Skip to content

Latest commit

 

History

History
37 lines (25 loc) · 1.39 KB

README.md

File metadata and controls

37 lines (25 loc) · 1.39 KB
Raw

Third-parties (LLM Ecosystems)

Dependencies of Auto Evaluation Toolkit, see tools/evaluator/README.md for more details.

Installation

The auto-evaluation toolkit requires customized Megatron-LM and HELM. To avoid dependency problems when installing those packages, we recommend using NGC's PyTorch container (nvcr.io/nvidia/pytorch:22.12-py3). Assuming the path to your shared file system (where your data and model checkpoints are saved) is /mnt/shared, start the docker container with following commands.

docker pull nvcr.io/nvidia/pytorch:22.12-py3
docker run --gpus all -it --rm -v /mnt/shared:/workspace

After starting the docker container, run the following scripts in the container to install Megatron-LM or HELM.

The training machines only need to install Megatron-LM:

./setup_megatron.sh

The evaluation machine needs to install both Megatron-LM and HELM

./setup_megatron.sh
./setup_helm.sh

The toolkit use W&B (wandb) to monitor the trend of metrics during training. Above steps have installed wandb, and you only need to run wandb login and enter your wandb API key. If you have your own instance of wandb, run the following script.

wandb login --host <URL of your wandb instance>
# enter your api key