Running Inference on Dolly 2.0 Model with Hugging Face

Introduction​

Dolly 2.0, the groundbreaking, open-source, instruction-following Large Language Model (LLM) that has been fine-tuned on a human-generated instruction dataset, licensed for both research and commercial purposes. Developed using the EleutherAI Pythia model family, this 12-billion-parameter language model is built exclusively on a high-quality, human-generated instruction following dataset, contributed by Databricks employees.

Dolly 2.0 package is open source, including the training code, dataset, and model weights, all available for commercial use. This unprecedented move empowers organizations to create, own, and customize robust LLMs capable of engaging in human-like interactions, without the need for API access fees or sharing data with third parties.

Running locally​

Prerequisites​

1. A NVIDIA GPU

2. Python

Installing dependencies​

pip -q install git+https://github.com/huggingface/transformers # need to install from github
pip -q --upgrade install accelerate # ensure you are using version higher than 0.12.0

Create an inference.py file with following code:

# content of the inference.py file
import argparse
import torch
from transformers import pipeline

def main(prompt_string, model_version):

    # use dolly-v2-12b if you're using Colab Pro+, using pythia-2.8b for Free Colab
    generate_text = pipeline(model=model_version, 
                            torch_dtype=torch.bfloat16, 
                            trust_remote_code=True,
                            device_map="auto")

    print(generate_text(prompt_string))

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--prompt", type=str, required=True, help="The prompt to be used in the GPT model")
    parser.add_argument("--model_version", type=str, default="./databricks/dolly-v2-12b", help="The model version to be used")
    args = parser.parse_args()
    main(args.prompt, args.model_version)

Building the container (optional)​

You may want to create your own container for this kind of task. In that case, use the instructions for creating and publishing your own image in the docker hub. Use huggingface/transformers-pytorch-deepspeed-nightly-gpu as base image, install dependencies listed above and copy the inference.py into it. So your Dockerfile will look like this:

FROM huggingface/transformers-pytorch-deepspeed-nightly-gpu
RUN apt-get update -y
RUN pip -q install git+https://github.com/huggingface/transformers
RUN pip -q install accelerate>=0.12.0 
COPY ./inference.py .

Running Inference on Bacalhau​

Prerequisite​

To get started, you need to install the Bacalhau client, see more information here

Structure of the command​

1. export JOB_ID=$( ... ): Export results of a command execution as environment variable

2. bacalhau docker run: Run a job using docker executor.

3. --gpu 1: Flag to specify the number of GPUs to use for the execution. In this case, 1 GPU will be used.

4. -w /inputs: Flag to set the working directory inside the container to /inputs.

5. -i gitlfs://huggingface.co/databricks/dolly-v2-3b.git: Flag to clone the Dolly V2-3B model from Hugging Face's repository using Git LFS. The files will be mounted to /inputs/databricks/dolly-v2-3b.

6. -i https://gist.githubusercontent.com/js-ts/d35e2caa98b1c9a8f176b0b877e0c892/raw/3f020a6e789ceef0274c28fc522ebf91059a09a9/inference.py: Flag to download the inference.py script from the provided URL. The file will be mounted to /inputs/inference.py.

7. jsacex/dolly_inference:latest: The name and the tag of the Docker image.

8. The command to run inference on the model: python inference.py --prompt "Where is Earth located ?" --model_version "./databricks/dolly-v2-3b". It consists of:

   1. inference.py: The Python script that runs the inference process using the Dolly V2-3B model.

   2. --prompt "Where is Earth located ?": Specifies the text prompt to be used for the inference.

   3. --model_version "./databricks/dolly-v2-3b": Specifies the path to the Dolly V2-3B model. In this case, the model files are mounted to /inputs/databricks/dolly-v2-3b.

When a job is submitted, Bacalhau prints out the related job_id. We store that in an environment variable so that we can reuse it later on.

export JOB_ID=$(bacalhau docker run \
    --gpu 1 \
    --id-only \
    -w /inputs \
    -i gitlfs://huggingface.co/databricks/dolly-v2-3b.git \
    -i https://gist.githubusercontent.com/js-ts/d35e2caa98b1c9a8f176b0b877e0c892/raw/3f020a6e789ceef0274c28fc522ebf91059a09a9/inference.py \
    jsacex/dolly_inference:latest \
    -- python inference.py --prompt "Where is Earth located ?" --model_version "./databricks/dolly-v2-3b")

Checking the State of your Jobs​

Job status: You can check the status of the job using bacalhau job list:

bacalhau job list --id-filter ${JOB_ID}

When it says Completed, that means the job is done, and we can get the results.

Job information: You can find out more information about your job by using bacalhau job describe:

bacalhau job describe ${JOB_ID}

Job download: You can download your job results directly by using bacalhau job get. Alternatively, you can choose to create a directory to store your results. In the command below, we created a directory and downloaded our job output to be stored in that directory.

rm -rf results && mkdir results
bacalhau job get ${JOB_ID} --output-dir results

Viewing your Job Output​

After the download has finished, we can see the results in the results/outputs folder.