[rollout,vllm] Fix DP args and local_rank for Ray NOSET_VISIBLE_DEVICES

[JohnConnor123]

Summary

• Fix A: ensure vLLM receives DP CLI args when data_parallel_size > 1 (not only when EP is enabled).
• Fix B: when running under Ray with RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1 and data_parallel_size > 1,
  compute the correct vLLM-local LOCAL_RANK (local rank within TP×PP) to avoid DP adjusted local rank ... out of bounds
  and incorrect GPU binding.
• This is a refinement of the comments from PR [rollout,vllm] Fix DP args and local_rank for Ray NOSET_VISIBLE_DEVICES #5152

Context / Motivation

In DP configs (typically dp=2,tp=1,pp=1,nnodes=1):

• vLLM could start without DP settings if the condition was tied only to EP.
• Under Ray NOSET, using ray.get_runtime_context().get_accelerator_ids() to derive local rank can produce a rank that becomes
  out-of-bounds after vLLM DP adjustment.

Checklist

• Bug fixed
• Repro / test plan included
• Regression test (real Ray+vLLM e2e)

[gemini-code-assist]

Code Review

This pull request introduces two important fixes for running vLLM with data parallelism under Ray. The first fix ensures that data parallelism arguments are correctly passed to vLLM, which was previously only done for expert parallelism. The second fix aims to correct the local_rank for workers when using Ray's NOSET_VISIBLE_DEVICES mode to prevent out-of-bounds errors.

My review focuses on the implementation of the local_rank adjustment. While the change to include DP arguments is correct, the logic for local_rank correction in verl/workers/rollout/vllm_rollout/utils.py contains critical bugs related to argument parsing that would prevent it from functioning. I've provided a detailed comment with a corrected implementation.