[faresobeid]

Under rl, an explicit value for a launcher-derived field must now equal the derived value; a different value fails validation instead of being silently overwritten.

What

• trainer.weight_broadcast / orchestrator.weight_broadcast: must match the shared [weight_broadcast], except host and inference_world_size, which the launcher fills per deployment.
• orchestrator.pad_to_multiple_of: derived from trainer.model.cp.
• orchestrator.num_train_workers: derived from the deployment when it is set by the launcher (more than 1 train GPU, or multi-node).
• New helper set_derived in utils/validation.py.
• validate_enough_devices_for_nccl now runs after the shared broadcast is copied in and reads the shared weight_broadcast, so [weight_broadcast] type = "nccl" on 1 GPU fails validation.

Equal values are allowed: the single-node SLURM launch re-parses the resolved rl.json, which contains every derived value (test_single_node_nccl_resolved_json_roundtrips). host is not compared because the single-node NIXL template overrides --weight-broadcast.host on that re-parse.

Not changed: dp_replicate from deployment.nodes_per_fsdp_group (#3992 deletes it); orchestrator.inference_metrics_roles (already filled only when unset). No shipped TOML under configs/ or examples/ sets any of these fields.

Validation

• tests/unit/test_configs.py: parametrized conflict test, resolved-JSON roundtrip, test_rl_nccl_broadcast_needs_two_gpus (fails on main), test_load_configs over every shipped TOML.
• Branch merged with main 1810d5b9a: pytest tests/unit/orchestrator tests/unit/test_configs.py tests/unit/utils tests/unit/eval -m "not gpu": 491 passed. ruff check, ruff format --check and uv lock --check are clean.

Overlaps

• feat(trainer): remove CP token padding and duplicated RL microbatches #3724 changes the pad_to_multiple_of derivation; wrap its expression in set_derived.
• feat(weights): add lossless sparse delta synchronization #3881, feat: add ModelExpress weight updates #3487, Add ModelExpress public-client weight broadcast #3890 add broadcast types/fields; they flow through the shared block.

---

Note

Low Risk
Changes are config validation only at parse time; they prevent misconfiguration but do not alter runtime training or broadcast behavior when configs are valid.

Overview
RL config validation now treats launcher-derived fields as read-only: an explicit value is accepted only if it matches what the launcher would compute; otherwise validation fails with a clear error instead of silently replacing the user's setting.

This applies to orchestrator.pad_to_multiple_of, orchestrator.num_train_workers (when derived from deployment), and trainer / orchestrator.weight_broadcast (must align with shared [weight_broadcast], except host and inference_world_size, which stay launcher-filled). A new check_derived helper in validation.py implements the equality check so resolved rl.json round-trips still work.

NCCL GPU check runs after shared weight broadcast is copied onto sub-configs, so [weight_broadcast] type = "nccl" on a single-GPU layout fails validation even when set only at the top level.

Unit tests cover conflicting overrides, the NCCL two-GPU rule, and existing round-trip behavior.

^{Reviewed by Cursor Bugbot for commit 9590070. Bugbot is set up for automated code reviews on this repo. Configure here.}