Per-stage init_from / freeze for rollout-capable partial-retrain checkpoints #42

New Issue

2026-08-13T15:08:02+02:00

lars commented

2026-08-13 15:08:02 +02:00

active = false already builds one stage alone — but the resulting checkpoint
can't be rolled out at all, so the interesting ablation isn't runnable end to
end. The obvious next step given that v0.3 split the config per stage:

[stage1_model]
init_from = "checkpoints/…/best.pt"
freeze = true

Then "retrain only the secondary decoder against a fixed, known-good stage
1" — precisely the experiment the 2026-08-03 species failure calls for — is a
config change rather than a bespoke script, and the output is a complete,
rollout-capable checkpoint. Cheap: build_models already constructs stages
independently, and checkpoint_io already knows how to load a partial state
dict for the migration path.

Migrated from issues.md (v0.3.0 branch review, 2026-08-13), Issue 17.

`active = false` already builds one stage alone — but the resulting checkpoint can't be rolled out at all, so the interesting ablation isn't runnable end to end. The obvious next step given that v0.3 split the config per stage: ```toml [stage1_model] init_from = "checkpoints/…/best.pt" freeze = true ``` Then "retrain **only** the secondary decoder against a fixed, known-good stage 1" — precisely the experiment the 2026-08-03 species failure calls for — is a config change rather than a bespoke script, and the output is a complete, rollout-capable checkpoint. Cheap: `build_models` already constructs stages independently, and `checkpoint_io` already knows how to load a partial state dict for the migration path. --- Migrated from `issues.md` (v0.3.0 branch review, 2026-08-13), Issue 17.

lars added the feature label 2026-08-13 15:08:02 +02:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: lars/giant#42