[OMNIML-5899] Add CUDA kernels for IQ packing - #2448

Merged

cjluo-nv merged 13 commits into

mainfrom

hungyuehc/omniml-5899-kernel-v2

Sep 17, 2026

Commits on Sep 16, 2026

Add CUDA kernels for IQ packing
hychiang-git
committed
51b6c19
Document IQ CUDA format constants
hychiang-git
committed
b39dca3
Clarify IQ CUDA extension failure messages
hychiang-git
committed
d554c36
Merge branch 'main' into hungyuehc/omniml-5899-kernel-v2
hychiang-git
authored
d2eacfb
Enforce per-row IQ CUDA block alignment
hychiang-git
committed
333ef0e
Test IQ CUDA extension boundaries
hychiang-git
committed
b8b3047
Merge branch 'main' into hungyuehc/omniml-5899-kernel-v2
hychiang-git
committed
1b4f9e6
Deduplicate IQ extension test cases
hychiang-git
committed
46afe88

Commits on Sep 17, 2026

Harden IQ CUDA packing contract
hychiang-git
committed
7643d73
Share IQ CUDA packing helpers and harden the input contract

cjluo-nv
and
claude
committed
74240e1
Scope the GGML attribution to the constants it covers

cjluo-nv
and
claude
committed
74e94db
Validate the IQ kernels' encoding, not just their boundaries

cjluo-nv
and
claude
committed
bb35d00
Align the reduction helpers' barriers and tighten the IQ2_XS scale contract

cjluo-nv
and
claude
committed
7cc1220