cuda: MoE prefill — fused SwiGLU epilogue and fused weighted reduction - #107

Open

bri-prism wants to merge 48 commits into

prismfrom

perf/moe-prefill-cda

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 17, 2026

Commits on Jul 18, 2026

ggml-cpu: x86 AVX512-VNNI repack GEMV/GEMM for Q1_0 and Q2_0 (#86)
bri-prism
authored
9fcaed7

Commits on Jul 20, 2026

force qwen35 to treat IGPU like a GPU so DSpark does not fall back to CPU (#88)
thadreber-web
authored
7529fda

Commits on Jul 31, 2026

server: do not propagate kv-mean-center bias to draft model contexts (#81)
cnndabbler
authored
4dd1656

Commits on Aug 3, 2026