Metal: Reconcile DSpark runtime (#480) + B2 rejection sampling/adaptive block sizing (#482) - #502

Open

stephenlthorn wants to merge 10 commits into

antirez:mainfrom

stephenlthorn:dspark-integration

Commits on Jul 5, 2026

Add guarded DSpark loader plumbing

audreyt
authored and
stephenlthorn
committed
Add DeepSpec nonseq draft-head schema gate

audreyt
authored and
stephenlthorn
committed
Add DeepSpec DSpark scaffold

audreyt
authored and
stephenlthorn
committed
Add DSpark speculative draft runtime

audreyt
authored and
stephenlthorn
committed
Make DSpark nonseq runtime ready

audreyt
authored and
stephenlthorn
committed
Commit DSpark partial accepts from prefix checkpoints

audreyt
authored and
stephenlthorn
committed
[dspark] Add B2 rejection sampling for lossless speculative decode

authored and
stephenlthorn
committed
[dspark] Fix off-by-one in B2 target logits + persist RNG state

authored and
stephenlthorn
committed
[dspark] Add adaptive block sizing (DS4_DSPARK_ADAPTIVE)

authored and
stephenlthorn
committed
[dspark] Stage adaptive escalation, document argmax-replay invariant
stephenlthorn
committed