Enforce a 90% unified-memory cap + purge KV on unload#363

Merged

Gajesh2007 merged 19 commits into

masterfrom

feat/unified-memory-cap

Jun 16, 2026

Commits on Jun 16, 2026

Add UnifiedMemoryCap + ResidentModelLedger (90%-cap policy foundation)
anupsv
committed
Route the KV budgets through UnifiedMemoryCap (drop competing reserves)
anupsv
committed
Apply the hardCap floor to live KV headroom (close the small-box OOM gap)
anupsv
committed
Make the model-load gate + doctor honor the 90% cap
anupsv
committed
Purge KV from RAM + SSD on unload, sweep stale KV at startup
anupsv
committed
Address review: load gate honors activation reserve; clear cache on standalone evict
anupsv
committed
Address re-review: clearCache on all standalone unload paths; fix ledger docs
anupsv
committed
Drop stale load-gate wording from a ResidentModelLedger test comment
anupsv
committed
Add post-load measured-headroom guard (close the estimate-vs-measured gap)
anupsv
committed
Derive defaultLoadHeadroomGb from the cap (remove the admit-then-reject trap)
anupsv
committed
Address bot review: remove unused ResidentModelLedger; cast-guard tokenBudgetMax
anupsv
committed
Gate the saveBlock test seam behind #if DEBUG (drop production surface)
anupsv
committed
Count VLM media-decode RAM against the 90% unified-memory cap
anupsv
committed
Reserve VLM vision-path generation KV against the 90% cap
anupsv
committed
Size vision-path KV by full token span; don't let env bricks the cap
anupsv
committed
Merge remote-tracking branch 'origin/master' into tmp/363-merge
Gajesh2007
committed
fix(memory): reserve loading model's footprint during load (serve-while-load OOM)

Gajesh2007
and
anupsv
committed
Merge remote-tracking branch 'origin/feat/unified-memory-cap' into tmp/363-merge
Gajesh2007
committed
fix(memory): honor operator memory_reserve_gb in the live KV gate (Codex P2)

Gajesh2007
and
anupsv
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Enforce a 90% unified-memory cap + purge KV on unload#363

Enforce a 90% unified-memory cap + purge KV on unload#363
Gajesh2007 merged 19 commits into
masterfrom
feat/unified-memory-cap

Commits on Jun 16, 2026

Add UnifiedMemoryCap + ResidentModelLedger (90%-cap policy foundation)

Route the KV budgets through UnifiedMemoryCap (drop competing reserves)

Apply the hardCap floor to live KV headroom (close the small-box OOM gap)

Make the model-load gate + doctor honor the 90% cap

Purge KV from RAM + SSD on unload, sweep stale KV at startup

Address review: load gate honors activation reserve; clear cache on standalone evict

Address re-review: clearCache on all standalone unload paths; fix ledger docs

Drop stale load-gate wording from a ResidentModelLedger test comment

Add post-load measured-headroom guard (close the estimate-vs-measured gap)

Derive defaultLoadHeadroomGb from the cap (remove the admit-then-reject trap)

Address bot review: remove unused ResidentModelLedger; cast-guard tokenBudgetMax

Gate the saveBlock test seam behind #if DEBUG (drop production surface)

Count VLM media-decode RAM against the 90% unified-memory cap

Reserve VLM vision-path generation KV against the 90% cap

Size vision-path KV by full token span; don't let env bricks the cap

Merge remote-tracking branch 'origin/master' into tmp/363-merge

fix(memory): reserve loading model's footprint during load (serve-while-load OOM)

Merge remote-tracking branch 'origin/feat/unified-memory-cap' into tmp/363-merge

fix(memory): honor operator memory_reserve_gb in the live KV gate (Codex P2)