[rollout] feat: enable Async RL for trtllm rollout

[hchings]

What does this PR do?

Requires verl's trtllm version to be updated to include NVIDIA/TensorRT-LLM#12272 [Merged].

**Note that this MR only enables e2e async RL functionalities for trtllm rollout and tested convergence, it hasn't included perf optimizations for the rollout itself. Currently, we do see python overheads in _update_requests of trtllm rollout (especially when wtih TorchSampler). We'll have separate MRs to address rollout's perf.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

Convergence of Qwen3-8B-base and Qwen3-30B-A3B-Base of this MR aligns with vllm with the same config.

1. Convergence - Qwen3-8B-base

bypass=False + rollout_IS=token. e2e script.

• Green line: trtllm https://wandb.ai/nvidia/DAPO_fully_async_erinh/runs/b9sez762?nw=nwusererinh
• The other two lines: vllm as reference (one with bypass=False + IS=token, the other with bypass=True and IS=null)

*Note that token_multi_prob_error is a tentatively metric adopting from nemo-rl doc.

2. Convergence - Qwen3-30B-A3B-Base

https://wandb.ai/nvidia/DAPO_fully_async_30B_erinh?nw=nwusererinh
Two trtllm curves, two vllm curves, for bypass=False + rollout_IS=token and bypass=True + rollout_IS=null correspondingly.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[gemini-code-assist]

Code Review

This pull request introduces abort and resume functionality for AsyncLLM in trtllm rollouts. The core logic changes in trtllm_async_server.py correctly implement this by mapping pause_generation and resume_generation calls and handling the cancelled finish reason. A new test file, test_trtllm_abort.py, is added to validate this functionality. My review focuses on the new test file, where I've identified a couple of areas for improvement to enhance its robustness and maintainability, particularly around configuration path resolution and Ray cluster cleanup.

[gemini-code-assist]

The method for locating the configuration directory is fragile as it depends on the current working directory from which pytest is executed. This can lead to test failures if run from a different directory (e.g., inside the tests/ directory). A more robust approach would be to determine the project's root directory programmatically and construct the path from there. This will make the test more reliable and easier to run in different environments.

Please also add from pathlib import Path to the top-level imports of the file.

Suggested change

	config_dir = os.path.abspath("verl/verl/trainer/config")
	if not os.path.exists(config_dir):
	config_dir = os.path.abspath("verl/trainer/config")
	from pathlib import Path
	repo_root = Path(__file__).resolve().parents[4]
	config_dir_path = repo_root / "verl/verl/trainer/config"
	if not config_dir_path.exists():
	config_dir_path = repo_root / "verl/trainer/config"
	if not config_dir_path.exists():
	raise FileNotFoundError(f"Config directory not found. Searched under {repo_root / 'verl'}")
	config_dir = str(config_dir_path)

[gemini-code-assist]

The use of subprocess.run(["ray", "stop"]) is risky for test cleanup. It's a forceful command that can interfere with other parallel tests using Ray and depends on the ray CLI being in the PATH. The ray.shutdown() call on the previous line is the standard and safer way to clean up a local Ray cluster and should be sufficient. Relying on ray stop can mask resource leak issues that should be addressed directly.

[CLAassistant]

All committers have signed the CLA.