CUDA: dedup MoE gate/up activation quantization - #25441

Merged

JohannesGaessler merged 8 commits into

ggml-org:masterfrom

praneshgo:pgonegandla/moe_quant_optimize

Jul 16, 2026

Commits on Jul 8, 2026

CUDA: dedup MoE gate/up activation quantization (fp4)
praneshgo
committed
6f286cf
CUDA: bounds-check token ids in build_tok2c_kernel
praneshgo
committed
a155c70

Commits on Jul 9, 2026

Refactor the code based on review comments
praneshgo
committed
1327f42

Commits on Jul 11, 2026

Adding q8_1 support for dedup and addressing review comments
praneshgo
committed
1c687ce

Commits on Jul 14, 2026

Add pragma unrolls
praneshgo
committed
5fda553
Remove redundant cudaMemsetAsync call
praneshgo
committed
afd0239
Removing follow up redundancies
praneshgo
committed
b93a436

Commits on Jul 15, 2026

Merge branch 'master' of https://github.com/ggml-org/llama.cpp into pgonegandla/moe_quant_optimize
praneshgo
committed
bc67b43