Closed yzh119 closed 4 months ago
First step towards #199 .
Group gemm should also be helpful for MoE.
Test passed, I'll merge this PR first, for the next steps, we need to compile more shapes (for lora shrink and expand), and integrate punica's bgmv and sgmv implementations for extremes shapes (vector, etc).
First step towards #199 .
Group gemm should also be helpful for MoE.