| Blog | Roadmap |

News 🔥

• [2023/09] Medusa v0.1 is released! 🎉

Medusa is a simple framework that democratizes the acceleration techniques for LLM generation with multiple decoding heads.

Medusa on Vicuna-7b.

We aim to tackle the three pain points of popular acceleration techniques like speculative decoding:

• Requirement of a good draft model.
• System complexity.
• Inefficiency when using sampling-based generation.

Medusa adds extra "heads" to LLMs to predict multiple future tokens simultaneously. When augmenting a model with Medusa, the original model stays untouched, and only the new heads are fine-tuned during training. During generation, these heads each produce multiple likely words for the corresponding position. These options are then combined and processed using a tree-based attention mechanism. Finally, a typical acceptance scheme is employed to pick the longest plausible prefix from the candidates for further decoding.

We aim to solve the challenges associated with speculative decoding by implementing the following ideas:

• Instead of introducing a new model, we train multiple decoding heads on the same model.
• The training is parameter-efficient so that even the "GPU-Poor" can do it. And since there is no additional model, there is no need to adjust the distributed computing setup.
• Relaxing the requirement of matching the distribution of the original model makes the non-greedy generation even faster than greedy decoding.

• Introduction
• Contents
• Installation
  • Method 1: With pip
  • Method 2: From source
  • Model Weights
  • Inference
  • Training
    • Prepare the data
    • Train the model
    • Push to Hugging Face Hub
• Citation
• Codebase Guide
• Contributing
• Acknowledgements

pip install medusa-llm

git clone https://github.com/FasterDecoding/Medusa.git
cd Medusa
pip install -e .

We currently support single-GPU inference with a batch size of 1, which is the most common setup for local model hosting. We are actively working to extend Medusa's capabilities by integrating it into other inference frameworks; please don't hesitate to reach out if you are interested in contributing to this effort.

You can use the following command for launching a CLI interface:

CUDA_VISIBLE_DEVICES=0 python -m medusa.inference.cli --model [path of medusa model]

You can also pass --load-in-8bit or --load-in-4bit to load the base model in quantized format. If you download the base model elsewhere, you may override base model name or path with --base-model [path of base model].

For training, please install:

pip install -e ".[train]"

We take a public version of the ShareGPT dataset, which is a subset of the Vicuna training data. For other models, you can use the corresponding training dataset.

git clone https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered

Remark: If you haven't installed git-lfs, please install it before clone:

git lfs install

We follow the training setup from FastChat, but with a much larger learning rate because we freeze the original model and only train the new heads. Here is the training command for the Vicuna-7b model on 4 GPUs. Since we are only training the new heads, the training does not require a lot of memory, and only data parallelism is needed. You can modify the script to fit your own setup. For larger models, we use the same setup. You can also use --load_in_8bit or --load_in_4bit to load the base model in quantized format.

torchrun --nproc_per_node=4 medusa/train/train.py --model_name_or_path lmsys/vicuna-7b-v1.3 \
    --data_path ShareGPT_Vicuna_unfiltered/ShareGPT_V4.3_unfiltered_cleaned_split.json \
    --bf16 True \
    --output_dir test \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "no" \
    --learning_rate 1e-3 \
    --weight_decay 0.0 \
    --warmup_ratio 0.1 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --lazy_preprocess True \
    --medusa_num_heads 3 \
    --medusa_num_layers 1

You can use the following command to push your model to the Hugging Face Hub:

python -m medusa.hf_utils --folder [path of the model folder] --repo [name of the repo]

@misc{medusa,
  author = {Tianle Cai and Yuhong Li and Zhengyang Geng and Hongwu Peng and Tri Dao},
  title = {Medusa: Simple Framework for Accelerating LLM Generation with Multiple Decoding Heads},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/FasterDecoding/Medusa}},
}

medusa/model/medusa_model.py is the key file for Medusa. It contains the MedusaModel class, which is a wrapper of the original model and the new heads. This class also has implementation of a streaming generation method. If you want to dive into the details of Medusa, this is the place to start.

We also provide some illustrative notebooks in notebooks/ to help you understand the codebase.

We welcome community contributions to Medusa. If you have an idea for how to improve it, please open an issue to discuss it with us. When submitting a pull request, please ensure that your changes are well-tested. Please split each major change into a separate pull request. We also have a Roadmap summarizing our future plans for Medusa. Don't hesitate to reach out if you are interested in contributing to any of the items on the roadmap.

This codebase is influenced by remarkable projects from the LLM community, including FastChat, TinyChat, vllm and many others.