ggml-cpu: enable Q2_0 VNNI kernel on AVX-VNNI-only CPUs - #76

Open

gondoi wants to merge 40 commits into

PrismML-Eng:prismfrom

gondoi:cpu-avx-vnni-q2_0

Commits on Jun 8, 2026

Optimized ARM NEON q1_0 dot (#33)
pl752
authored
64a6d4f
Q1_0 repack kernels for Arm NEON+DP (#34)

pl752
and
Copilot
authored
720e06b

Commits on Jun 9, 2026

Add release-prism workflow
khosravipasha
committed
ac71e9f
Add Q2_0 quantization: type definition and CPU backend
khosravipasha
committed
984bf97
Add Q2_0 Metal backend
khosravipasha
committed
f51b5aa

Commits on Jun 10, 2026

cuda: Q2_0
khosravipasha
committed
9c0edea
release-prism: install spirv-headers for ubuntu-arm64 vulkan build
khosravipasha
committed
5f33471
vulkan: Q2_0

Vort3xed
authored and
khosravipasha
committed
34dc581
fix: fp32_to_fp16 saturates finite overflow to inf

Vort3xed
authored and
khosravipasha
committed
98e900b
ggml-cpu: AVX-512-VNNI dot-products for Q1_0/Q2_0 (#37)

authored and
khosravipasha
committed
0e0353d
cpu: move Q2_0 AVX-512-VNNI dot into arch/x86, restore generic scalar

bri-prism
authored and
khosravipasha
committed
4d88cd4

Commits on Jun 12, 2026

cuda: opt-in Hopper wgmma path for Q1_0 prefill (sm_90a)
bri-prism
committed
21eff18
cuda: extend Hopper wgmma path to Q2_0
bri-prism
committed
6f25c34
cuda: hopper path review follow-up — multi-GPU correctness + hardening
bri-prism
committed
02860ff
cuda: clang-format the hopper-q1 path
bri-prism
committed
fdb74a0
cuda: reject Blackwell in the hopper wgmma gate (#47)
bri-prism
committed
3ff0bfc

Commits on Jul 7, 2026

ci(release): resolve OpenMP redist DLL by glob instead of pinned MSVC version (#50)

khosravipasha
and
claude
committed
6700b53
ci(release): disable LLAMA_BUILD_APP in ios-xcode-build

khosravipasha
and
claude
committed
0ad1dab

Commits on Jul 9, 2026

kv-cache: optional per-channel K-cache mean-centering for Q4_0 (#51)
bri-prism
authored
afc74b7

Commits on Jul 10, 2026

Commits on Jul 12, 2026

Commits on Jul 13, 2026

Commits on Jul 14, 2026

Commits on Jul 15, 2026

ggml-cpu: enable Q2_0 VNNI kernel on AVX-VNNI-only CPUs

gondoi
and
claude
committed
d152571