neel04/ReAct_Jax

ReAct architecture and training loop - now in Jax!

ReAct_Jax

ReAct architecture and training loop - now in Jax!

Setup

run.sh is the runner script for the Docker container. It pulls the latest version of the code from the dev branch, and runs train_model.py with the arguments specified in TRAIN_ARGS.

In the commands below, I download the script off a GitHub gist. You're better off forking this repo and changing the repo to directly wget from your fork.

The script takes care of actually pulling the docker container, the repo, and running the training loop across a TPU Pod slice automagically.

Important

Make sure to modify & rebuild the Dockerfile with your own API tokens (for WandB & HuggingFace)

Caution

Replace the docker image path (the envar IMAGE_PATH) in run.sh with your own DockerHub URL.

1. Run run.sh on your TPU pod slice (tested extensively with TPUv4-32)
2. Declare your INSTANCE_NAME from GCP via export INSTANCE_NAME=<your_instance_name>
3. Enjoy

Inferencing

python3 inferencer.py --checkpoint_path '/Users/neel/Documents/research/ReAct_Jax/ReAct/outputs/model 5000.eqx' --num_blocks 3 --width 256 --n_heads 4 --seqlen 196  --prompt "Sam is sad because"

Doing a training run

1. First, get a preemptible TPUv4 node as a queued resource:

   gcloud alpha compute tpus queued-resources create $INSTANCE_NAME \
   --node-id react-node \
   --project react-jax \
   --zone us-central2-b \
   --accelerator-type v4-32 \
   --runtime-version tpu-vm-v4-base \
   --best-effort

2. (Optiona) Setup the TPU pod slice with basics (nvim + NvChad):

   gcloud compute tpus tpu-vm ssh $INSTANCE_NAME \
   --zone=us-central2-b --worker=all --command="\
       sudo apt-get update; \
       sudo snap install nvim --classic; \
       git clone https://github.com/NvChad/starter ~/.config/nvim && nvim; \
       echo 'Setup done!'"

3. And then actually kickoff the training by downloading the run.sh script and running it:

   gcloud compute tpus tpu-vm ssh $INSTANCE_NAME \
   --zone=us-central2-b --worker=all --command="\
       tmux kill-server; sudo rm -rf ./ReAct_Jax/ReAct/outputs/; \
       sleep 3s && wget https://gist.githubusercontent.com/neel04/3bfc7e4d9cd746829b7e72f1b6fac5de/raw/run.sh; \
       sleep 5s && tmux new-session -d 'bash run.sh &> out.log'; \
       sudo rm ./run.sh"

Troubleshooting

If you get errors regarding workers not being able to sync up at the distributed barrier, do:

gcloud compute tpus tpu-vm ssh --zone "us-central2-b" $INSTANCE_NAME --worker 'all' --project "react-jax" --command 'sudo docker system prune -f && sudo rm -rf ~/.cache;'

If Docker is unresponsive, just restart docker service:

gcloud compute tpus tpu-vm ssh --zone "us-central2-b" $INSTANCE_NAME --worker 'all' --project "react-jax" --command 'sudo systemctl restart docker'