Question about the recommended local setup for H100

[Doow501]

Hi, thank you for open-sourcing this project.

I’m trying to set up a local environment on an H100 machine, and I wanted to ask about the recommended dependency combination for the default training workflow.

While reading the repo, I saw that:

• the README suggests torch==2.5.1 with cu124 for Ampere/Hopper
• the default config seems to use vllm rollout
• setup.py allows vllm>=0.8.5,<=0.12.0 and requires numpy<2.0.0
• requirements.txt pins numpy==2.1.0
• the helper install script pins vllm==0.11.0

I may be misunderstanding the intended setup, but it seems possible to end up with a local torch 2.5.1 + cu124 environment first, and then hit version changes again once vllm is needed by the default rollout path.

Could you please clarify what the recommended local setup is for H100, especially for the default PPO/SDPO training path?

In particular, it would be very helpful to know the suggested versions for:

• torch
• CUDA
• numpy
• vllm

Also, which source should I follow as the main reference for local installation: the README, setup.py, requirements.txt, or the helper install script?

Thanks a lot for your help.

[Doow501]

I solved this problem by using verl's image.

[joejajo]

I solved this problem by using verl's image.

Hi im using a100 can you help with installation i also face the same issue

[Doow501]

I solved this problem by using verl's image.

Hi im using a100 can you help with installation i also face the same issue

It might help to pull the latest image:
docker pull verlai/verl:vllm017.latest

[joejajo]

I solved this problem by using verl's image.

Hi im using a100 can you help with installation i also face the same issue

It might help to pull the latest image: docker pull verlai/verl:vllm017.latest

thanks it did resolve my evn issues