xuzhenqi comments

Repositories
Issues
Comments

Results 2 comments of


                                            xuzhenqi

[QST] how to use groupwise scaling along M for FP8 gemm to impelement per-token-per-128-channel and blockwise?

@yizhang2077 I am also working on this using similar codes, which can be compiled successfully, but produce incorrect results. Have you solved this problem?

[QST] how to use groupwise scaling along M for FP8 gemm to impelement per-token-per-128-channel and blockwise?

Vllm's version works.