metaseq VS YaLM-100B

This is a super important issue that affects the pace and breadth of iteration of AI almost as much as the raw hardware improvements do. The blog is fun but somewhat shallow and not technical or very surprising if you’ve worked with clusters of GPUs in any capacity over the years. (I liked the perspective of a former googler, but I’m not sure why past colleagues would recommend Jax over pytorch for LLMs outside of Google.) I hope this newco eventually releases a more technical report about their training adventures, like the PDF file here: https://github.com/facebookresearch/metaseq/tree/main/projec...

The number of people who can actually pre-train a true LLM is very small.

It remains a major feat with many tweaks and tricks. Case in point: the 114 pages of OPT175B logbook [1]

[1] https://github.com/facebookresearch/metaseq/blob/main/projec...

Reliability at scale: take a look at the OPT training log book for their 175B model run. It needed a lot of babysitting. In my experience, that scale of TPU training run requires a restart about once every 1-2 weeks—and they provide the middleware to monitor the health of the cluster and pick up on hardware failures.

I really appreciated this log of Facebook training a large language model of how troublesome AI development can be: https://github.com/facebookresearch/metaseq/tree/main/projects/OPT/chronicles

Stable Diffusion will run on any decent gaming GPU or a modern MacBook, meanwhile LLMs comparable to GPT-3/ChatGPT have had pretty insane memory requirements - e.g., <https://github.com/facebookresearch/metaseq/issues/146>

It seems so, check this log book from Meta: https://github.com/facebookresearch/metaseq/blob/main/projec...

Its probably just open source software/training sets repurposed... https://github.com/yandex/YaLM-100B

There is a 100b model published on pache 2.0 license. Though there is no information about finetuning it or using in 4-bit with smth like llama.cpp. Trying to figure out how to try it without renting extremely expensive gpu set. https://github.com/yandex/YaLM-100B

Hey everyone! I just discovered an open-source 100 billion parameter language model called YaLM, which is published under the Apache 2.0 license. The model is trained on more than 1 TB of Russian and English text. Here's the GitHub repo: https://github.com/yandex/YaLM-100B and an article explaining how it was trained: https://medium.com/yandex/yandex-publishes-yalm-100b-its-the-largest-gpt-like-neural-network-in-open-source-d1df53d0e9a6

Yandex has already released a LLM: https://github.com/yandex/YaLM-100B

So there is this open project YaLM 100B require 200 GB of disk space, it is trained on 1.7 TB of text

You'd think so, but to train a model like the one CAI uses, it would require truly jaw-breaking amount of funds. That's why CAI is so suspicious tbh. Just to give you an example, YaML (100 billion parameters which is probably less than CAI) took 65 days to train, and 800 A100 graphics cards. 175 billion parameters would not be 1.75 times higher because it's not a linear function. It would probably be 10x or even more. IIRC, "Open"Ai could only afford to train GPT-3 a single time...

I don't much follow AI news beyond what I randomly happen to see on HN, but this might still be the largest open source model: https://github.com/yandex/YaLM-100B . There's discussion of it here: https://old.reddit.com/r/MachineLearning/comments/vpn0r1/d_h... - at the bottom of that page is a comment from someone who actually ran it in the cloud.

Hilariously, Yandex Alisa runs circles around it, because it's not just a collection of gimmicks but has an actual 100B-class language model (YaLM, opensourced) as its core, plus lots of decent engineering. It's helpful, skillful and feels alive, almost like ChatGPT.