add Temp Scheduler

[hallerite]

Note

Introduces scheduled sampling temperature and ensures it flows end-to-end from rollout to training.

• Config/Changelog: Adds sampling.temperature_schedule (constant/linear/cosine) with validation; documents in CHANGELOG.md
• Runtime: Computes step-wise temperature (compute_temperature) and injects into get_sampling_args; logs sampling/temperature
• Workers/Scheduler: EnvWorker now receives sampling_args and temperature; responses tag rollouts with temperature; Scheduler.set_sampling_args(...) updates args and temperature used for requests
• Transport/Trajectories: TrainingSample gains optional temperature; interleave_rollout/branch_rollout attach it
• Batching/Packer: prepare_batch now takes temperatures list; packing sorts and groups microbatches by temperature to avoid mixing; Single/Multi packers propagate and group by temperature
• Trainer: Scales logits by microbatch temperature when computing logprobs/entropy
• Tests: Update/add unit tests for batching by temperature and EnvWorker API

^{Written by Cursor Bugbot for commit 9ffb15e. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[Jackmin801]

hey @hallerite! nice PR, excited to be supporting temperature scheduling in the orchestrator. some quick comments:

1. You're probably going to have resolve a lot of merge conflicts when [Refactor] rename Runs -> MultiRunManager and cleanup hooks #1619 merges soon. Sorry about that. Its mostly renaming stuff so I would think the agents are capable enough to figure out the merge conflict resolution but yea sorry for the trouble.

2. Im not sure if grouping by temperature in the packer is necessary atm. We already group by run and implicitly by step (we never pack across step boundary for a run in a step) and so we will never be in a case where the prepare_batch loop sees any unique run having different temps in the same batch.

Though with async I guess we can have multiple sampling temperatures within a batch for a run with max unique temps <= max async. But in that case, I would think the better solution is to make the loss support passing a tensor of temperatures rather than packing such that similar temperatures are together.

[samsja]

not sure what is up with your ruff , did you used the pre-commit hook ?

[hallerite]

I did now