[vllm, rollout, cfg, doc] feat: Accelerate RL rollouts with EAGLE/EAGLE3 speculative decoding

[alekseymalakhov11]

What does this PR do?

Add inference-only speculative decoding for vLLM rollouts using EAGLE / EAGLE3 draft models. This PR adds the vLLM rollout/config wiring for the new speculative decoding path, reload handling for the draft model during weight sync, rollout metrics for speculative decoding, and documentation.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: [rollout] feat: support eagle3 speculative decode in rollout
  [rollout] feat: support eagle3 speculative decode in rollout #5509

• 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

We used qwen3-8b together with RedHatAI/Qwen3-8B-speculator.eagle3 on verl-team/lighteval-MATH-preprocessed.

In our experiments, the final reward is comparable to the baseline. The screenshots also highlights two implementation details that are important for making this work correctly in RL rollout: reloading the draft-model weights after actor weight sync, and rebuilding the RoPE cos_sin cache after reload. Also it shows speculative decoding metrics such as acceptance rate and mean acceptance length so the behavior and performance of EAGLE/EAGLE3 can be analyzed directly during training.

In some settings, performance of draft model may degrade (acceptance length drops). Even with that limitation, this change opens a useful direction for more practical research on speculative decoding in RL for LLMs, and makes it easier to explore follow-up ideas such as retraining or adapting the draft model.

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 support for inference-only speculative decoding in vLLM rollout, including necessary configuration updates, metrics tracking, and weight loading adjustments. I have identified two issues: an incorrect handling of the acceptance rate when no draft tokens are generated, and a malformed f-string in a validation error message.

[gemini-code-assist]

The calculation for acceptance_rate when spec_delta["num_draft_tokens"] is 0 results in float("inf"). This is likely incorrect and can cause issues with metric aggregation (e.g., np.mean over inf will be inf). When no draft tokens are generated, the acceptance rate should be 0.0.

Suggested change

	spec_delta["num_accepted_tokens"] / spec_delta["num_draft_tokens"]
	if spec_delta["num_draft_tokens"] > 0
	else float("inf")
	spec_delta["num_accepted_tokens"] / spec_delta["num_draft_tokens"]
	if spec_delta["num_draft_tokens"] > 0
	else 0.0

[gemini-code-assist]

The f-string for the ValueError message is malformed. The variable self.tensor_model_parallel_size is outside the curly braces, so it will be printed literally instead of its value being interpolated. This will produce a confusing error message for users.

Suggested change

	raise ValueError(
	f"draft_tensor_parallel_size={self.speculative_decoding.draft_tensor_parallel_size} "
	"cannot be other value than 1 or target model "
	"tensor_parallel_size={self.tensor_model_parallel_size} "
	)
	raise ValueError(
	f"draft_tensor_parallel_size={self.speculative_decoding.draft_tensor_parallel_size} "
	"cannot be other value than 1 or target model "
	f"tensor_parallel_size={self.tensor_model_parallel_size}"
	)