fix(#1157): NemotronH decoded with STALE input ids, and the recurrent carry was never the cause - #1221

Merged

localai-bot merged 9 commits into

mainfrom

row/MODEL-NEMOTRON-H-ABI-A2P-1157-fix

Aug 18, 2026

Commits on Aug 18, 2026

diag(#1157): instrument the NemotronH recurrent carry and add a fresh-prefill discriminator
mudler
committed
ef3bcfd
diag(#1157): a device-MoE bypass and a both-modes driver, so one GPU slot can bisect
mudler
committed
ed049a7
test(#1157): the device MoE gate covered two PREFILL widths and never the decode one
mudler
committed
a75ab96
test(#1157): the FA-2 d128 decode gate measured two GQA ratios, and the failing model is a third
mudler
committed
65b0759
fix(#1157): the paged forward embedded the STALE host token ids, so every decode step saw the same token
mudler
committed
8b566b6
record(#1157): the A3 gate has a host PASS, a named device cause, and the seam that allowed it
mudler
committed
0fba1af
merge: origin/main into the #1157 NemotronH decode repair
mudler
committed
f8e083f
docs(#1157): the diagnostic is not scaffolding, so its comment should not call itself temporary
mudler
committed
6e9e895
merge: origin/main into the #1157 stale-input-ids fix
mudler
committed
967e263