search

facebookresearch / fairseq

Facebook AI Research Sequence-to-Sequence Toolkit written in Python.
MIT License
30.43k stars 6.4k forks source link

broken xlmr.large.tar.gz file #1487

Closed rlebret closed 4 years ago

rlebret commented 4 years ago

Hi,

It seems that there is a problem with the new XLM-R large archive available from: https://dl.fbaipublicfiles.com/fairseq/models/xlmr.large.tar.gz It returns a file of size 1.1GB and there is a gzip error when uncompressing the file.

Best, Rémi

stefan-it commented 4 years ago

I could extract it with tar -xf xlmr.large.tar.gz, so the .gz extension is a bit misleading 😅

myleott commented 4 years ago

@ngoyal2707

ngoyal2707 commented 4 years ago

Thanks for flagging, sorry about the wrong format. It's fixed now.

rlebret commented 4 years ago

Thanks for the update. Out of curiosity, how do you explain that xlmr.large.v0 model size was 5.3GB while xlmr.large.v1 is now 1.1GB? In the meantime, the size of xlmr.base remains the same (2.6GB).

myleott commented 4 years ago

This was because the v0 checkpoints mistakenly included the optimizer state, which tripled the file size. Similarly the xlmr.base release also had optimizer states, which we've now removed.