Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/EleutherAI/cookbook

Deep learning for dummies. All the practical details and useful utilities that go into working with real models.
https://github.com/EleutherAI/cookbook

Last synced: 14 days ago
JSON representation

Deep learning for dummies. All the practical details and useful utilities that go into working with real models.

Awesome Lists containing this project

README 

        
# The Cookbook

Deep learning for dummies, by Quentin Anthony, Jacob Hatef, Hailey Schoelkopf, and Stella Biderman



All the practical details and utilities that go into working with real models! If you're just getting started, we recommend jumping ahead to [Basics](#basics) for some introductory resources on transformers.



## Table of Contents



- [The Cookbook](#the-cookbook)

  * [Utilities](#utilities)

    + [Calculations](#calculations)

    + [Benchmarks](#benchmarks)

  * [Reading List](#reading-list)

    + [Basics](#basics)

    + [How to do LLM Calculations](#how-to-do-llm-calculations)

    + [Distributed Deep Learning](#distributed-deep-learning)

    + [Best Practices](#best-practices)

    + [Data/Model Directories](#data-and-model-directories)

  * [Minimal Repositories for Educational Purposes](#minimal-repositories-for-educational-purposes)

  * [Contributing](#contributing)



## Utilities



### Calculations



For training/inference calculations (e.g. FLOPs, memory overhead, and parameter count)

- **[calc](./calc/)**



Useful external calculators include



[Cerebras Model Lab](https://www.cerebras.net/model-lab/). User-friendly tool to apply Chinchilla scaling laws.



[Transformer Training and Inference VRAM Estimator](https://vram.asmirnov.xyz/) by Alexander Smirnov. A user-friendly tool to estimate VRAM overhead.



### Benchmarks



Communication benchmarks

- **[benchmarks/communication](./benchmarks/communication)**



Transformer sizing and GEMM benchmarks

- **[benchmarks/sizing](./benchmarks/sizing)**



## Reading List



### Basics



[LLM Visualizations](https://bbycroft.net/llm). Clear LLM visualizations and animations for basic transformer understanding.



[Annotated PyTorch Paper Implementations](https://nn.labml.ai/)



[Jay Alammar's blog](https://jalammar.github.io/blog) contains many blog posts pitched to be accessible to a wide range of backgrounds. We recommend his posts [the Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/), and [the Illustrated GPT-2](https://jalammar.github.io/illustrated-gpt2/) in particular.



[The Annotated Transformer](https://nlp.seas.harvard.edu/annotated-transformer/) by Sasha Rush, Austin Huang, Suraj Subramanian, Jonathan Sum, Khalid Almubarak, and Stella Biderman. A walk through of the seminal paper "Attention is All You Need" along with in-line implementations in PyTorch.



[Transformer Explainer](https://poloclub.github.io/transformer-explainer/) by Aeree Cho, Grace C. Kim, Alexander Karpekov, Alec Helbling, Jay Wang, Seongmin Lee, Benjamin Hoover, and Polo Chau. It runs a gpt-2 model live in your browser, with visualization and customization options.



### How to do LLM Calculations



[Transformers Math 101](https://blog.eleuther.ai/transformer-math/). A blog post from EleutherAI on training/inference memory estimations, parallelism, FLOP calculations, and deep learning datatypes.



[Transformer Inference Arithmetic](https://kipp.ly/transformer-inference-arithmetic/) by Carol Chen. A breakdown on the memory overhead, FLOPs, and latency of transformer inference



[LLM Finetuning Memory Requirements](https://blog.scottlogic.com/2023/11/24/llm-mem.html) by Alex Birch. A practical guide on the memory overhead of finetuning models.



### Distributed Deep Learning



[Everything about Distributed Training and Efficient Finetuning](https://sumanthrh.com/post/distributed-and-efficient-finetuning/) by Sumanth R Hegde. High-level descriptions and links on parallelism and efficient finetuning.



[Efficient Training on Multiple GPUs](https://huggingface.co/docs/transformers/main/en/perf_train_gpu_many) by Hugging Face. Contains a detailed walk-through of model, tensor, and data parallelism along with the ZeRO optimizer.



Papers

- [Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM](https://arxiv.org/abs/2104.04473)

- [Demystifying Parallel and Distributed Deep Learning: An In-Depth Concurrency Analysis](https://arxiv.org/abs/1802.09941)

- [ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://arxiv.org/abs/1910.02054)

- [PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel](https://arxiv.org/abs/2304.11277)

- [PyTorch Distributed: Experiences on Accelerating Data Parallel Training](https://arxiv.org/abs/2006.15704)



### Best Practices



[ML-Engineering Repository](https://github.com/stas00/ml-engineering). Containing community notes and practical details of everything deep learning training led by Stas Bekman



[Common HParam Settings](https://docs.google.com/spreadsheets/d/14vbBbuRMEHoqeuMHkTfw3uiZVmyXNuoSp8s-aHvfvZk/edit?usp=sharing) by Stella Biderman. Records common settings for model training hyperparameters and her current recommendations for training new models.



[The Practitioner's Guide to muP](https://blog.eleuther.ai/mutransfer/) is a joint project by [Cerebras](https://cerebras.ai/) and EleutherAI. Provides a clear picture of muP basics, with an accompanying basic implementation in [nanoGPT](https://github.com/EleutherAI/nanoGPT-mup)



### Data and Model Directories



[Directory of LLMs](https://docs.google.com/spreadsheets/d/1gc6yse74XCwBx028HV_cvdxwXkmXejVjkO-Mz2uwE0k/edit?usp=sharing) by Stella Biderman. Records details of trained LLMs including license, architecture type, and dataset.



[Data Provenance Explorer](https://dataprovenance.org/) A tool for tracing and filtering on data provenance for the most popular open source finetuning data collections.



## Minimal Repositories for Educational Purposes



Large language models are frequently trained using very complex codebases due to the need to optimize things to work at scale and support a wide variety of configurable options. This can make them less useful pedagogical tools, so some people have developed striped-down so-called "Minimal Implementations" that are sufficient for smaller scale work and more pedagogically useful.



GPT Inference

- https://github.com/pytorch-labs/gpt-fast/tree/main



GPT Training

- https://github.com/karpathy/minGPT



Architecture-Specific Examples

- https://github.com/zphang/minimal-gpt-neox-20b

- https://github.com/zphang/minimal-llama

- https://github.com/zphang/minimal-opt



[RWKV](https://www.rwkv.com/)

- https://github.com/Hannibal046/nanoRWKV/tree/main



## Contributing



If you found a bug, typo, or would like to propose an improvement please don't hesitate to open an [Issue](https://github.com/EleutherAI/cookbook/issues) or contribute a [PR](https://github.com/EleutherAI/cookbook/pulls).



## Cite As



If you found this repository helpful, please consider citing it using



```bibtex

@misc{anthony2024cookbook,

    title = {{The EleutherAI Model Training Cookbook}},

    author = {Anthony, Quentin and Hatef, Jacob and Schoelkopf, Hailey and Biderman, Stella},

    howpublished = {GitHub Repo},

    url = {https://github.com/EleutherAI/cookbook},

    year = {2024}

}

```