Pluto.jl VS transformers

I thought that notebook based development and package based development were diametrically opposed in the past, but Pluto.jl notebooks have changed my mind about this.

A Pluto.jl notebook is a human readable Julia source file. The Pluto.jl package is itself developed via Pluto.jl notebooks.

https://github.com/fonsp/Pluto.jl

Also, the VSCode Julia plugin tooling has really expanded in functionality and usability for me in the past year. The integrated debugging took some work to setup, but is fast enough to drop into a local frame.

https://code.visualstudio.com/docs/languages/julia

Julia is the first language I have achieved full life cycle integration between exploratory code to sharable package. It even runs quite well on my Android. 2023 is the first year I was able to solve a differential equation or render a 3D surface from a calculated mesh with the hardware in my pocket.

Re Julia there is also pluto.jl that is another notebook-like environment for julia. It's been a few years since I played with it but it looked cool, for example it handles state differently so you don't get into the same messes as with ipython notebooks. https://plutojl.org/

For this, Notebooks are often used. Julia offers a uniquely nice and interactive Pluto notebook for the web https://github.com/fonsp/Pluto.jl

I believe this is what Pluto sets out to do for Julia.

I used it as part of the “Computational Thinking” with Julia course a year or two back. Even then the beta software was very good and some of the demos the Pluto dev showed were nothing short of amazing

https://plutojl.org/

I think what you want is Pluto!

I like how Pluto.jl handles this:

> Pluto offers an environment where changed code takes effect instantly and where deleted code leaves no trace. Unlike Jupyter or Matlab, there is no mutable workspace, but rather, an important guarantee:

> At any instant, the program state is completely described by the code you see.

[1] https://github.com/fonsp/Pluto.jl

Is t5x an encoder/decoder architecture?

Some more general options.

The Flax ecosystem

https://github.com/google/flax?tab=readme-ov-file

or dm-haiku

https://github.com/google-deepmind/dm-haiku

were some of the best developed communities in the Jax AI field

Perhaps the “trax” repo? https://github.com/google/trax

Some HF examples https://github.com/huggingface/transformers/tree/main/exampl...

Sadly it seems much of the work is proprietary these days, but one example could be Grok-1, if you customize the details. https://github.com/xai-org/grok-1/blob/main/run.py

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.