You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The currently fp8 kernel can only support standard matmul (A~[M, K], B~[K, N]), however, MoE usually implemented as a batched matmul (A~[B, M, K], [B, K, N]) where B is the size of expert, and the current fp8 linear does not support this batched matmul. So I am wondering is there any plan to support the fp8 batched matmul?
The text was updated successfully, but these errors were encountered:
The currently fp8 kernel can only support standard matmul (A~[M, K], B~[K, N]), however, MoE usually implemented as a batched matmul (A~[B, M, K], [B, K, N]) where B is the size of expert, and the current fp8 linear does not support this batched matmul. So I am wondering is there any plan to support the fp8 batched matmul?
The text was updated successfully, but these errors were encountered: