feat(#810 A2-Q1): NemotronH's 23 Mamba2 blocks reach the device on the FP8 W8A8 seam (#517) - #1289

Draft

localai-bot wants to merge 20 commits into

mainfrom

feat(#810 A2-Q1): NemotronH's 23 Mamba2 blocks reach the device on the FP8 W8A8 seam (#517)
mudler
committed
0f33640
test(#810 A2-Q1): the A3 gate reports its per-token time with the terms it divided, or refuses
mudler
committed
681504c
fix(#810 A2-Q1): the carry gate banded a defect smaller than the noise it had to accept, so it gates the STATE
mudler
committed
5921853
docs(#810 A2-Q1): record what Thor measured, including the question #960 left open
mudler
committed
d6d643c
fix(#810 A2-Q1): the busy fraction stopped quoting a GB10 baseline beside a Thor number
mudler
committed
a3f09a0
docs(#810 A2-Q1): record the Thor A3 pass, and say plainly which baseline it cannot be read against
mudler
committed
1b0b6a7
fix(#810 A2-Q1): the busy fraction sampled the model load, so it answered the wrong question (#1290)
mudler
committed
5ad4158
merge: origin/main into the A2-Q1 landing (#810)
mudler
committed
bcb3964
fix(#1290): the host-arm divergence is an sm_110 result, and GB10 is a counter-measurement
mudler
committed
a483898
test(#810 A2-Q1): pin the per-token helper's refusals, because three of its four behaviours ARE refusals
mudler
committed
898d81a
fix(#810 A2-Q1): the "is this resolvable" diagnostic printed 1, because doctest stringifies char* as bool
mudler
committed
b5cb5e7

Navigation Menu