feat(kvflash): snapshot ledger + QK-pool restore-seed + consume-restored-KV#446

Closed

dusterbloom wants to merge 11 commits into

Luce-Org:mainfrom

dusterbloom:pr/kvflash-ledger-consume

Commits on Jun 24, 2026

perf(qwen35moe): pool-sized KV reservation keeps experts hot at high max_ctx
dusterbloom
committed
feat(kvflash): pager serialize/deserialize + critical-chunk pinning
dusterbloom
committed
feat(qwen35moe): pooled chunked prefill + pooled snapshot/restore over KVFlash
dusterbloom
committed
refactor(qwen35moe): dedup chunked-prefill loops + inline for_each_segment
dusterbloom
committed
fix(kvflash): save pooled prefix snapshot at chunk boundary so ≥128K restore works
dusterbloom
committed
feat(kvflash): snapshot ledger + QK-pool restore-seed (lib + tests) for QK×PR372 composition
dusterbloom
committed
feat(kvflash): consume restored KV on pooled restore instead of re-prefill (opt-in)
dusterbloom
committed
fix(kvflash): reconstruct restored-prefix drafter-history from ledger; enable consume default-on
dusterbloom
committed
test(server): cold_prefix_boundary returns 4000 after layout_known_ guard removal
dusterbloom
committed
fix(cubic): address production P1/P2 review findings (bounded loop, OOB, UAF, atoi-UB, re-prefill fallback, pin_range, gguf type guards)
dusterbloom
committed

Commits on Jun 25, 2026

fix(kvflash): reconstruct pooled-K from ledger on QK warm-restore
dusterbloom
committed