feat: add online ppo with value-model critic and gae advantages

- register online_ppo train type backed by PPOStrategy: token-level clipped surrogate over GAE advantages plus masked value regression against rollout-pinned returns, with explained-variance metrics - fold the reference-KL penalty (k3 estimator) into per-token rewards before GAE and pin advantages/returns on RolloutResult so replayed gradient steps optimize fixed targets - add self-contained ValueModel critic with a zero-initialized value head and backbone warm-started from policy weights; AutoRegressiveLM stays untouched and trunk parity is pinned by tests - step the critic's own optimizer outside the policy-version lock with the same max_grad_norm clipping as the policy - persist critic state as value_model.pt/value_optimizer.pt checkpoint extras; resume restores it, fails loudly when missing, and the train.sh completeness check requires the extras for online_ppo configs - extract shared rollout sequence/logprob helpers from GRPO (behavior unchanged) and add ppo_gamma/ppo_gae_lambda/ppo_vf_coef CLI options
2026-09-05 01:59:50 +08:00
parent 816b96a58a
commit 350e4a1849
17 changed files with 1390 additions and 72 deletions
@@ -186,6 +186,10 @@ scheduler.pt
 manifest.json
 ```

+`online_ppo` jobs additionally require `value_model.pt` and
+`value_optimizer.pt` (the critic state); the completeness check derives this
+from the training config's `train_type`.
+
 New checkpoints write `manifest.json` after every payload file, sync the complete
 staging directory, and then atomically rename that directory into place. Legacy
 checkpoints without a manifest remain resumable when the original required files