This repository records EleutherAI's library 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. This library is in widespread use in academic, industry, and government labs, including by researchers at Oak Ridge National Lab, CarperAI, Stability AI, Together.ai, Korea University, Carnegie Mellon University, and the University of Tokyo among others. Uniquely among similar libraries GPT-NeoX supports a wide variety of systems and hardwares, including launching via Slurm, MPI, and the IBM Job Step Manager, and has been run at scale on AWS, CoreWeave, ORNL Summit, ORNL Frontier, LUMI, and others.
If you are not looking to train models with billions of parameters from scratch, this is likely the wrong library to use. For generic inference needs, we recommend you use the Hugging Face transformers
library instead which supports GPT-NeoX models.
GPT-NeoX leverages many of the same features and technologies as the popular Megatron-DeepSpeed library but with substantially increased usability and novel optimizations. Major features include:
- Distributed training with ZeRO and 3D parallelism
- A wide variety of systems and hardwares, including launching via Slurm, MPI, and the IBM Job Step Manager, and has been run at scale on AWS, CoreWeave, ORNL Summit, ORNL Frontier, LUMI, and others.
- Cutting edge architectural innovations including rotary and alibi positional embeddings, parallel feedforward attention layers, and flash attention.
- Predefined configurations for popular architectures including Pythia, PaLM, Falcon, and LLaMA 1 & 2
- Curriculum Learning
- Easy connections with the open source ecosystem, including Hugging Face's tokenizers and transformers libraries, logging via WandB, and evaluation via our Language Model Evaluation Harness.
[8/10/2023] We now support checkpointing with AWS S3! Activate with the s3_path
config option (for more detail, see the PR)
[9/20/2023] As of EleutherAI#1035, we have deprecated Flash Attention 0.x and 1.x, and migrated support to Flash Attention 2.x. We don't believe this will cause problems, but if you have a specific use-case that requires old flash support using the latest GPT-NeoX, please raise an issue.
[8/10/2023] We have experimental support for LLaMA 2 and Flash Attention v2 supported in our math-lm project that will be upstreamed later this month.
[5/17/2023] After fixing some miscellaneous bugs we now fully support bf16.
[4/11/2023] We have upgraded our Flash Attention implementation to now support Alibi positional embeddings.
[3/9/2023] We have released GPT-NeoX 2.0.0, an upgraded version built on the latest DeepSpeed which will be regularly synced with going forward.
Prior to 3/9/2023, GPT-NeoX relied on DeeperSpeed, which was based on an old version of DeepSpeed (0.3.15). In order to migrate to the latest upstream DeepSpeed version while allowing users to access the old versions of GPT-NeoX and DeeperSpeed, we have introduced two versioned releases for both libraries:
- Version 2.0 of GPT-NeoX and DeeperSpeed are the latest versions built on the latest DeepSpeed, and will be maintained going forward.
- Version 1.0 of GPT-NeoX and DeeperSpeed maintain snapshots of the old stable versions that GPT-NeoX-20B and the Pythia Suite were trained on.
- GPT-NeoX
- Contents
- Quick Start
- Configuration
- Datasets
- Training and Finetuning
- Inference
- Evaluation
- Exporting to Hugging Face
- Monitoring
- Running on multi-node
- Adoption and Publications
- Administrative Notes
First make sure you are in an environment with Python 3.8 with an appropriate version of PyTorch 1.8 or later installed. Note: Some of the libraries that GPT-NeoX depends on have not been updated to be compatible with Python 3.10 . Python 3.9 appears to work, but this codebase has been developed and tested for Python 3.8.
To install the remaining basic dependencies, run:
pip install -r requirements/requirements.txt
pip install -r requirements/requirements-wandb.txt # optional, if logging using WandB
pip install -r requirements/requirements-tensorboard.txt # optional, if logging via tensorboard
python ./megatron/fused_kernels/setup.py install # optional, if 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.
To use Flash-Attention, install the additional dependencies in ./requirements/requirements-flashattention.txt
and set the attention type in your configuration accordingly (see configs). This can provide significant speed-ups over regular attention on certain GPU architectures, including Ampere GPUs (such as A100s); see the repository for more details.
NeoX and Deep(er)Speed support training on multiple different nodes and you have the option of using a variety of different launchers to orchestrate multi-node jobs.
In general there needs to be a "hostfile" somewhere accessible with the format:
node1_ip slots=8
node2_ip slots=8
where the first column contains the IP address for each node in your setup and the number of slots is the number of GPUs that node has access to. In your config you must pass in the path to the hostfile with "hostfile": "/path/to/hostfile"
. Alternatively the path to the hostfile can be in the environment variable DLTS_HOSTFILE
.
pdsh
is the default launcher, and if you're using pdsh
then all you must do (besides ensuring that pdsh is installed in your environment) is set {"launcher": "pdsh"}
in your config files.
If using MPI then you must specify the MPI library (DeepSpeed/GPT-NeoX currently supports mvapich
, openmpi
, mpich
, and impi
, though openmpi
is the most commonly used and tested) as well as pass the deepspeed_mpi
flag in your config file:
{
"launcher": "openmpi",
"deepspeed_mpi": true
}
With your environment properly set up and the correct configuration files you can use deepy.py
like a normal python script and start (for example) a training job with:
python3 deepy.py train.py /path/to/configs/my_model.yml
Using Slurm can be slightly more involved. Like with MPI, you must add the following to your config:
{
"launcher": "slurm",
"deepspeed_slurm": true
}
If you do not have ssh access to the compute nodes in your Slurm cluster you need to add {"no_ssh_check": true}
There are many cases where the above default launching options are not sufficient
- Many clusters have their own unique job scheduler or specific MPI/Slurm arguments necessary for launching jobs such as Summit JSRun or LLNL Flux
- While the above Slurm/MPI/pdsh default options are enough for most job runs, advanced users may want to add arguments for optimization or debugging purposes
In these cases, you will need to modify the DeepSpeed multinode runner utility to support your usecase. Broadly, these enhancements fall under two categories:
In this case, you must add a new multinode runner class to deepspeed/launcher/multinode_runner.py
and expose it as a configuration option in GPT-NeoX. Examples on how we did this for Summit JSRun are in this DeeperSpeed commit and this GPT-NeoX commit, respectively.
We have encountered many cases where we wish to modify the MPI/Slurm run command for an optimization or to debug (e.g. to modify the Slurm srun CPU binding or to tag MPI logs with the rank). In this case, you must modify the multinode runner class' run command under its get_cmd
method (e.g. mpirun_cmd for OpenMPI). Examples on how we did this to provide optimized and rank-tagged run commands using Slurm and OpenMPI for the Stability cluster are in this DeeperSpeed branch
In general you will not be able to have a single fixed hostfile, so you need to have a script to generate one dynamically when your job starts. An example script to dynamically generate a hostfile using Slurm and 8 GPUs per node is:
#!/bin/bash
GPUS_PER_NODE=8
mkdir -p /sample/path/to/hostfiles
# need to add the current slurm jobid to hostfile name so that we don't add to previous hostfile
hostfile=/sample/path/to/hostfiles/hosts_$SLURM_JOBID
# be extra sure we aren't appending to a previous hostfile
rm $hostfile &> /dev/null
# loop over the node names
for i in `scontrol show hostnames $SLURM_NODELIST`
do
# add a line to the hostfile
echo $i slots=$GPUS_PER_NODE >>$hostfile
done
$SLURM_JOBID
and $SLURM_NODELIST
being environment variables Slurm will create for you. See the sbatch documentation for a full list of available Slurm environment variables set at job creation time.
Then you can create an sbatch script from which to kick off your GPT-NeoX job. A bare-bones sbatch script on a Slurm-based cluster with 8 GPUs per node would look like this:
#!/bin/bash
#SBATCH --job-name="neox"
#SBATCH --partition=your-partition
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=8
#SBATCH --gres=gpu:8
# Some potentially useful distributed environment variables
export HOSTNAMES=`scontrol show hostnames "$SLURM_JOB_NODELIST"`
export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
export MASTER_PORT=12802
export COUNT_NODE=`scontrol show hostnames "$SLURM_JOB_NODELIST" | wc -l`
# Your hostfile creation script from above
./write_hostfile.sh
# Tell DeepSpeed where to find our generated hostfile via DLTS_HOSTFILE
export DLTS_HOSTFILE=/sample/path/to/hostfiles/hosts_$SLURM_JOBID
# Launch training
python3 deepy.py train.py /sample/path/to/your/configs/my_model.yml
You can then kick off a training run with sbatch my_sbatch_script.sh
We also provide a Dockerfile and docker-compose configuration if you prefer to run NeoX in a container.
Requirements to run the container are to have appropriate GPU drivers, an up-to-date installation of Docker, and nvidia-container-toolkit installed. To test if your installation is good you can use their "sample workload", which is:
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
Provided that will run, you need to export NEOX_DATA_PATH and NEOX_CHECKPOINT_PATH in your environment to specify your data directory and directory for storing and loading checkpoints:
export NEOX_DATA_PATH=/mnt/sda/data/enwiki8 #or wherever your data is stored on your system
export NEOX_CHECKPOINT_PATH=/mnt/sda/checkpoints
And then, from the gpt-neox directory, you can build the image and run a shell in a container with
docker compose run gpt-neox bash
After the build, you should be able to do this:
mchorse@537851ed67de:~$ echo $(pwd)
/home/mchorse
mchorse@537851ed67de:~$ ls -al
total 48
drwxr-xr-x 1 mchorse mchorse 4096 Jan 8 05:33 .
drwxr-xr-x 1 root root 4096 Jan 8 04:09 ..
-rw-r--r-- 1 mchorse mchorse 220 Feb 25 2020 .bash_logout
-rw-r--r-- 1 mchorse mchorse 3972 Jan 8 04:09 .bashrc
drwxr-xr-x 4 mchorse mchorse 4096 Jan 8 05:35 .cache
drwx------ 3 mchorse mchorse 4096 Jan 8 05:33 .nv
-rw-r--r-- 1 mchorse mchorse 807 Feb 25 2020 .profile
drwxr-xr-x 2 root root 4096 Jan 8 04:09 .ssh
drwxrwxr-x 8 mchorse mchorse 4096 Jan 8 05:35 chk
drwxrwxrwx 6 root root 4096 Jan 7 17:02 data
drwxr-xr-x 11 mchorse mchorse 4096 Jan 8 03:52 gpt-neox
For a long-running job, you should run
docker compose up -d
to run the container in detached mode, and then, in a separate terminal session, run
docker compose exec gpt-neox bash
You can then run any job you want from inside the container.
Concerns when running for a long time or in detached mode include
- You will have to terminate the container manually when you are no longer using it
- If you want processes to continue running when your shell session ends, you will need to background them.
- If you then want logging, you will have to make sure to pipe logs to disk or set up wandb.
If you prefer to run the prebuilt container image from dockerhub, you can run the docker compose commands with -f docker-compose-dockerhub.yml
instead, e.g.,
docker compose run -f docker-compose-dockerhub.yml gpt-neox bash
All functionality should be launched using deepy.py
, a wrapper around the deepspeed
launcher.
We currently offer three main functions:
train.py
is used for training and finetuning models.eval.py
is used to evaluate a trained model using the language model evaluation harness.generate.py
is used to sample text from a trained model.
which can be launched with:
./deepy.py [script.py] [./path/to/config_1.yml] [./path/to/config_2.yml] ... [./path/to/config_n.yml]
For example, to launch training you can run
./deepy.py train.py ./configs/20B.yml ./configs/local_cluster.yml
For more details on each entry point, see the Training and Finetuning, Inference and Evaluation respectively.
GPT-NeoX parameters are defined in a YAML configuration file which is passed to the deepy.py launcher. We have provided some example .yml files in configs, showing a diverse array of features and 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 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 enwik8 dataset with the GPT2 Tokenizer, saving them to ./data
you can run:
python prepare_data.py -d ./data
or a single shard of the pile (pile_subset
) 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 pile_subset
The tokenized data will be saved out to two files: [data-dir]/[dataset-name]/[dataset-name]_text_document.bin
and [data-dir]/[dataset-name]/[dataset-name]_text_document.idx
. 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/enwik8/enwik8_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 used.
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):
(alternatively, you can provide any tokenizer file that can be loaded by Hugging Face's tokenizers library with the Tokenizer.from_pretrained()
command)
You can now pretokenize your data using tools/datasets/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/datasets/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 arbitrary 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 125M.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/125M.yml
) and the data path parameters in another (e.g configs/local_setup.yml
).
GPT-NeoX-20B is a 20 billion parameter autoregressive language model trained on the Pile. Technical details about GPT-NeoX-20B can be found in the associated paper. The configuration file for this model is both available at ./configs/20B.yml
and 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://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://the-eye.eu/public/AI/models/GPT-NeoX-20B/full_weights/ -P 20B_checkpoints
Weights can be alternatively be downloaded using a BitTorrent client. Torrent files can be downloaded here: slim weights, full weights.
We additionally have 150 checkpoints saved throughout training, one every 1,000 steps. We are working on figuring out how to best serve these at scale, but in the meanwhile people interested in working with the partially trained checkpoints can email us at [email protected] to arrange access.
The Pythia Scaling Suite is a suite of models ranging from 70M parameters to 12B parameters trained on the Pile intended to promote research on interpretability and training dynamics of large language models. Further details about the project and links to the models can be found in the in the paper and on the project's GitHub.
The Polyglot Project is an effort to train powerful non-English pretrained language models to promote the accessibility of this technology to researchers outside the dominant powerhouses of machine learning. EleutherAI has trained and released 1.3B, 3.8B, and 5.8B parameter Korean language models, the largest of which outpreforms all other publicly available language models on Korean language tasks. Further details about the project and links to the models can be found here.
For most uses we recommend deploying models trained using the GPT-NeoX library via the Hugging Face Transformers library which is better optimized for inference.
We support three types of generation from a pretrained model:
- Unconditional generation
- Conditional generation based on an input read from a file
- 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 125M.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 eval.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.
GPT-NeoX is optimized heavily for training only, and GPT-NeoX model checkpoints are not compatible out of the box with other deep learning libraries. To make models easily loadable and shareable with end users, and for further exporting to various other frameworks, GPT-NeoX supports checkpoint conversion to the Hugging Face Transformers GPTNeoXModel format.
To convert a NeoX checkpoint (with pipeline-parallel-size>=1) to Hugging Face-loadable format, run:
python ./tools/ckpts/convert_module_to_hf.py --input_dir /path/to/model/global_stepXXX --config_file your_config.yml --output_dir hf_model/save/location
To convert a sequential model to Hugging Face format, run:
python ./tools/ckpts/convert_sequential_to_hf.py --input_dir /path/to/model/global_stepXXX --config_file your_config.yml --output_dir hf_model/save/location
(Note: this script should be used for v2.0 checkpoints saved on a v2.0 commit prior to EleutherAI#866 and which used pipe-parallel-size=1
. Using pipe-parallel-size=0
will also save models in this format.)
Then to upload a model to the Hugging Face Hub, run:
huggingface-cli login
python ./tools/ckpts/upload.py
and input the requested information, including HF hub user token.
Note, however, that this compatibility is not one-to-one, and only certain configurations from GPT-NeoX are supported in the Hugging Face GPTNeoXModel class. Advanced features such as alternative positional embeddings may require new Transformers modeling code and new conversion script tweaks.
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 need to supply a hostfile for use with the MPI-based DeepSpeed launcher, you can set the environment variable DLTS_HOSTFILE
to point to the hostfile.
The GPT-NeoX library was been widely adopted by academic and industry researchers and ported on to many HPC systems.
If you have found this library useful in your research, please reach out and let us know! We would love to add you to our lists.
EleutherAI and our collaborators have used it in the following publications:
- Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, Shivanshu Purohit, Laria Reynolds, Jon Tow, Ben Wang, and Samuel Weinbach. "GPT-NeoX-20B: An Open-Source Autoregressive Language Model." In Proceedings of the ACL Workshop on Challenges & Perspectives in Creating Large Language Models, 2022.
- Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, Oskar van der Wal. "Pythia: A suite for analyzing large language models across training and scaling." In International Conference on Machine Learning, pp. 2397-2430. PMLR, 2023.
- Zhangir Azerbayev, Bartosz Piotrowski, Hailey Schoelkopf, Edward W. Ayers, Dragomir Radev, and Jeremy Avigad. "Proofnet: Autoformalizing and formally proving undergraduate-level mathematics. arXiv preprint arXiv:2302.12433, 2023.
- Stella Biderman, USVSN Sai Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Anthony, Shivanshu Purohit, and Edward Raff. "Emergent and predictable memorization in large language models." In Neural Information Processing Systems, 2023.
- Hyunwoong Ko, Kichang Yang, Minho Ryu, Taekyoon Choi, Seungmu Yang, and Sungho Park. "A Technical Report for Polyglot-Ko: Open-Source Large-Scale Korean Language Models." arXiv preprint arXiv:2306.02254, 2023.
- Kshitij Gupta, Benjamin Thérien, Adam Ibrahim, Mats Leon Richter, Quentin Anthony, Eugene Belilovsky, Irina Rish, and Timothée Lesort. "Continual Pre-Training of Large Language Models: How to re-warm your model?" In Workshop on Efficient Systems for Foundation Models @ ICML, 2023.
- Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q Jiang, Jia Deng, Stella Biderman, and Sean Welleck. "Llemma: An open language model for mathematics" In Math-AI Workshop @ NeurIPS, 2023.
- Alexander Havrilla, Maksym Zhuravinskyi, Duy Phung, Aman Tiwari, Jonathan Tow, Stella Biderman, Quentin Anthony, and Louis Castricato. "trlX: A Framework for Large Scale Reinforcement Learning from Human Feedback." In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023.
The following publications by other research groups use this library:
- Ta-Chung Chi, Ting-Han Fan, Peter J. Ramadge, and Alexander Rudnicky. "KERPLE: Kernelized Relative Positional Embedding for Length Extrapolation." In Advances in Neural Information Processing Systems 35 (2022).
- Sameera Horawalavithana, Ellyn Ayton, Shivam Sharma, Scott Howland, Megha Subramanian, Scott Vasquez, Robin Cosbey, Maria Glenski, and Svitlana Volkova. "Foundation Models of Scientific Knowledge for Chemistry: Opportunities, Challenges and Lessons Learned." In Proceedings of the ACL Workshop on Challenges & Perspectives in Creating Large Language Models, 2022.
- Sophia Kolak, Ruben Martins, Claire Le Goues, and Vincent J. Hellendoorn. "Patch Generation with Language Models: Feasibility and Scaling Behavior"." In Proceedings of the Deep Learning for Code Workshop at ICLR, 2022.
- Frank F. Xu, Uri Alon, Graham Neubig, and Vincent J. Hellendoorn. "A Systematic Evaluation of Large Language Models of Code." In Proceedings of the ICLR Workshop on Deep Learning For Code, 2022.
- Eghbal A. Hosseini, Martin A. Schrimpf, Yian Zhang, Samuel Bowman, Noga Zaslavsky, and Evelina Fedorenko. "Artificial neural network language models align neurally and behaviorally with humans even after a developmentally realistic amount of training." BioRxiv, 2022.
- Byung-Doh Oh and William Schuler. "Transformer-Based LM Surprisal Predicts Human Reading Times Best with About Two Billion Training Tokens." In Findings of the Association for Computational Linguistics, 2023.
- Ta-Chung Chi, Ting-Han Fan, Alexander Rudnicky, and Peter Ramadge. "Dissecting Transformer Length Extrapolation via the Lens of Receptive Field Analysis." In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 13522-13537, 2023.
- Ta-Chung Chi, Ting-Han Fan, Li-Wei Chen, Alexander Rudnicky, and Peter Ramadge. "Latent Positional Information is in the Self-Attention Variance of Transformer Language Models Without Positional Embeddings." In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 13522-13537 (2023).
- Xidong Feng, Yicheng Luo, Ziyan Wang, Hongrui Tang, Mengyue Yang, Kun Shao, David Mguni, Yali Du, and Jun Wang. "ChessGPT: Bridging Policy Learning and Language Modeling." arXiv preprint arXiv:2306.09200, 2023.
- Orion Walker Dollar, Sameera Horawalavithana, Scott Vasquez, W. James Pfaendtner, and Svitlana Volkova. "MolJET: Multimodal Joint Embedding Transformer for Conditional de novo Molecular Design and Multi-Property Optimization." preprint under review, 2023.
- Jean Kaddour and Qi Liu. "Text Data Augmentation in Low-Resource Settings via Fine-Tuning of Large Language Models." arXiv:2310.01119, 2023.
- Alon Albalak, Liangming Pan, Colin Raffel, and William Yang Wang. "Efficient Online Data Mixing For Language Model Pre-Training." In NeurIPS Workshop on R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models, 2023.
- Eghbal A. Hosseini and Evelina Fedorenko. "Large language models implicitly learn to straighten neural sentence trajectories to construct a predictive representation of natural language." bioRxiv, 2023.
- Junqi Yin, Sajal Dash, Feiyi Wang, and Mallikarjun Shankar. "FORGE: Pre-Training Open Foundation Models for Science. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 1-13, 2023.
- Jean Kaddour and Qi Liu. "Text Data Augmentation in Low-Resource Settings via Fine-Tuning of Large Language Models." arXiv preprint arXiv:2310.01119, 2023.
- Peng Di, Jianguo Li, Hang Yu, Wei Jiang, Wenting Cai, Yang Cao, Chaoyu Chen, Dajun Chen, Hongwei Chen, Liang Chen, Gang Fan, Jie Gong, Zi Gong, Wen Hu, Tingting Guo, Zhichao Lei, Ting Li, Zheng Li, Ming Liang, Cong Liao, Bingchang Liu, Jiachen Liu, Zhiwei Liu, Shaojun Lu, Min Shen, Guangpei Wang, Huan Wang, Zhi Wang, Zhaogui Xu, Jiawei Yang, Qing Ye, Gehao Zhang, Yu Zhang, Zelin Zhao, Xunjin Zheng, Hailian Zhou, Lifu Zhu, and Xianying Zhu. "CodeFuse-13B: A Pretrained Multi-lingual Code Large Language Model." arXiv preprint arXiv:2310.06266, 2023.
- Nikitha Rao, Kush Jain, Uri Alon, Claire Le Goues, and Vincent J Hellendoorn. "CAT-LM Training Language Models on Aligned Code And Tests." 38th IEEE/ACM International Conference on Automated Software Engineering (ASE), pp. 409-420. IEEE, 2023.
The following models were trained using this library:
- EleutherAI's GPT-NeoX-20B, Pythia (70M through 13B), and LLeMMA (34B)
- CarperAI's FIM-NeoX-1.3B
- StabilityAI's StableLM (3B and 7B)
- Together.ai's RedPajama-INCITE (3B and 7B)
- Carnegie Mellon University's proofGPT (1.3B and 6.7B)
- Dampish's StellarX (2.8B and 4B)
- Oak Ridge National Lab's FORGE (26B)
- EleutherAI's Polyglot-Ko (1.3B through 12.8B) (Korean)
- Korea University's KULLM-Polyglot (5.8B and 12.8B) (Korean)
- Stability AI's Japanese Stable LM (7B)
- LearnItAnyway's LLaVA-Polyglot-Ko (1.3B) (Korean)
- Rinna Co.'s japanese-gpt-neox-3.6b (Japanese) and bilingual-gpt-neox-4b (English / Japanese)
- CyberAgent's Open-CLM (125M through 7B) (Japanese)
- The Hungarian Research Centre for Linguistics's PULI GPTrio (6.7B) (Hungarian / English / Chinese)
- The University of Tokyo's weblab-10b and weblab-10b-instruct (Japanese)
- nolando.ai's Hi-NOLIN (9B) (English, Hindi)
- Carnegie Mellon University's PolyCoder (160M through 2.7B) and CAT-LM (2.7B)
- StabilityAI's StableCode (1.3B) and StableCode-Completion-Alpha (3B)
- CodeFuse AI's CodeFuse (13B)
- University College London's ChessGPT-3B
- Gretel's Text-to-Table (3B)
If you have found the GPT-NeoX library helpful in your work, you can cite this repository as
@software{gpt-neox-library,
title = {{GPT-NeoX: Large Scale Autoregressive Language Modeling in PyTorch}},
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 Phang, Jason and Purohit, Shivanshu and Schoelkopf, Hailey and Stander, Dashiell and Songz, Tri and Tigges, Curt and Thérien, Benjamin and Wang, Phil and Weinbach, Samuel},
url = {https://www.github.com/eleutherai/gpt-neox},
doi = {10.5281/zenodo.5879544},
month = {9},
year = {2023},
version = {2.0.0},
}
To cite the 20 billion parameter model named GPT-NeoX-20B
, please use
@inproceedings{gpt-neox-20b,
title={{GPT-NeoX-20B}: An Open-Source Autoregressive Language Model},
author={Black, Sid and Biderman, Stella and Hallahan, Eric and Anthony, Quentin and Gao, Leo and Golding, Laurence and He, Horace and Leahy, Connor and McDonell, Kyle and Phang, Jason and Pieler, Michael and Prashanth, USVSN Sai and Purohit, Shivanshu and Reynolds, Laria and Tow, Jonathan and Wang, Ben and Weinbach, Samuel},
booktitle={Proceedings of the ACL Workshop on Challenges \& Perspectives in Creating Large Language Models},
url={https://arxiv.org/abs/2204.06745},
year={2022}
}
GPT-NeoX is built by the open-source AI community, and relies on our amazing contributors! Please see our contributing guide for more details on our CLA, code formatting, testing, etc.
This repository hosts code that is part of EleutherAI's GPT-NeoX project. Copyright (c) 2024, EleutherAI. 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
http://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 the exclusive copyright of EleutherAI. 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.
This repository also contains code written by a number of other authors. Such contributions are marked and the relevant licensing is included where appropriate.
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 provided by CoreWeave and a Slurm cluster provided by Stability AI. We are thankful to the DeepSpeed team for their advice and consultation.