[OMNIML-5570, OMNIML-5569] 1/2 Add layer-wise KV-cache AutoQuant with forward KL - #2272

Merged

meenchen merged 30 commits into

mainfrom

agent/kv-cache-autoquant-core

Sep 11, 2026

Commits on Aug 18, 2026

Add layer-wise KV cache AutoQuant search
meenchen
committed
0cbc2c1
Add three-format mixed KV export
meenchen
committed
abf5959
Harden KV cache AutoQuant search
meenchen
committed
c6e9ac6
Test KV AutoQuant report export
meenchen
committed
d26d912
Fix KV AutoQuant state and export handling
meenchen
committed
300c95f

Commits on Aug 19, 2026

fix: harden KV-cache AutoQuant workflows
meenchen
committed
b57b483
fix: guard KV-cache AutoQuant execution
meenchen
committed
722ab98

Commits on Aug 20, 2026

fix: validate KV-cache AutoQuant candidates
meenchen
committed
faa1f95

Commits on Aug 23, 2026

Fix KV AutoQuant architecture edge cases
meenchen
committed
b907f1c

Commits on Aug 25, 2026

Fix KV AutoQuant cost validation
meenchen
committed
34a2495
Fix KV AutoQuant export validation
meenchen
committed
c06fb0e

Commits on Aug 27, 2026

Use calibration-free KV AutoQuant candidates
meenchen
committed
efc25b3

Commits on Sep 1, 2026

Merge main into KV-cache AutoQuant core
meenchen
committed
d1b5e47
Limit FSDP2 rejection to KV AutoQuant
meenchen
committed
d6e0a66
Clarify KV AutoQuant recipe CLI help
meenchen
committed
e75d06e

Commits on Sep 2, 2026

Unify KV-cache search with AutoQuantize
meenchen
committed
7a1f35a

Commits on Sep 3, 2026

Fix KV AutoQuant replay and export metadata
meenchen
committed
f0f7dad
Merge main into KV-cache AutoQuant core
meenchen
committed
9fbe0a1
Avoid per-forward KV score synchronization
meenchen
committed
1c9279e
Preserve AutoQuant format type compatibility
meenchen
committed
3e2d100
Accept covariant AutoQuant format sequences
meenchen
committed
fdec562

Commits on Sep 4, 2026

Fix KV AutoQuant export compatibility
meenchen
committed
19fe678
Fix KV AutoQuant export and restore edge cases
meenchen
committed
1cd026c

Commits on Sep 5, 2026

Address KV AutoQuant review feedback
meenchen
committed
f5fc9a5

Commits on Sep 9, 2026

Address KV AutoQuantize review feedback
meenchen
committed
4bfee89

Commits on Sep 10, 2026

Preserve BF16 metadata for KV-only export
meenchen
committed
dfd1386
Merge main into KV AutoQuantize core
meenchen
committed
0d25480
Use production model in KV AutoQuantize example
meenchen
committed
1da6da2
Merge export refactor into KV AutoQuantize core
meenchen
committed
311cc9d
Move KV candidate naming into AutoQuantize
meenchen
committed
ee1ac3e