rexwang8

/

qilin-lit-6b

like 6

Hey there, This appears to be an issue on Huggingface's side. I uploaded a 12.2gb file and the SHA256 is correct on the website but is 11.7gb when downloaded with a different SHA256, I've tried reuploading but there is no change, I will try a few other solutions later.

Hi @rexwang8

I just deleted the file in Hugging Face's S3 storage, so that it can get uploaded once more. You should be able to reupload it.

@coyotte508 Unfortunately, that didn't seem to help. I reuploaded the file and it is still 11.4gb (0.8gb less) when I download it again.

The file uploaded is 12,216,970,719 bytes, which is 12.22 GB or 11.38 GiB. 1GiB = 1024 * 1024 * 1024 bytes, 1GB = 1000 * 1000 * 1000 bytes.

original file is 12216970719 bytes but the corrupted version is 11719138605 bytes, i managed to get the correct uncorrupted version from the huggingface cli but using koboldai and direct download both yield the corrupted version strangely

Could this be a caching issue where the new file was succesfully uploaded but cached copies are still being distributed until they expire?

Yes, maybe it's a caching issue either on Cloudfront's side or the browser's, if the corrupted file was uploaded with the correct file's info. Cloudfront caches files for a maximum of 24 hours for a geographic zone.

Downloading it in my zone (EU) gets the correct file. If trying to download it in a private browser window doesn't work, you may need to wait for the 24 hours.

you said you get the correct uncorrupted file on EU zone? It should be cloudfront then. I've tried with koboldAI and direct here, private window doesn't work yet. I'll check back later, thanks for the help.

Hi there is it fixed already? I really would like to try it.

Henk says he got it running on his side with koboldai now. @candymint23 , it should be ready to try.

We fixed an additional bug in the TPU version of KoboldAI that impacted the model. This model is now fully supported in KoboldAI.

@rexwang8 @Henk717 Many thanks!