Skip to content

Latest commit

 

History

History
133 lines (88 loc) · 5.15 KB

README.md

File metadata and controls

133 lines (88 loc) · 5.15 KB
Raw

Nyuntam

Nyuntam is NyunAI's advanced suite designed to optimize, adapt, and accelerate a wide array of deep learning models across various domains. The repository is structured into several submodules, each targeting specific tasks:

  • Nyuntam Text Generation: Focuses on compressing large language models for text generation tasks.
  • Nyuntam Vision: Tailored for compressing and optimizing vision models.
  • Nyuntam Adapt: A robust module for fine-tuning and transfer learning across both vision and language models leveraging SoTA PEFT, full-finetuning and GPU parallelism.

Key Features

  • State-of-the-Art Compression: Includes techniques such as pruning, quantization, distillation, and more, to ensure model efficiency without compromising performance.
  • Adaptation: Fine-tune and adapt vision and language models to specific tasks and datasets using methods like (Q)LoRA, (Q)SSF, and others.
  • Multi-Platform Support: Run experiments seamlessly on various platforms using Docker or virtual environments.
  • CLI Integration: Simplify your workflow with the NyunZero CLI.

Installation

NyunZero CLI (Recommended)

The fastest way to install and use Nyuntam is through the NyunZero CLI. It streamlines the setup process and gets you running experiments in no time.

For more details, visit the NyunZero CLI Documentation.

Git + Docker

  1. Clone the Repository:

    git clone --recursive https://github.com/nyunAI/nyuntam.git
cd nyuntam

  2. Pull the Required Docker Image:

    docker pull <docker_image_name_or_id>

  3. Run the Docker Container:

    docker run -it -d --gpus all -v $(pwd):/workspace --name {CONTAINER_NAME} --network=host <docker_image_name_or_id> bash

NOTE: Ensure that nvidia-container-toolkit is installed before running the Docker container. For access to gated repositories within the containers, add Hugging Face tokens.

Git + Virtual Environment

  1. Clone the Repository:

    git clone --recursive https://github.com/nyunAI/nyuntam.git
cd nyuntam

  2. Create and Activate a Virtual Environment:

    python3 -m venv {ENVIRONMENT_NAME}
source {ENVIRONMENT_NAME}/bin/activate

  3. Install the Required Packages:

    # for text_generation
pip install -r text_generation/requirements.txt 
# for vision
pip install -r vision/requirements.txt 
# for nyuntam_adapt
pip install -r nyuntam_adapt/requirements.txt

Docker for Submodules

For running each submodule independently via Docker, use the following commands:

  • Nyuntam Text Generation:

    docker pull nyunadmin/nyunzero_text_generation:v0.1
docker run -it -d --gpus all -v $(pwd):/workspace --name {CONTAINER_NAME} --network=host nyunadmin/nyunzero_text_generation:v0.1 bash

  • Nyuntam Vision:

    docker pull nyunadmin/nyunzero_kompress_vision:v0.1
docker run -it -d --gpus all -v $(pwd):/workspace --name {CONTAINER_NAME} --network=host nyunadmin/nyunzero_kompress_vision:v0.1 bash

  • Nyuntam Adapt:

    docker pull nyunadmin/nyunzero_adapt:v0.1
docker run -it -d --gpus all -v $(pwd):/workspace --name {CONTAINER_NAME} --network=host nyunadmin/nyunzero_adapt:v0.1 bash

Usage

Setting Up YAML Configuration Files

Each submodule in Nyuntam requires a YAML file that defines all the necessary hyperparameters and configurations for running experiments. Examples and templates for these YAML files can be found within each submodule.

  • Nyuntam Text Generation:

  • Nyuntam Vision:

    • Example YAML configurations are can be found here.
    • Details on setting up model & dataset paths, compression algorithms, and other parameters can be found in the documentation here.

  • Nyuntam Adapt:

    • For fine-tuning and transfer learning tasks, example YAML configurations are available here
    • Detailed explanations on model & dataset path setups, params, and algorithms are available in the documentation.

Running Experiments

To run an experiment, use the following command from within the appropriate submodule directory:

python main.py --yaml_path {path/to/recipe.yaml}

This command will execute the main script using the configurations specified in the YAML file.

Examples

For detailed examples and use cases, refer to examples

Documentation

For complete details checkout NyunAI Docs