diff --git a/include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp b/include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp index 15c4c21c86..946b81c146 100644 --- a/include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp +++ b/include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp @@ -1264,7 +1264,7 @@ struct MoeFlatmmKernel scatter_token_id = scatter_token_id * kargs.TopK + (fused_token >> token_id_offset); c_scatter_offsets[mIter][m0] = scatter_token_id * kargs.stride_C; - c_scatter_valids[mIter][m0] = (scatter_token_id < kargs.NumTokens); + c_scatter_valids[mIter][m0] = (scatter_token_id < (kargs.NumTokens * (IsInputGemm? kargs.TopK : 1))); }); });