[CK_TILE] moe sorting ex kernel to support expert > 128 (#1840)

* moe sorting ex

* fix bug for race condition

* fix bug and optimze large expert

* fix

* optimize with sub_token_oneshot

* support skip empty tokens for expert sorting

* update moe_sorting

* tidy code
This commit is contained in:
carlushuang
2025-02-11 17:49:17 +08:00
committed by GitHub
parent 2312f4aaf9
commit c0adab4850
12 changed files with 936 additions and 99 deletions

View File

@@ -7,6 +7,7 @@
#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp"
#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp"
#include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp"
#include "ck_tile/ops/fused_moe/kernel/moe_sorting_problem.hpp"
#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp"
#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp"
#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp"
@@ -14,7 +15,6 @@
#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
#include "ck_tile/ops/common/utils.hpp"