[faresobeid]

Summary

SFTConfig duplicated most of TrainerConfig. This PR adds BaseTrainerConfig in configs/trainer.py with the shared fields and validators. TrainerConfig and SFTConfig both subclass it.

• Shared fields: model, tokenizer, optim, scheduler, ckpt, log, matmul_precision, max_steps, memory_profiler_path, gc, trace_path, dist_timeout_seconds, heartbeat, control, env_vars.

• Shared validators: deepep_disables_grad_clipping, full_optimizer_offload_requires_supported_optimizer, full_optimizer_offload_disables_grad_clipping, vlm_freeze_incompatible_with_lora, dont_do_massive_traces, validate_scheduler_steps, auto_setup_tokenizer.

• The three "auto" settings that resolve differently are class variables, resolved by one resolve_auto_defaults validator (replaces three validators per class):

  class variable	RL (TrainerConfig)	SFT (SFTConfig)
  auto_moe_router_dtype	float32	bfloat16
  auto_freeze_moe_router	True	False
  auto_weight_decay	0.0	0.01

• resume, monitors, output_dir and weight_broadcast stay per class: their types or meanings differ.

Behavior

No runtime change. Defaults, resolved values, warnings and validation errors are the same.

Visible differences:

• SFT help text uses the shared wording for two fields: max_steps ("Maximum number of training steps.") and env_vars ("...for the trainer process(es)...").
• Field order. Shared fields come first in model_fields, --help and resolved-config dumps. The set of fields and their values do not change.
• After-validator order. Shared validators run before class-specific ones. No class-specific validator reads a value that a moved validator sets. If a config has several errors, a different one may be reported first.

Proof

• Differential dump against main at the PR's original base: every TOML under configs/ and examples/ (78 files; there is no k8s/ TOML), standalone and composed with each sibling TOML, plus 18 CLI override cases (auto overrides, trace/max_steps, full offload, Muon + FSDP offload, cosine scheduler, DeepEP, LoRA, tokenizer). Each case was parsed with all 7 config classes, with run.name uuids pinned. That is 2310 cases, 245 of which parse. model_dump, error messages and warnings are identical.
• model_json_schema() is identical for all 7 config classes.
• --help: TrainerConfig and RLConfig have the same lines (order only). SFTConfig differs only in the two descriptions above.
• On current main (52455b4): every TrainerConfig / SFTConfig field has the same default and type as on main (including dist_timeout_seconds = 900 from feat(trainer): 900 s collective timeout, flight recorder dumps, master-only waits off NCCL #4019).
• pytest tests/unit/test_configs.py tests/unit/orchestrator -m "not gpu": 336 passed.
• ruff check and ruff format --check are clean.

---

Note

Low Risk
Refactor-only config hierarchy with parity testing claimed; no training logic changes, only consolidation and validator ordering.

Overview
Introduces BaseTrainerConfig in trainer.py so RL TrainerConfig and SFTConfig share one definition of model/tokenizer/optim/scheduler, logging, profiling, distributed timeout, and the common cross-cutting validators (DeepEP vs grad clip, full optimizer offload, VLM+LoRA, trace limits, scheduler steps, tokenizer auto-setup).

SFTConfig now subclasses the base instead of re-declaring those fields and copy-pasted validators. RL- vs SFT-specific behavior for "auto" settings is expressed as subclass ClassVar defaults (auto_moe_router_dtype, auto_freeze_moe_router, auto_weight_decay) resolved in a single resolve_auto_defaults validator on the base.

RL-only and SFT-only pieces stay on the respective subclasses (e.g. RL loss / rollout_transport; SFT renderer, deployment, online eval wiring). Intended runtime behavior is unchanged; help text and field/validator ordering may differ slightly.

^{Reviewed by Cursor Bugbot for commit 1b459c1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[samsja]

lgtm, i thought this was already the case tho

@snimu something to adapt into the pretraining entrypoint