Re-release consolidated OPT / OPT-IML checkpoints

[suchenzang]

After #459 and #556, we can now release updated checkpoints that are consolidated from FSDP shards with different model parallelism as well. We should update all of our checkpoints as a start to help address some of the following painpoints that users are facing:

• Why does it unreadable generated by OPT-30B  #595
• How to load an OPT-IML checkpoint and use it for inference? #594
• How do I get reshard.pt to run OPT175B? #574
• Error when run reshard_fsdp on opt-IML 30b for inference #567
• Can not load OPT-175B model. #475
• convert_to_singleton seems to hang for OPT-66B #407
• Convert to singleton script fails for 1.3B checkpoint #233
• How to load pretrained weight in finetuning? #211

and previous issues:

• How to load sharded checkpoints? #31
• shard_metadata not found #125

We have internal consolidated versions for 2.7B and 30B to check against, and will also need to confirm that generation looks roughly sane after consolidation.

[EIFY]

I can work around it, but could the following issue be considered related as well?

• hub_utils.py assumes a different sharding convention #552

[andrewPoulton]

Yeah looks like it, at least tangentially - the loading logic there could probably do with simplifying. It should be possible to identify naming convention by just reading the checkpoint directory.