Gemma4 mtp recover#1

Draft

fhnmor21 wants to merge 13 commits into

feature/turboquant-kv-cachefrom

gemma4-mtp-recover

Commits on May 23, 2026

chore: ignore local worktree directory
fhnmor21
committed
feat: add Gemma 4 MTP assistant support

Ooooze
authored and
fhnmor21
committed
feat: implement async MTP draft pipeline for improved speculative decoding

Ooooze
authored and
fhnmor21
committed
feat: enhance speculative decoding with pipeline depth-2 and async MTP handling

Ooooze
authored and
fhnmor21
committed
feat: add MTP acceptance tracing and enhance draft processing

Ooooze
authored and
fhnmor21
committed
feat: enhance MTP performance with in-graph argmax and acceptance tracing

Ooooze
authored and
fhnmor21
committed
feat: enhance MTP tensor handling and introduce centroid routing

Ooooze
authored and
fhnmor21
committed
feat: add TurboQuant KV cache and weight compression support

Ooooze
authored and
fhnmor21
committed
llama: prefix MTP assistant tensors with 'mtp.' on load

sujitvasanth
authored and
fhnmor21
committed
fix: Gemma 4 time-to-first-token drops from 8-12s to <1s by unblocking cache reuse

sujitvasanth
authored and
fhnmor21
committed
fix: Gemma 4 + TurboQuant KV no longer crashes on second prompt when --cache-reuse enabled

sujitvasanth
authored and
fhnmor21
committed
fix: restore gemma4 mtp compatibility overlay
fhnmor21
committed
fix: restore Gemma4 MTP CLI aliases
fhnmor21
committed