Let's start discussing implementation.
- Need to expose the quantization scripts (either included here or add
doc on how to use https://github.com/qwopqwop200/GPTQ-for-LLaMa)
- Make sure GPTQ works for multiple models (priority to Falcon).
Currently it means that every place we use `get_{tensor|sharded}` to
check for quantization.
My idea is to reintegrate as much as possible into `utils/layer.py` by
expanding `load_multi` to be a bit more generic.
This might require some thinking, but ultimately the
`qweight,qzeros,scales,g_idx` should be in a single place, and
independant of bias presence.
# What does this PR do?
Reworked the loading logic. Idea is to use cleaner loading code:
- Remove need for `no_init_weights`
- Remove all weird `bnb_linear` and `load_weights` and
`post_load_weights`.
New code layout:
- New class `Weights` in charge of handling loading the weights from
multiple files into appropiate tensors (potentially sharded)
- TP layers now are "shells", they contain the code to know what kind of
sharding we need + eventual `all_reduce`. They do not inherit from
linear, but they contain some kind of Linear instead
- the contained linear can be either FastLinear, BnbLinear or GPTq
Linear next.
- All modeling code is explictly made for sharding, process group is
just no-ops for non sharded code (removes a lot of test cases)
---------
Co-authored-by: Ubuntu <ubuntu@ip-172-31-41-161.taildb5d.ts.net>
Co-authored-by: Ubuntu <ubuntu@ip-172-31-41-161.ec2.internal>
Co-authored-by: OlivierDehaene <olivier@huggingface.co>
Co-authored-by: OlivierDehaene <23298448+OlivierDehaene@users.noreply.github.com>
