Execution API¶
Maturity: verified
Execution APIs define how callers start a unit of work and inspect its result without making provider lifecycle part of the application contract.
Pretraining interfaces¶
The current surface executes one complete fresh or resumed pretraining attempt. The earlier single-loader executor and partial model/optimizer checkpoint format have been removed; the application now has one training and evidence path.
Application lifecycle¶
Symbol |
Contract |
|---|---|
caller-supplied run, attempt, repository, mount, and optional provider provenance |
|
execute the application-owned train/validation loop and publish terminal evidence |
|
successful attempt identities, artifact URIs, final step, and loss |
Checkpoints and evidence¶
Symbol |
Contract |
|---|---|
advance one completed training step and any due full-validation boundary |
|
exact train/validation metric-event vocabulary |
|
append metric events and bind checkpoint-safe byte prefixes |
|
save model, optimizer, scheduler, loader, loop, random, and metric-prefix state |
|
reuse one checkpoint already verified by the checkpoint writer in terminal evidence |
|
bind the frozen metric cursor and exact latest/best checkpoints into one immutable result |
|
verify the metric records and checkpoint selections used for exact resume |
The executor performs full validation at configured boundaries, follows the fixed warmup-stable-decay learning-rate schedule, and records exact latest and best checkpoint references. On resume, it verifies the prior attempt before creating the child workspace, restores the full PyTorch and loader state, copies the checkpoint-bound metric prefix, and continues with a new attempt identity. The caller still owns retry and provider lifecycle; distributed execution is outside this contract.