mesh-transformer-jax VS hivemind

GPT-J is a LLM case study with two goals: Training a LLM with a data source containing unique material, and using the training frameworkMesh Transformer JAX to achieve a high training efficiency through parallelization. There is no research paper about GPT-J, but on its GitHub pages, the model, different checkpoints, and the complete source code for training is given.

This idea has already been proposed in ViT-22B and GPT-J-6B.

Sure. Here's the repo I used for the fine-tuning: https://github.com/kingoflolz/mesh-transformer-jax. I used 5 epochs, and appart from that I kept the default parameters in the repo.

You can skip to step 4 using something like GPT-J as far as I understand: https://github.com/kingoflolz/mesh-transformer-jax#links

The pretrained model is already available.

The Github repo suggests it's possible you can change the number of checkpoints to make it run on a GPU.

GPT-J, which uses mesh-transformer-jax: https://github.com/kingoflolz/mesh-transformer-jax

They hinted at it in the screenshot, but the goods are linked from the https://6b.eleuther.ai page: https://github.com/kingoflolz/mesh-transformer-jax#gpt-j-6b (Apache 2)

https://github.com/learning-at-home/hivemind is also relevant

I really hope you'll join me, for the Petals support, at least! A single docker-compose.yml file is all we need, for now. If we are able to find enough people willing to host some smaller models, perhaps we could expand into the Hivemind, and create our own, custom foundation model one day?

Decentralized deep learning: https://github.com/learning-at-home/hivemind

https://github.com/learning-at-home/hivemind - same people behind it, was made before petals I think.

not if we rise to meet that challenge. here's a few tools that facilitate AI research in the face of an advanced persistent threat: Hivemind- a distributed Pytorch framework

yeah, there's Hivemind. and there's research wrt how to chunk out training workload so it can be scaled up. not sure why there's commentary that latency issues would limit this sort of enterprise, the architecture typically isn't designed for liveness. other subfields of distributed training/inference include zero-knowledge machine learning. besides all of that, there's also adversarial computation like SafetyNets and refereed delegation of computation.

We already have the software for it. There are some projects, but the one I'm most familiar with is https://github.com/learning-at-home/hivemind for training and it's sister project https://petals.ml/ for running large models distributed.

I'm not entirely how the approach they're using works [0], but I study federated learning and one of the highly-cited survey papers has several chapters (5 and 6 in particular) addressing potential attacks, failure modes, and bias [1].

0: https://github.com/learning-at-home/hivemind

1: https://arxiv.org/abs/1912.04977

The Hivemind project is just that

https://github.com/learning-at-home/hivemind