[sergiopaniego]

What

Updates the vLLM server command in tutorial/examples/wordle.py from the deprecated trl vllm-serve to vLLM's native server, matching the migration in huggingface/trl#6765:

CUDA_VISIBLE_DEVICES=0 VLLM_SERVER_DEV_MODE=1 vllm serve Qwen/Qwen3-1.7B --host 0.0.0.0 --port 8000 \
    --weight-transfer-config '{"backend": "nccl"}' \
    --logprobs-mode processed_logprobs \
    --max-logprobs -1

TRL version

huggingface/trl#6765 shipped in TRL v1.11.0, so this command needs TRL >= 1.11.0. The tutorial's uv header installs trl[vllm] without a version pin, so it resolves to the latest release, which includes the change.

This PR was prepared with AI assistance.

---

Note

Low Risk
Documentation-only change to example shell commands; no runtime or training code is modified.

Overview
Updates the Wordle GRPO tutorial’s Option 2 setup so the separate vLLM server uses vLLM’s native vllm serve instead of the deprecated trl vllm-serve.

The new command sets VLLM_SERVER_DEV_MODE=1 and adds flags for NCCL weight transfer, processed_logprobs, and unlimited max logprobs so the server matches TRL’s post–v1.11.0 expectations for GRPO training against a remote vLLM instance.

^{Reviewed by Cursor Bugbot for commit 6236537. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.