Matt Wong

icon location Austin, TX icon location Stanford '23 BS Math + MS EE. Now working on ML frameworks at AMD. I love hiking and, to a lesser extent, working on my beater. Sometimes at the same time :--]

Results 1 issues of Matt Wong

[Kernel] Layernorm performance optimization

1 comment

This PR primarily creates optimized specializations of fused_add_rms_norm_kernel, used in many layernorms. It also includes a slightly optimized version of blockReduceSum/warpReduceSum which slightly reduce the number of shuffles done when...

action-required