Label Knight
Label Knight
I meet the same question, I have changed the config and it show right if only use matplotlib, but pandas profiling doesn't work @sbrugman
我今天又看了下，我测试的是14B，如果是72B full的话，即使是80G的显卡也撑不过model.float()
我也遇到同样的问题
> torch.where call the torch.nonzer o function
> 可以尝试用sat微调：https://github.com/THUDM/CogVLM/tree/cogvlm2_dev ，把微调代码的CogVLMModel换成CogVLM2Model就可以了。后续会merge到这个仓库里。 merge了吗？现在还是波动较大
@Victarry Thank you for your work and support, but it feels a bit confusing now. (1) I found 1671 (**https://github.com/NVIDIA/Megatron-LM/pull/1671**) in the pull request, which seems to have implemented this...
try set --init-method-std to 0.02 ?