Note
23nd Aug 2024 Update:
- Demo: https://demo.homebrew.ltd/
- Processes single-sound instruction data, under 10s, English
Warning
llama3-s is an open research experiment
- Join us in the
#researchchannel in Homebrew's Discord - We livestream training runs in
#research-livestream
llama3-s is an open, ongoing research experiment to extend a text-based LLM to have native "listening" ability. Think of it as an open data, open weight, on device Siri.
It uses an early fusion technique inspired by Meta's Chameleon paper.
We build train in public:
- 23 Aug: We’re excited to share llama3.1-s-instruct-v0.2, our latest multimodal checkpoint with improved speech understanding by enhancing the model's audio instruction-following capabilities through training on interleaving synthetic data.
- 17 Aug: We pre-trained our LLaMA 3.1 model on continuous speech data, tokenized using WhisperSpeechVQ. The final loss converged to approximately 1.9, resulting in our checkpoint: llama3.1-s-base-v0.2
- 2 Aug: Retrained phase 1 with llama3.1 and fixes to hyperparameters, achieving significant improvement (MMLU: 0.66 -> 0.61)
- 1 Aug: Identified typo in original training recipe, causing significant degradation (MMLU: 0.6 -> 0.2), proposed fixes.
- 30 July: Presented llama3-s progress at: AI Training: From PyTorch to GPU Clusters
- 19 July: llama3-s-2024-07-19 understands synthetic voice with limited results
- 1 July: llama3-s-2024-07-08 showed converging loss (1.7) with limited data
llama3-s is an open research project. We're looking for collaborators, and will likely move towards crowdsourcing speech datasets in the future.
Checkout this notebook to try our latest model:
For detailed information on synthetic generation, please refer to the Synthetic Generation Guide.
- First Clone the Repo from github:
git clone --recurse-submodules https://github.com/homebrewltd/llama3-s.git
- The folder structure is as follows:
llama3-s
├── HF_Trainer # HF training code (deprecated)
├── synthetic_data # Synthetic data generation pipeline
├── configs # Audio pipeline configs
├── audio_to_audio # Parler audio (.wav) to semantic tokens
├── synthetic_generation_config # TTS semantic tokens
├── scripts # Setup scripts for Runpod
├── torchtune # Submodule: our fork of fsdp with checkpointing
├── model_zoo # Model checkpoints
│ ├── LLM
│ │ ├── Meta-Llama-3-8B-Instruct
│ │ ├── Meta-Llama-3-70B-Instruct
├── demo # Selfhost this demo (vllm)
├── inference # Google Colab
- Install Dependencies
python -m venv hf_trainer
chmod +x scripts/install.sh
./scripts/install.sh
Restart shell now
chmod +x scripts/setup.sh
./scripts/setup.sh
source myenv/bin/activate
- Logging Huggingface
huggingface-cli login --token=<token>
- Training
export CUTLASS_PATH="cutlass"
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
accelerate launch --config_file ./accelerate_config.yaml train.py
- Install Package
python -m venv torchtune
pip install torch torchvision tensorboard
cd ./torchtune
pip install -e .
You can also download the model using tune:
tune download homebrewltd/llama3.1-s-whispervq-init --hf-token <token> --output-dir ../model_zoo/llama3.1-s-whispervq-init --ignore-patterns "original/consolidated*"
Setup the Dataset from HF path by change the path and change the name of the model in the following YAML file.
nano torchtune/recipes/configs/jan-llama3-s/8B_full.yaml
- Training Multi GPU (1-8GPUs Supported)
tune run --nproc_per_node 4 full_finetune_fsdp2 --config recipes/configs/jan-llama3-1-s/8B_full.yaml
We offer code for users to create a web UI demo. Please follow the instructions below:
python -m venv demo
source demo/bin/activate
# First install all required packages
pip install --no-cache-dir -r ./demo/requirements.txt
Then run the command below to launch a Gradio demo locally. You can add the variables use-4bit and use-8bit for quantized usage:
python -m demo.app --host 0.0.0.0 --port 7860 --max-seq-len 1024
You can also host a demo using vLLM for faster inference but its not support streaming output:
python -m demo.app_vllm
Alternatively, you can easily try our demo on HuggingFace 🤗
@misc{chameleonteam2024chameleonmixedmodalearlyfusionfoundation,
title={Chameleon: Mixed-Modal Early-Fusion Foundation Models},
author={Chameleon Team},
year={2024},
eprint={2405.09818},
archivePrefix={arXiv},
primaryClass={cs.CL},
journal={arXiv preprint}
}
@misc{zhang2024adamminiusefewerlearning,
title={Adam-mini: Use Fewer Learning Rates To Gain More},
author={Yushun Zhang and Congliang Chen and Ziniu Li and Tian Ding and Chenwei Wu and Yinyu Ye and Zhi-Quan Luo and Ruoyu Sun},
year={2024},
eprint={2406.16793},
archivePrefix={arXiv},
primaryClass={cs.LG},
journal={arXiv preprint}
}
@misc{defossez2022highfi,
title={High Fidelity Neural Audio Compression},
author={Défossez, Alexandre and Copet, Jade and Synnaeve, Gabriel and Adi, Yossi},
year={2022},
eprint={2210.13438},
archivePrefix={arXiv},
journal={arXiv preprint}
}
@misc{WhisperSpeech,
title={WhisperSpeech: An Open Source Text-to-Speech System Built by Inverting Whisper},
author={Collabora and LAION},
year={2024},
url={https://github.com/collabora/WhisperSpeech},
note={GitHub repository}
}- Torchtune: The codebase we built upon
- Accelerate: Library for easy use of distributed training
- WhisperSpeech: Text-to-speech model for synthetic audio generation
- Encodec: High-fidelity neural audio codec for efficient audio compression
- Llama3: the Family of Models that we based on that has the amazing language capabilities !!!
