规避,下一版本xSGL镜像会全面修复这个逻辑错误。目前修复后训推diff合理,rollout阶段耗时主要集中在decode,故耗时基本无影响,性能与之前相近
image

@grootZ
grootZ requested a review from Yangruipis as a code owner September 7, 2026 10:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant