This repository records EleutherAI's work-in-progress for training large-scale language models on GPUs. Our current framework is based on NVIDIA's Megatron Language Model and has been augmented with techniques from DeepSpeed as well as some novel optimizations.

We aim to make this repo a centralized and accessible place to gather techniques for training large-scale autoregressive language models, and accelerate research into large-scale training. Additionally, we hope to train and open source a 175B parameter GPT-3 replication along the way. Please note, however, that this is a research codebase that is primarily designed for performance over ease of use. We endeavour to make it as easy to use as is feasible, but if there's anything in the readme that is unclear or you think you've found a bug, please open an issue.

If you are interested in contributing, please join our Discord and head to the #gpt-neox channel. We're working with cloud compute provider CoreWeave for training, and hope to release the weights of smaller models as we progress up to 175B parameters.

If you're looking for our TPU codebase, see GPT-Neo.

• Pretrained Models
• Quick Start
• Configuration
• Datasets
  • Preconfigured Datasets
  • Using Custom Data
  • Using and Training Tokenizers
• Training and Finetuning
• Inference
• Evaluation
• Monitoring
  • Weights & Biases
  • Tensorboard
• Administrative Notes
  • Citing GPT-NeoX
  • Licensing
  • Acknowledgements

A 20 billion Parameter autoregressive language model trained on the pile. For technical details about the model, see our paper here.

The configuration file for the model is available here, and is also included in the download links below.

Slim weights - (No optimizer states, for inference or finetuning, 39GB)

To download from the command line to a folder named 20B_checkpoints, use the following command:

wget --cut-dirs=5 -nH -r --no-parent --reject "index.html*" https://mystic.the-eye.eu/public/AI/models/GPT-NeoX-20B/slim_weights/ -P 20B_checkpoints

Full weights - (Including optimizer states, 268GB)

To download from the command line to a folder named 20B_checkpoints, use the following command:

wget --cut-dirs=5 -nH -r --no-parent --reject "index.html*" https://mystic.the-eye.eu/public/AI/models/GPT-NeoX-20B/full_weights/ -P 20B_checkpoints

Alternatively, the models can be downloaded using a BitTorrent client. Torrent files can be downloaded here: slim weights, full_weights.

First make sure you are in an environment with Python 3.8 or later with an appropriate version of PyTorch 1.8 or later installed.

To install the remaining basic dependencies, run:

pip install -r requirements/requirements.txt
python ./megatron/fused_kernels/setup.py install # optional if not using fused kernels

from the repository root.

Warning: Our codebase relies on DeeperSpeed, our fork of the DeepSpeed library with some added changes. We strongly recommend using Anaconda, a virtual machine, or some other form of environment isolation before continuing. Failure to do so may cause other repositories that rely on DeepSpeed to break.

We also provide a Dockerfile if you prefer to run NeoX in a container. To use this option, first build an image named gpt-neox from the repository root directory with docker build -t gpt-neox -f Dockerfile .. We also host pre-built images on Docker Hub at leogao2/gpt-neox.

You can then run a container based on this image. For instance, the below snippet mounts the cloned repository (gpt-neox) directory to /gpt-neox in the container and uses nvidia-docker to make four GPUs (numbers 0-3) accessible to the container. As noted by the NCCL documentation, both --shm-size=1g and --ulimit memlock=-1 are important to prevent Docker from allocating too little shared memory.

nvidia-docker run --rm -it -e NVIDIA_VISIBLE_DEVICES=0,1,2,3 --shm-size=1g --ulimit memlock=-1 --mount type=bind,src=$PWD,dst=/gpt-neox gpt-neox

GPT-NeoX-20B (currently the only pretrained model we provide) is a very large model. The weights alone take up around 40GB in GPU memory and, due to the tensor parallelism scheme as well as the high memory usage, you will need at minimum 2 GPUs with a total of ~45GB of GPU VRAM to run inference, and significantly more for training. Unfortunately the model is not yet possible to use on a single consumer GPU.

GPT-NeoX parameters are defined in a YAML configuration file which is passed to the deepy.py launcher. For more details on the configuration file, see Configuration. The configuration file for GPT-NeoX-20B is at ./configs/20B.yml - but you may need to edit some fields to specify where your model and tokenizer are saved. In the config file edit the following fields:

  "vocab-file": "./20B_checkpoints/20B_tokenizer.json",
  "save": "./20B_checkpoints",
  "load": "./20B_checkpoints",

changing ./20B_checkpoints to the path to the root folder of the downloaded checkpoints. If the checkpoints exist at ./20B_checkpoints you can leave this as is.

Depending on the number of GPUs you're using, you may also need to change the parallelism settings. To run inference on the 20B model on 2 GPUs, change:

   "pipe-parallel-size": 4,

to:

   "pipe-parallel-size": 1,

If you're using 8 GPUs, you can leave this unchanged.

All functionality (inference included), should be launched in parallel using deepy.py, a wrapper around the deepspeed launcher.

We currently offer three main functions:

1. train.py is used for training and finetuning models.
2. evaluate.py is used to evaluate a trained model using the language model evaluation harness.
3. generate.py is used to sample text from a trained model.

and can be launched with:

./deepy.py [script.py] [./path/to/config_1.yml] [./path/to/config_2.yml] ... [./path/to/config_n.yml]

E.G To generate text unconditionally with the GPT-NeoX-20B model, you can use the following:

./deepy.py generate.py ./configs/20B.yml

Or optionally pass in a text file (e.g prompt.txt) to use as the prompt, which should be a plain .txt file with each prompt separated by newline characters, also passing in the path to an output file.

./deepy.py generate.py ./configs/20B.yml -i prompt.txt -o sample_outputs.txt

To reproduce our evaluation numbers on, for example, lambada and PIQA use:

./deepy.py evaluate.py ./configs/20B.yml --eval_tasks lambada piqa

You can add an arbitrary list of evaluation tasks here, for details of all tasks available, see lm-evaluation-harness.

For more details on each entry point, see the Training and finetuning, Inference and Evaluation sections.

GPT-NeoX parameters are defined in a YAML configuration file which is passed to the deepy.py launcher. We have provided some example .yaml files in configs, including one for GPT-NeoX-20B, and example configuration files for other model sizes.

These files are generally complete, but non-optimal. For example, depending on your specific GPU configuration, you may need to change some settings such as pipe-parallel-size, model-parallel-size to increase or decrease the degree of parallelisation, train_micro_batch_size_per_gpu or gradient-accumulation-steps to modify batch size related settings, or the zero_optimization dict to modify how optimizer states are parallelised across workers.

For a more detailed guide to all the features available and how to configure them, see the configuration README, and for documentation of every possible argument, see configs/neox_arguments.md.

Several preconfigured datasets are available, including most components from the Pile, as well as the Pile train set itself, for straightforward tokenization using the prepare_data.py entry point.

E.G, to download and tokenize the Enron emails corpus with the GPT2 Tokenizer, saving them to ./data you can run:

python prepare_data.py -d ./data

or with the GPT-NeoX-20B tokenizer (assuming you have it saved at ./20B_checkpoints/20B_tokenizer.json):

python prepare_data.py -d ./data -t HFTokenizer --vocab-file ./20B_checkpoints/20B_tokenizer.json

The tokenized data will be saved out to two files at [data-dir]/[dataset-name]/[dataset-name]_text_document.bin & [data-dir]/[dataset-name]/[dataset-name]_text_document.bin. You will need to add the prefix that both these files share to your training configuration file under the data-path field. E.G:

  "data-path": "./data/enron/enron_text_document",

To prepare your own dataset for training with custom data, format it as one large jsonl-formatted file with each item in the list of dictionaries being a separate document. The document text should be grouped under one JSON key, i.e "text". Any auxiliary data stored in other fields will not be

Next make sure to download the GPT2 tokenizer vocab, and merge files from the following links:

• Vocab: https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-vocab.json
• Merge: https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-merges.txt

Or use the 20B tokenizer (for which only a single Vocab file is needed):

• Vocab: https://mystic.the-eye.eu/public/AI/models/GPT-NeoX-20B/slim_weights/20B_tokenizer.json

(alternatively, you can provide any tokenizer file that can be loaded by Huggingface's tokenizers library with the Tokenizer.from_pretrained() command)

You can now pretokenize your data using tools/preprocess_data.py, the arguments for which are detailed below:

usage: preprocess_data.py [-h] --input INPUT [--jsonl-keys JSONL_KEYS [JSONL_KEYS ...]] [--num-docs NUM_DOCS] --tokenizer-type {HFGPT2Tokenizer,HFTokenizer,GPT2BPETokenizer,CharLevelTokenizer} [--vocab-file VOCAB_FILE] [--merge-file MERGE_FILE] [--append-eod] [--ftfy] --output-prefix OUTPUT_PREFIX
                          [--dataset-impl {lazy,cached,mmap}] [--workers WORKERS] [--log-interval LOG_INTERVAL]

optional arguments:
  -h, --help            show this help message and exit

input data:
  --input INPUT         Path to input jsonl files or lmd archive(s) - if using multiple archives, put them in a comma separated list
  --jsonl-keys JSONL_KEYS [JSONL_KEYS ...]
                        space separate listed of keys to extract from jsonl. Defa
  --num-docs NUM_DOCS   Optional: Number of documents in the input data (if known) for an accurate progress bar.

tokenizer:
  --tokenizer-type {HFGPT2Tokenizer,HFTokenizer,GPT2BPETokenizer,CharLevelTokenizer}
                        What type of tokenizer to use.
  --vocab-file VOCAB_FILE
                        Path to the vocab file
  --merge-file MERGE_FILE
                        Path to the BPE merge file (if necessary).
  --append-eod          Append an <eod> token to the end of a document.
  --ftfy                Use ftfy to clean text

output data:
  --output-prefix OUTPUT_PREFIX
                        Path to binary output file without suffix
  --dataset-impl {lazy,cached,mmap}
                        Dataset implementation to use. Default: mmap

runtime:
  --workers WORKERS     Number of worker processes to launch
  --log-interval LOG_INTERVAL
                        Interval between progress updates

For example:

python tools/preprocess_data.py \
            --input ./data/mydataset.jsonl.zst \
            --output-prefix ./data/mydataset \
            --vocab ./data/gpt2-vocab.json \
            --merge-file gpt2-merges.txt \
            --dataset-impl mmap \
            --tokenizer-type GPT2BPETokenizer \
            --append-eod

You would then run training with the following settings added to your configuration file:

  "data-path": "data/mydataset/mydataset",

Training is launched using deepy.py, a wrapper around DeepSpeed's launcher, which launches the same script in parallel across many GPUs / nodes.

The general usage pattern is:

python ./deepy.py train.py [path/to/config1.yml] [path/to/config2.yml] ...

You can pass in an arbritrary number of configs which will all be merged at runtime.

You can also optionally pass in a config prefix, which will assume all your configs are in the same folder and append that prefix to their path.

E.G:

python ./deepy.py train.py -d configs small.yml local_setup.yml

This will deploy the train.py script on all nodes with one process per GPU. The worker nodes and number of GPUs are specified in the /job/hostfile file (see parameter documentation), or can simply be passed in as the num_gpus arg if running on a single node setup.

Although this is not strictly necessary, we find it useful to define the model parameters in one config file (e.g configs/small.yml) and the data path parameters in another (e.g configs/local_setup.yml).

We support three types of generation from a pretrained model:

1. Unconditional generation
2. Conditional generation based on an input read from a file
3. Interactive generation, which allows for multiple rounds of back-and-forth between a user and the language model via a command line interface

All three types of text generation can be launched via python ./deepy.py generate.py -d configs small.yml local_setup.yml text_generation.yml with the appropriate values set in configs/text_generation.yml.

GPT-NeoX supports evaluation on downstream tasks through the language model evaluation harness.

To evaluate a trained model on the evaluation harness, simply run:

python ./deepy.py evaluate.py -d configs your_configs.yml --eval_tasks task1 task2 ... taskn

where --eval_tasks is a list of evaluation tasks followed by spaces, e.g --eval_tasks lambada hellaswag piqa sciq. For details of all tasks available, refer to the lm-evaluation-harness repo.

In addition to storing logs locally, we provide built-in support for two popular experiment monitoring frameworks: Weights & Biases and TensorBoard

EleutherAI is currently using Weights & Biases to record our experiments. If you are logged into Weights & Biases on your machine—you can do this by executing wandb login—your runs will automatically be recorded. There are two optional fields associated with Weights & Biases: wandb_group allows you to name the run group and wandb_team allows you to assign your runs to an organization or team account.

We also support using TensorBoard via the tensorboard-dir field. Dependencies required for TensorBoard monitoring can be found in and installed from ./requirements/requirements-tensorboard.txt.

If you have found GPT-NeoX helpful in your work, you can cite this repository as

@software{gpt-neox,
  author = {Andonian, Alex and Anthony, Quentin and Biderman, Stella and Black, Sid and Gali, Preetham and Gao, Leo and Hallahan, Eric and Levy-Kramer, Josh and Leahy, Connor and Nestler, Lucas and Parker, Kip and Pieler, Michael and Purohit, Shivanshu and Songz, Tri and Wang, Phil and Weinbach, Samuel},
  title = {{GPT-NeoX}: Large Scale Autoregressive Language Modeling in PyTorch},
  url = {https://github.com/eleutherai/gpt-neox},
  year = {2021}
}

In the above BibTex entry, names are in alphabetical order, and the year corresponds to the project's open-source release.

This repository hosts code that is part of EleutherAI's GPT-NeoX project. Copyright © 2021, EleutherAI contributors (in alphabetical order): Alex Andonian, Quentin Anthony, Stella Biderman, Sid Black, Preetham Gali, Leo Gao, Eric Hallahan, Josh Levy-Kramer, Connor Leahy, Lucas Nestler, Kip Parker, Michael Pieler, Shivanshu Purohit, Tri Songz, Phil Wang, Samuel Weinbach. Licensed under the Apache License:

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    https://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

This repository is based off code written by NVIDIA that is licensed under the Apache License, Version 2.0. In accordance with the Apache License, all files that are modifications of code originally written by NVIDIA maintain a NVIDIA copyright header. All files that do not contain such a header are original to EleutherAI contributors. When the NVIDIA code has been modified from its original version, that fact is noted in the copyright header. All derivative works of this repository must preserve these headers under the terms of the Apache License.

For full terms, see the LICENSE file. If you have any questions, comments, or concerns about licensing please email us at [email protected].

We run our experiments on a Kubernetes cluster generously provided by CoreWeave.