transformers VS Ray

The HuggingFace transformers library already has support for a similar method called prompt lookup decoding that uses the existing context to generate an ngram model: https://github.com/huggingface/transformers/issues/27722

I don't think it would be that hard to switch it out for a pretrained ngram model.

Most of this tutorial is based on Hugging Face course about Transformers and on Niels Rogge's Transformers tutorials: make sure to check their work and give them a star on GitHub, if you please ❤️

* Superconvergence + LR range finder + Fast AI's Ranger21 optimizer was the goto optimizer for CNNs, and worked fabulously well, but on transformers, the learning rate range finder sadi 1e-3 was the best, whilst 1e-5 was better. However, the 1 cycle learning rate stuck. https://github.com/huggingface/transformers/issues/16013

Thanks! :) I'm pushing them into transformers, pytorch-gemma and collabing with the Gemma team to resolve all the issues :)

The RoPE fix should already be in transformers 4.38.2: https://github.com/huggingface/transformers/pull/29285

My main PR for transformers which fixes most of the issues (some still left): https://github.com/huggingface/transformers/pull/29402

If you want to tinker with the architecture Hugging Face has a FOSS implementation in transformers: https://github.com/huggingface/transformers/blob/main/src/tr...

If you want to reproduce the training pipeline, you couldn't do that even if you wanted to because you don't have access to thousands of A100s.

I am aware that using mixed precision reduces the stability of weight and there will be little consistency but don't expect it to be this much. I have attached the graph of evaluation metrics. If someone can give me some insight into this issue, that would be great.

In transformers, they tried really hard to have a single function or method to deal with both self and cross attention mechanisms, masking, positional and relative encodings, interpolation etc. While it allows a user to use the same function/method for any model, it has led to severe parameter bloat. Just compare the original implementation of llama by FAIR with the implementation by HF to get an idea.

22. Ray | Github | tutorial

Training times for GSM8k are mentioned here: https://github.com/ray-project/ray/tree/master/doc/source/te...

Here's a reference from RLlib that points to docs and an example, and here's one from one of my projects that includes all my own implementations

RLlib

⦿ AWS Glue Data Quality - Automatic data quality rule recommendations based on your data AWS Glue for Ray - Data integration with Ray (ray.io), a popular new open-source compute framework that helps you scale Python workloads

https://ray.io (just dropping the link)

I've wondered whether it's easier to add data analyst stuff to Elixir that Python seems to have, or add features to Python that Erlang (and by extension Elixir) provides out of the box.

By what I can see, if you want multiprocessing on Python in an easier way (let's say running async), you have to use something like ray core[0], then if you want multiple machines you need redis(?). Elixir/Erlang supports this out of the box.

Explorer[1] is an interesting approach, where it uses Rust via Rustler (Elixir library to call Rust code) and uses Polars as its dataframe library. I think Rustler needs to be reworked for this usecase, as it can be slow to return data. I made initial improvements which drastically improves encoding (https://github.com/elixir-nx/explorer/pull/282 and https://github.com/elixir-nx/explorer/pull/286, tldr 20+ seconds down to 3).

[0] https://github.com/ray-project/ray

Adding these for anyone who had a similar question about Ray vs dask 1, 2, 3