@CheeJerry
@qingyao_sun
@volokuleshov
@chrismdesa
It looks like a lot of people can't see the second post:
(2/n TLDR) We’ve released quantized LLaMA, Mistral, and OpenHermes models and a full codebase that you can quantize and deploy your own models with at .
👉QuIP# is even better now! Updated QuIP# adds fine-tuning for SOTA quantized 2, 3, & 4 bit LLMs. With QuIP#, 3 bit scales better than 4 bit for the first time!
💻
📜
Deep models need data, and annotating lots of behavior data is costly. In our
@CVPR
paper, we introduce AutoSWAP, a framework that uses program synthesis and weak supervision to train models with a fraction of the labels!
📄&👊🃏:
(1/n)
Turns out QuIP# works pretty well at 1 bit too. In fact, it's close to original QuIP at 2 bits.
Example generation from 1 bit 2-70b: "Cornell University is located in Ithaca, NY, which is known for its beautiful scenery. Ithaca is a small city of about 16,000 people ..."
👉QuIP# is even better now! Updated QuIP# adds fine-tuning for SOTA quantized 2, 3, & 4 bit LLMs. With QuIP#, 3 bit scales better than 4 bit for the first time!
💻
📜
🧵 (2/n)
QuIP# crushes all publicly available 2 bit PTQ methods on language modeling & zero shot tasks while being conceptually clean and simple.
We’ve released quantized LLaMA, Mistral, and OpenHermes models, and a full codebase at
Special thanks to co-authors
@JenJSun
and
@yisongyue
! Our poster will also make an appearance at , make sure to check out MABe 2022!
Also, why is there no code emoji?
(5/n)
In AutoSWAP, domain experts provide domain knowledge in a reusable domain-specific language (from program synthesis) and domain-level labeling functions (LFs, noisy heuristic functions from weak supervision).
(2/n)