feat(mdp): run per-layer CUDA graphs on the MDP decoder - #49

Open

BestJuly wants to merge 5 commits into

lit/mdp-partial-cudagraph-v2

Commits on Aug 20, 2026

feat(mdp): greedy token-budget packing and fixed-shape THD batches

BestJuly
and
claude
committed
8892dc4
fix(mdp): reserve a cu_seqlens slot for the static padding tail

BestJuly
and
claude
committed
4813e7e

Commits on Aug 21, 2026

test(mdp): assert the derived pad_between_seqs, not the old hardcoded True

BestJuly
and
claude
committed
8a51361
test(mdp): assert FLOPs honesty through the emitted override

BestJuly
and
claude
committed
5200c89
feat(mdp): run per-layer CUDA graphs on the MDP decoder

BestJuly
and
claude
committed
3c8e72c