YHPeter comments

Results 5 comments of


                                            YHPeter

[BUG] ZERO++ | AssertionError: ZeRO parameter intra parallel group is already initialized

same error

VSCode GitHub Copilot Edit tool cannot be used to edit file open from Overleaf Workshop

I'm using MacOS, get the same problem. Within overleaf worksapce, the Copilot is not sendable and request. No error, warns occurs, just cannot send the request.

[Bug]: Data Parallel with embed task hangs.

I also face such problem.

[Bug]: Data Parallel with embed task hangs.

Fallback to V0, also not working.

微调qwen3vl8B, 除第一个batch，后面的loss均为0是为什么，数据均按对应格式构造的

问题比较奇怪应该不是qwen3-vl的问题我在三个不同的集群上面都是H100 基本上一样的环境两个有一样的现象很神奇 Gemma2, Gemma3, Qwen3, Qwen2都一样 {'loss': 14.6752, 'grad_norm': 1.4142135381698608, 'learning_rate': 0.0, 'epoch': 0.0} 其中一个把FA3->FA2就好了还有个实在无法解决均是Z1/Z2