[webgpu-native] opt matmulnbits

Open qjia7 opened this issue 1 year ago • 2 comments

Oct 17 '24 02:10 qjia7

All native cases can pass now. @guschmue @fs-eire Please take a look, thanks.

Oct 17 '24 10:10 qjia7

also tested and did some perf comparison - yeah, it's good on Xe: Phi3 on tlk token/sec went up from 8.5 -> 13.1

Oct 21 '24 04:10 guschmue