[Bugfix] torchrun compatibility

[hiyouga]

This PR fixes the vllm's offline inference when using torchrun as the external launcher in veRL broken after #13591

When we launch vllm using torchrun with tp = 2 on 8 GPUs, the layout of the distributed world should be (DP = 4 * PP = 1 * TP = 2). However, we find setting VLLM_DP_SIZE to 4 causes the process to hang after generation. So, we use the parallel_config to determine how to compute the dp size. (actually we don't know why, maybe #13591 is not fully compatible with external launchers)

This modification is also the same as https://verl.readthedocs.io/en/latest/README_vllm0.7.html#use-vllm-v1-engine described

Regarding the get_pp_indices function, we take the modulus of the pp rank by the pp size to ensure that the pp rank is always smaller than the pp size.

This PR has been validated using the latest veRL on the GSM8K task.

I'm a newbie in vLLM. Please feel free to let me know if there are any problems.

cc @youkaichao

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[youkaichao]

thanks for the investigation!