Repository navigation
Async trainers write lora_name into a private attribute that RolloutWorkerProtocol does not declare #8008
Description
With PEFT adapter sync on, AsyncGRPOTrainer sets the adapter name on an injected rollout worker through self.rollout_worker._loop_kwargs["lora_name"]. #7984 ports the same write to AsyncDistillationTrainer.
RolloutWorkerProtocol declares only the queues and the lifecycle/version methods. _loop_kwargs is a private detail of AsyncRolloutWorker. Building the trainer fails with AttributeError when all of these hold:
- the injected worker implements the protocol but is not an
AsyncRolloutWorker, - the model is a PEFT model,
- the vLLM server is started with
--enable-lora, - no
weight_transferis injected.
Root cause: the protocol has no way to pass the adapter name. An injected worker is built before the trainer probes the server, so the trainer can only tell it the name afterwards, and the only place it can write it is AsyncRolloutWorker's private _loop_kwargs.
Raised in the review of #7984: #7984 (comment) and #7984 (comment)
Activity
dundysm commentedon Oct 31, 2026
hi @albertvillanova; i'd like to take this one. plan: add a
set_lora_name(lora_name)method toRolloutWorkerProtocolin both async_grpo and async_distillation, implement it on bothAsyncRolloutWorkers (it writes the name into the loop kwargs beforestart()), and have both trainers call it instead of touching_loop_kwargs. tests cover an injected worker that only implements the protocol. pr shortly unless you'd prefer a different shape.albertvillanova commentedon Nov 5, 2026
Thanks @dundysm for picking this up, but I'd hold off on the
set_lora_namePR. #8066 and #8107 (both approved) remove therollout_workerargument andRolloutWorkerProtocolfrom both async trainers, and this_loop_kwargswrite goes away with them. Custom rollout logic moves to subclassing the rollout loop instead (see #7827). Linking this issue to those PRs so it closes when they land.