huggingface / candle

Minimalist ML framework for Rust
Apache License 2.0
15.12k stars 884 forks source link

Support for sharded checkpoints #1564

Open zlstringham opened 8 months ago

zlstringham commented 8 months ago

For instance, there are examples for BLIP, but BLIP2 uses sharded checkpoints; e.g. https://huggingface.co/Salesforce/blip2-opt-2.7b

It would be good to support this use-case out of the box.

LaurentMazare commented 8 months ago

In this case, I'm not sure the difficulty is supporting sharded checkpoints as we already have numerous example using these but rather that blip2 uses a different text model than the original blip so this model will have to be ported to candle.