Nicolas Patry
Nicolas Patry
@dvrogozh Can you stop calling it a bug everywhere, since it's not a bug, it's breaking change you are proposing, that you introduced in torch==2.5. The new behavior may be...
I made a cleaner implementation imho: https://github.com/huggingface/safetensors/pull/509 Can you check that it fixes your issue ? You're also more than welcome to steal the code from said PR so we...
The logic you kept is the logic I want to get rid of. It's still wrong to depend on torch internals (here the string representation of resolved device) on that...
Perfect done. Thanks again for raising awareness about the upcoming new behavior !
Superseeded by #509
Can you try with `1.6.0` to check ? We have been upgrading dependencies quite heavily recently, it's possible the version you had included an issue, even though I cannot reproduce...
1.7.0 is out, and works perfectly on our end, and even with 1.6.1 we're not able to reproduce the issue.
Fixed https://github.com/huggingface/hf-hub/pull/110
Thanks for this. I agree looking them up in other OSes would be nice, but since I don't own a keyboard with such high F I'm not going to bother...
We shouldn't need this tooling for the sync variant, because you should be issuing single calls per file in sync therefore the 429 are entirely on the end user to...