[rollout] feat: use dummy load_format when init AsyncServer

[vermouth1992]

What does this PR do?

• Loading weights in AsyncServer is duplicated and is time-consuming for large models
• Use dummy weights instead as the actual weights will be transferred by the trainer

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, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • 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

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

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).)

[gemini-code-assist]

Code Review

This pull request changes the load_format for the vLLM engine to dummy to optimize startup time by not loading model weights. While this is a good optimization, it introduces a critical race condition where the server can process requests with a randomly-initialized model, leading to incorrect outputs. My review includes a critical comment with a suggested fix to add a synchronization mechanism to ensure the server only handles requests after valid weights have been loaded.

[gemini-code-assist]

Using load_format="dummy" is a good optimization for startup time, as it avoids loading model weights during initialization. However, this introduces a critical race condition.

The AsyncvLLMServer becomes ready to serve requests via its FastAPI endpoints as soon as init_engine completes. At this point, the underlying vLLM engine has a model with random weights. Any call to generate or chat_completion before the trainer has loaded the actual weights will be processed by this dummy model, resulting in garbage output without any error. This can lead to silent failures that are hard to debug.

To ensure correctness, the server must prevent generation with dummy weights. I recommend adding a synchronization mechanism to block requests until the real weights are loaded.

You could use an asyncio.Event for this:

1. In AsyncvLLMServer.__init__, add the event:

   self._weights_loaded = asyncio.Event()

2. Add a new method to AsyncvLLMServer that the trainer can call after loading the weights:

   async def mark_weights_as_loaded(self):
       """Signal that the model weights have been loaded."""
       self._weights_loaded.set()

3. In generate and chat_completion methods, wait for this event before proceeding:

   async def generate(self, ...):
       await self._weights_loaded.wait()
       # ... rest of the method
   
   async def chat_completion(self, raw_request: Request):
       await self._weights_loaded.wait()
       # ... rest of the method

This will make the server robust by ensuring it only processes requests after it's in a valid state.

[wuxibin89]

We may need individual rollout for debugging, e.g custom agent loop develop, in which case the trainer may not be started and no weight transfer.

[vermouth1992]

We may need individual rollout for debugging, e.g custom agent loop develop, in which case the trainer may not be started and no weight transfer.

Yes. Let me refactor the config of AsyncLLM first.

[ETOgaosion]

Maybe "dummy" load format, this breaks vllm test.