This repository contains code for training, finetuning, evaluating, and deploying LLMs for inference with Composer and the MosaicML platform. Designed to be easy-to-use, efficient and flexible, this codebase enables rapid experimentation with the latest techniques.

You'll find in this repo:

• llmfoundry/ - source code for models, datasets, callbacks, utilities, etc.
• scripts/ - scripts to run LLM workloads
  • data_prep/ - convert text data from original sources to StreamingDataset format
  • train/ - train or finetune HuggingFace and MPT models from 125M - 70B parameters
    • train/benchmarking - profile training throughput and MFU
  • inference/ - convert models to HuggingFace or ONNX format, and generate responses
    • inference/benchmarking - profile inference latency and throughput
  • eval/ - evaluate LLMs on academic (or custom) in-context-learning tasks
• mcli/ - launch any of these workloads using MCLI and the MosaicML platform
• TUTORIAL.md - a deeper dive into the repo, example workflows, and FAQs

Mosaic Pretrained Transformers (MPT) are GPT-style models with some special features -- Flash Attention for efficiency, ALiBi for context length extrapolation, and stability improvements to mitigate loss spikes. As part of MosaicML's Foundation series, we have open-sourced several MPT models:

To try out these models locally, follow the instructions in scripts/inference/README.md to prompt HF models using our hf_generate.py or hf_chat.py scripts.

We've been overwhelmed by all the amazing work the community has put into MPT! Here we provide a few links to some of them:

• ReplitLM: replit-code-v1-3b is a 2.7B Causal Language Model focused on Code Completion. The model has been trained on a subset of the Stack Dedup v1.2 dataset covering 20 languages such as Java, Python, and C++
• LLaVa-MPT: Visual instruction tuning to get MPT multimodal capabilities
• ggml: Optimized MPT version for efficient inference on consumer hardware
• GPT4All: locally running chat system, now with MPT support!
• Q8MPT-Chat: 8-bit optimized MPT for CPU by our friends at Intel

Tutorial videos from the community:

• Using MPT-7B with Langchain by @jamesbriggs
• MPT-7B StoryWriter Intro by AItrepreneur
• Fine-tuning MPT-7B on a single GPU by @AIology2022
• How to Fine-tune MPT-7B-Instruct on Google Colab by @VRSEN

Something missing? Contribute with a PR!

• Blog: LLM Training and Inference with Intel Gaudi2 AI Accelerators
• Blog: Training LLMs at Scale with AMD MI250 GPUs
• Blog: Training LLMs with AMD MI250 GPUs and MosaicML
• Blog: Announcing MPT-7B-8K: 8K Context Length for Document Understanding
• Blog: Training LLMs with AMD MI250 GPUs and MosaicML
• Blog: MPT-30B: Raising the bar for open-source foundation models
• Blog: Introducing MPT-7B
• Blog: Benchmarking LLMs on H100
• Blog: Blazingly Fast LLM Evaluation
• Blog: GPT3 Quality for $500k
• Blog: Billion parameter GPT training made easy

This codebase has been tested with PyTorch 2.1 with NVIDIA A100s and H100s. This codebase may also work on systems with other devices, such as consumer NVIDIA cards and AMD cards, but we are not actively testing these systems. If you have success/failure using LLM Foundry on other systems, please let us know in a Github issue and we will update the support matrix!

We highly recommend using our prebuilt Docker images. You can find them here: https://hub.docker.com/orgs/mosaicml/repositories.

The mosaicml/pytorch images are pinned to specific PyTorch and CUDA versions, and are stable and rarely updated.

The mosaicml/llm-foundry images are built with new tags upon every commit to the main branch. You can select a specific commit hash such as mosaicml/llm-foundry:1.13.1_cu117-f678575 or take the latest one using mosaicml/llm-foundry:1.13.1_cu117-latest.

Please Note: The mosaicml/llm-foundry images do not come with the llm-foundry package preinstalled, just the dependencies. You will still need to pip install llm-foundry either from PyPi or from source.

This assumes you already have PyTorch, CMake, and packaging installed. If not, you can install them with pip install cmake packaging torch.

To get started, clone the repo and set up your environment. Instructions to do so differ slightly depending on whether you're using Docker.

We strongly recommend working with LLM Foundry inside a Docker container (see our recommended Docker image above). If you are doing so, follow these steps to clone the repo and install the requirements.

git clone https://github.com/mosaicml/llm-foundry.git
cd llm-foundry
pip install -e ".[gpu]"  # or pip install -e . if no NVIDIA GPU

If you choose not to use Docker, you should create and use a virtual environment.

git clone https://github.com/mosaicml/llm-foundry.git
cd llm-foundry

# Creating and activate a virtual environment
python3 -m venv llmfoundry-venv
source llmfoundry-venv/bin/activate

pip install cmake packaging torch  # setup.py requires these be installed

pip install -e ".[gpu]"  # or pip install -e . if no NVIDIA GPU

NVIDIA H100 GPUs have FP8 support; this additionally requires the following installations:

pip install flash-attn==1.0.7 --no-build-isolation
pip install git+https://github.com/NVIDIA/TransformerEngine.git@v0.10

See here for more details on enabling TransformerEngine layers and amp_fp8.

In our testing of AMD GPUs, the env setup includes:

git clone https://github.com/mosaicml/llm-foundry.git
cd llm-foundry

# Creating and activate a virtual environment
python3 -m venv llmfoundry-venv-amd
source llmfoundry-venv-amd/bin/activate

# installs
pip install cmake packaging torch
pip install -e .  # This installs some things that are not needed but they don't hurt
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2

Lastly, install the ROCm enabled flash attention (instructions here).

Notes:

1. attn_impl: triton does not work.
2. We don't yet have a Docker image where everything works perfectly. You might need to up/downgrade some packages (in our case, we needed to downgrade to numpy==1.23.5) before everything works without issue.

Support for LLM Foundry on Intel Gaudi devices is experimental, please use the branch habana_alpha and see the README on that branch which has install instructions and known issues.

For training and inference performance results on Intel Gaudi2 accelerators, see our blog: https://www.databricks.com/blog/llm-training-and-inference-intel-gaudi2-ai-accelerators

Note Make sure to go through the installation steps above before trying the quickstart!

Here is an end-to-end workflow for preparing a subset of the C4 dataset, training an MPT-125M model for 10 batches, converting the model to HuggingFace format, evaluating the model on the Winograd challenge, and generating responses to prompts.

(Remember this is a quickstart just to demonstrate the tools -- To get good quality, the LLM must be trained for longer than 10 batches 😄)

cd scripts

# Convert C4 dataset to StreamingDataset format
python data_prep/convert_dataset_hf.py \
  --dataset c4 --data_subset en \
  --out_root my-copy-c4 --splits train_small val_small \
  --concat_tokens 2048 --tokenizer EleutherAI/gpt-neox-20b --eos_text '<|endoftext|>'

# Train an MPT-125m model for 10 batches
composer train/train.py \
  train/yamls/pretrain/mpt-125m.yaml \
  data_local=my-copy-c4 \
  train_loader.dataset.split=train_small \
  eval_loader.dataset.split=val_small \
  max_duration=10ba \
  eval_interval=0 \
  save_folder=mpt-125m

# Convert the model to HuggingFace format
python inference/convert_composer_to_hf.py \
  --composer_path mpt-125m/ep0-ba10-rank0.pt \
  --hf_output_path mpt-125m-hf \
  --output_precision bf16 \
  # --hf_repo_for_upload user-org/repo-name

# Evaluate the model on a subset of tasks
composer eval/eval.py \
  eval/yamls/hf_eval.yaml \
  icl_tasks=eval/yamls/copa.yaml \
  model_name_or_path=mpt-125m-hf

# Generate responses to prompts
python inference/hf_generate.py \
  --name_or_path mpt-125m-hf \
  --max_new_tokens 256 \
  --prompts \
    "The answer to life, the universe, and happiness is" \
    "Here's a quick recipe for baking chocolate chip cookies: Start by"

Note: the composer command used above to train the model refers to the Composer library's distributed launcher.

If you have a write-enabled HuggingFace auth token, you can optionally upload your model to the Hub! Just export your token like this:

export HUGGING_FACE_HUB_TOKEN=your-auth-token

and uncomment the line containing --hf_repo_for_upload ... in the above call to inference/convert_composer_to_hf.py.

Check out TUTORIAL.md to keep learning about working with LLM Foundry. The tutorial highlights example workflows, points you to other resources throughout the repo, and answers frequently asked questions!

If you run into any problems with the code, please file Github issues directly to this repo.

If you want to train LLMs on the MosaicML platform, reach out to us at demo@mosaicml.com!