mtp: prefix-cache WARM hit (perfect + partial via range-warm)#221

Open

dusterbloom wants to merge 12 commits into

Luce-Org:mainfrom

dusterbloom:feat/mtp-prefix-warm-ghost

Commits on May 19, 2026

mtp: native-heads MTP speculator (Qwen3.6 NextN, γ-chain)
dusterbloom
committed
mtp: prefix-cache WARM hit (perfect + partial via range-warm)
dusterbloom
committed
mtp: load NextN head blocks onto GPU by default
dusterbloom
committed
qwen35: fix do_spec_decode argmax OOB on prefix-cache partial restore
dusterbloom
committed
pflash: propagate skip_park to daemon compress command
dusterbloom
committed
bench: add 2026-05-19 matrix — MTP/DFlash/PFlash agent + long-ctx
dusterbloom
committed
feat(speculator): per-request DFlash/MTP/auto dispatcher + PFlash drafter-arch protocol
dusterbloom
committed
bench: restore bench_matrix orchestrator (recovered from f031f08)
dusterbloom
committed
bench(results): archive 2026-05-19 matrix runs + dual-speculator summary update
dusterbloom
committed
docs(thoughts): dual-speculator architecture + drafter unification + resume plan
dusterbloom
committed
docs(thoughts): CLI client setup quickstart for dual-speculator server
dusterbloom
committed
docs(thoughts): HRM-Text-1B integration assessment memo
dusterbloom
committed