Add Flax weights

#19
by ArthurZ HF staff - opened
BigScience Workshop org

Automatically converted PyTorch weights to Flax weights. Using custom function

BigScience Workshop org

Those weights correspond to global step 91100 - I kept the tag just in case

BigScience Workshop org

I think that the weight map for the flax weights are missing my bad

BigScience Workshop org

@ArthurZ Are there tests that need to be run before this can be merged?

I hope it can work well with TPU-v3-32 in the future, this is probably the most cost-effective solution for using Bloom.

BigScience Workshop org

Oh sorry for the late reply! Not sure for the tests, I think it worked during the BLOOM FLAX sprint, I can double-check if needed

BigScience Workshop org

So what do we do this with this PR? Should we close as it's been a long time and still hasn't been merged without any development?

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment