Fix latest AITER failure and add more AITER tests in CK CI. (#2782)

* add aiter tests and move json_dump header * remove example/include path from cmake * extend time for aiter and pytorch stages [ROCm/composable_kernel commit: ef6c28e989]
2026-05-18 20:09:25 +00:00 · 2025-09-04 13:44:00 -07:00
parent e4ac6bca96
commit 80cc2b7bca
24 changed files with 30 additions and 24 deletions
--- a/example/ck_tile/01_fmha/fmha_bwd.cpp
+++ b/example/ck_tile/01_fmha/fmha_bwd.cpp
@@ -5,7 +5,7 @@
 #include "ck_tile/host.hpp"
 #include "mask.hpp"
 #include "utils.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 #include <array>
 #include <cstring>
--- a/example/ck_tile/01_fmha/fmha_fwd.cpp
+++ b/example/ck_tile/01_fmha/fmha_fwd.cpp
@@ -7,7 +7,7 @@
 #include "mask.hpp"
 #include "rotary.hpp"
 #include "utils.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 #include <array>
 #include <cstring>
--- a/example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
+++ b/example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
@@ -1,6 +1,6 @@
 #include "ck_tile/host.hpp"
 #include "layernorm2d_fwd.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include <algorithm>
 #include <cstring>

--- a/example/ck_tile/03_gemm/gemm_utils.hpp
+++ b/example/ck_tile/03_gemm/gemm_utils.hpp
@@ -9,7 +9,7 @@
 #include "ck_tile/host/kernel_launch.hpp"
 #include "ck_tile/ops/epilogue.hpp"
 #include "ck_tile/ops/gemm.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 #define CK_TILE_PIPELINE_COMPUTE_V3 1
 #define CK_TILE_PIPELINE_MEMORY 2
--- a/example/ck_tile/05_reduce/reduce.cpp
+++ b/example/ck_tile/05_reduce/reduce.cpp
@@ -3,7 +3,7 @@

 #include "ck_tile/host.hpp"
 #include "ck_tile/ops/reduce.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include <cstring>

 template <typename T>
--- a/example/ck_tile/09_topk_softmax/topk_softmax.cpp
+++ b/example/ck_tile/09_topk_softmax/topk_softmax.cpp
@@ -13,7 +13,7 @@
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/reduce.hpp"
 #include "topk_softmax_api.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 #if 0
 template <typename T>
--- a/example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp
+++ b/example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp
@@ -1,7 +1,7 @@
 #include "ck_tile/host.hpp"
 #include "rmsnorm2d_fwd.hpp"
 #include <cstring>
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 // different threshold for different dtype
 template <typename DataType>
--- a/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp
+++ b/example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp
@@ -1,7 +1,7 @@
 #include "ck_tile/host.hpp"
 #include "add_rmsnorm2d_rdquant_fwd.hpp"
 #include <cstring>
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 // different threshold for different dtype
 template <typename InputDataType>
--- a/example/ck_tile/12_smoothquant/smoothquant.cpp
+++ b/example/ck_tile/12_smoothquant/smoothquant.cpp
@@ -1,6 +1,6 @@
 #include "ck_tile/host.hpp"
 #include "smoothquant.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include <cstring>

 // different threshold for different dtype
--- a/example/ck_tile/13_moe_sorting/moe_sorting.cpp
+++ b/example/ck_tile/13_moe_sorting/moe_sorting.cpp
@@ -14,7 +14,7 @@
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/reduce.hpp"
 #include "moe_sorting_api.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 auto create_args(int argc, char* argv[])
 {
--- a/example/ck_tile/14_moe_smoothquant/moe_smoothquant.cpp
+++ b/example/ck_tile/14_moe_smoothquant/moe_smoothquant.cpp
@@ -1,6 +1,6 @@
 #include "ck_tile/host.hpp"
 #include "moe_smoothquant.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include <cstring>
 #include <set>

--- a/example/ck_tile/15_fused_moe/main.cpp
+++ b/example/ck_tile/15_fused_moe/main.cpp
@@ -5,7 +5,7 @@
 #include <set>

 #include "ck_tile/host.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include "fused_moe.hpp"

 // different threshold for different dtype
--- a/example/ck_tile/16_batched_gemm/batched_gemm.hpp
+++ b/example/ck_tile/16_batched_gemm/batched_gemm.hpp
@@ -9,7 +9,7 @@
 #include "ck_tile/host/kernel_launch.hpp"
 #include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
 #include "ck_tile/ops/elementwise/unary_element_wise_operation.hpp"
-#include <json_dump.hpp>
+#include "ck_tile/utility/json_dump.hpp"

 #define CK_TILE_PIPELINE_COMPUTE_V3 1
 #define CK_TILE_PIPELINE_MEMORY 2
--- a/example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
+++ b/example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
@@ -9,7 +9,7 @@
 #include "ck_tile/host/kernel_launch.hpp"
 #include "ck_tile/ops/gemm.hpp"
 #include "ck_tile/ops/elementwise/unary_element_wise_operation.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 #define CK_TILE_PIPELINE_COMPUTE_V3 1
 #define CK_TILE_PIPELINE_MEMORY 2
--- a/example/ck_tile/18_flatmm/run_flatmm_example.inc
+++ b/example/ck_tile/18_flatmm/run_flatmm_example.inc
@@ -2,7 +2,7 @@
 // Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
 #pragma once
 #include <type_traits>
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 template <typename T>
 constexpr const char* DataTypeToString()
 {
--- a/example/ck_tile/19_gemm_multi_d/run_gemm_multi_d_fp16_example.inc
+++ b/example/ck_tile/19_gemm_multi_d/run_gemm_multi_d_fp16_example.inc
@@ -3,7 +3,7 @@

 #pragma once
 #include <cstddef>
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"

 template <typename ADataType,
          typename BDataType,
--- a/example/ck_tile/21_elementwise/elementwise_example.cpp
+++ b/example/ck_tile/21_elementwise/elementwise_example.cpp
@@ -4,7 +4,7 @@
 #include "ck_tile/host.hpp"
 #include "ck_tile/ops/elementwise.hpp"
 #include "ck_tile/host/reference/reference_elementwise.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include "elementwise_common.hpp"

 auto create_args(int argc, char* argv[])
--- a/example/ck_tile/21_elementwise/elementwise_example_add_4d.cpp
+++ b/example/ck_tile/21_elementwise/elementwise_example_add_4d.cpp
@@ -4,7 +4,7 @@
 #include "ck_tile/host.hpp"
 #include "ck_tile/ops/elementwise.hpp"
 #include "ck_tile/host/reference/reference_elementwise.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include "elementwise_common.hpp"

 auto create_args(int argc, char* argv[])
--- a/example/ck_tile/21_elementwise/elementwise_example_transpose.cpp
+++ b/example/ck_tile/21_elementwise/elementwise_example_transpose.cpp
@@ -4,7 +4,7 @@
 #include "ck_tile/host.hpp"
 #include "ck_tile/ops/elementwise.hpp"
 #include "ck_tile/host/reference/reference_transpose.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include "elementwise_common.hpp"

 auto create_args(int argc, char* argv[])
--- a/example/ck_tile/21_elementwise/elementwise_example_unary.cpp
+++ b/example/ck_tile/21_elementwise/elementwise_example_unary.cpp
@@ -4,7 +4,7 @@
 #include "ck_tile/host.hpp"
 #include "ck_tile/ops/elementwise.hpp"
 #include "ck_tile/host/reference/reference_elementwise.hpp"
-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #include "elementwise_common.hpp"

 auto create_args(int argc, char* argv[])
--- a/example/ck_tile/35_batched_transpose/batched_transpose_example.cpp
+++ b/example/ck_tile/35_batched_transpose/batched_transpose_example.cpp
@@ -12,7 +12,7 @@

 #include "batched_transpose_example.hpp"

-#include "json_dump.hpp"
+#include "ck_tile/utility/json_dump.hpp"
 #if 0
 template <typename T>
 void dump_host_tensor_4d(const ck_tile::HostTensor<T>& x)