Running webshop in a distributional way

[DZ9]

Thanks for the great work! I found this repo support distributional training by accelerate. But for webshop task it seems tricky to train it in a distributional way because of the extra server installation.

Here are my questions:

1. Can I run webshop server on a separate machine and train archer? Are the following command must run within one machine?

   python -m PATH_TO_WEBSHOP/web_agent_site.app --log --attrs &
   cd scripts
   python run.py --config-name archer_webshop

2. Are there any other ways to run webshop env with accelerate that you have tried? I may run it with 16 or more gpu cards, which are 2 or more machines in nature.

[YifeiZhou02]

Thanks for your interest in our work. I have not tested whether existing code can work with multiple machines, and the current distributional training is only tested with 8 gpus.

For setting up webshop servers on a different server, it should suffice if you set up the server on a different machine and make use of ssh port forwarding to tunnel the corresponding ports (e.g. 3000) on both machines. For speeding up webshop, you may find #6 helpful.

[DZ9]

Thanks for your interest in our work. I have not tested whether existing code can work with multiple machines, and the current distributional training is only tested with 8 gpus.

For setting up webshop servers on a different server, it should suffice if you set up the server on a different machine and make use of ssh port forwarding to tunnel the corresponding ports (e.g. 3000) on both machines. For speeding up webshop, you may find #6 helpful.

This speedup method really helps, great thanks! May I know how many gpu hours will it cost to train webshop with gpt2 or mistral7B in your previous settings?

[YifeiZhou02]

I don't remember exactly but it took around 3 days on 4xA5000 GPUs to train on webshop with gpt2. This was without the speedup method so I imagine it should be made faster with the speedup.