[faresobeid]

Removes config knobs that nothing reads or that nobody sets:

• env_server.output_dir: no reader. src/prime_rl/entrypoints/env_server.py never uses it, and the launchers do not write it (write_env_server_config writes only env, serve, address_file and log).
• heartbeat.min_interval: no config, test or doc sets it. It becomes the constant MIN_INTERVAL_S = 30.0 in src/prime_rl/utils/heartbeat.py, the value that fix(utils): throttle BetterStack heartbeat pings #3715 picked for Better Stack's rate limit.
• The deployment.num_eval_gpus / deployment.num_eval_nodes aliases on SFT: both arrived together with num_infer_gpus / num_infer_nodes in feat: online evals during sft #3256, so they were never a legacy name. Nothing references them.
• SFT fake data.input_ids: fake data now always uses the increasing ids, which was already the config default. The random branch and the now-unused vocab_size argument of FakeDataset are removed, so no config changes behavior. increasing is the simpler path because it makes no generator draws. The exact-order dataloader tests already rely on it. One limitation already existed before this PR: each sample's token id is its sample index, so a very long fake-data run eventually goes past the vocab size.

I re-checked the other candidates from the config audit and left them out:

• dynamo.enabled: open feat(trainer): support VLM training with dynamo  #3483, feat(inference): launch managed Dynamo workers #3493, feat(dynamo): support python based dynamo vLLM workers administration #3629 and Add ModelExpress public-client weight broadcast #3890 read or validate it, e.g. by rejecting enabled = false for managed Dynamo.
• rollout_transport.hwm: refactor: orchestrator as hook-wired components with per-component configs #3652 derives it from max_off_policy_steps.
• trainer data.fake.generate_samples: still read (src/prime_rl/trainer/rl/data.py:77-87), and feat: integrate DSpark draft training, weight sync, and checkpoints #3729 sets it in an example.

Test plan (all run):

• pytest tests/unit/orchestrator tests/unit/test_configs.py tests/unit/train/sft tests/unit/utils -m "not gpu": 482 passed. Fake-dataset tests were updated for the removed arguments.
• Script checks:
  • five rapid Heartbeat.beat() calls send one ping;
  • num_eval_gpus is now rejected and num_infer_gpus still works;
  • EnvServerConfig(output_dir=...) is rejected.
• ruff check and ruff format --check on the changed files.

---

Note

Low Risk
Config surface and test-only fake-data paths shrink with no production readers of the removed fields; default fake-data behavior matches the former increasing default.

Overview
Removes several unused or redundant config knobs and aligns runtime code with the slimmer schemas.

Config removals: env_server.output_dir (never read by the env-server entrypoint or launchers); heartbeat.min_interval (replaced by a fixed 30s throttle in Heartbeat via MIN_INTERVAL_S); SFT deployment aliases num_eval_gpus / num_eval_nodes (only num_infer_* remain); and fake SFT data.input_ids.

Fake SFT data: FakeDataset always emits deterministic “increasing” token ids (sample index repeated per position); the random-id path and vocab_size plumbing are dropped. Variable-length fake data still randomizes sequence length only.

Docs on eval heartbeat behavior now refer to the hard-coded 30s interval. Unit tests were updated for the new FakeDataset API and configs.

^{Reviewed by Cursor Bugbot for commit 5fbf89e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[mikasenghaas]

u sure this is not set? env servers also log to output dir no?

[mikasenghaas]

i think @samsja added this intentionally

[mikasenghaas]

i feel like random might be useful?