tong wang
tong wang
they did not change the model structure, just finetuned on more data with objaverse-xl as suggested in objaverse-xl's preprint FYI: https://objaverse.allenai.org/objaverse-xl-paper.pdf
It is likely you ran out of RAM when loading shards. You can verify this by running htop and observing Mem and Swp. If this is the case, you can...
this issue makes vllm impossible for production use
512x512
you would need to cat all the parts first then unzip. for example: `zip -FF 1.zip --out 1-full.zip; unzip -FF 1-full.zip`
> Those codes changes the dimensional order of the original values. When the func is called at second time to same data, the dimensional order is wrong. I can confirm...
same here
> Looks like [5314e67](https://github.com/tlkh/asitop/commit/5314e67192e78e5391a98c8c806f19cd1c3145fe) may be the necessary fix Yes, or simply use gavi's fork fixed the issue `pip install git+https://github.com/gavi/asitop.git`
https://openreview.net/forum?id=XKeSauhUdJ 作者在openreview审稿阶段上传的Supplementary Material提到了具体的训练细节，我照着里面提供的超参训vae没遇到什么问题。 hope it helps
This work does not use cfg. But you can implement it by treating embeddings of a glyph image with no text drawn as the null prompt. By training the model...