diff --git a/include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp b/include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp index 830662216a..926f49fb4e 100644 --- a/include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp +++ b/include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp @@ -1682,6 +1682,8 @@ struct MoeSortingMultiPhaseKernel_P2 if(position < kargs.num_experts) { + index_t *p_sorted_expert_cnts = p_total_tokens_post_pad + 1 + position;//temp mock for p_sorted_expert_cnts, fixme:felix + p_sorted_expert_cnts[0] = out_0; p_expert_cumsum[position] = out_0 * kargs.unit_size_mdiv.divisor; } @@ -1710,6 +1712,7 @@ struct MoeSortingMultiPhaseKernel_P2 { auto total_tokens_post_pad = prev_cumsum_a * kargs.unit_size_mdiv.divisor; p_total_tokens_post_pad[0] = total_tokens_post_pad; + p_total_tokens_post_pad[kargs.num_experts+1] = prev_cumsum_a; //temp mock for p_sorted_expert_cnts, fixme:felix p_expert_cumsum[kargs.num_experts] = total_tokens_post_pad; } }