Optimize CUDA matmul for empty shards and MMQ dispatch#22170

Closed

nisparks wants to merge 2 commits intoggml-org:masterfrom

nisparks:pr3-mmq-cpasync

Commits on Apr 20, 2026

cuda: guard empty matmul shards and skip zero-sized RS no-ops

nisparks
and
Copilot
committed
cuda: gate MMQ cp.async to tensor-split MoE

nisparks
and
Copilot
committed