Hi @Junkai-Wu , the FMHA examples looks awesome. I wonder how hard would it be to adapt Hopper FMHA (example 88) to support MLA attn (i.e., v head dim != qk head dim) by utilizing the templates?
There seems to be a SM100 impl in FlashMLA PR!76, so I wonder if it is also applicable to support SM90 in a similar way.
Hi @Junkai-Wu , the FMHA examples looks awesome. I wonder how hard would it be to adapt Hopper FMHA (example 88) to support MLA attn (i.e., v head dim != qk head dim) by utilizing the templates?
There seems to be a SM100 impl in FlashMLA PR!76, so I wonder if it is also applicable to support SM90 in a similar way.