diff --git a/Jenkinsfile b/Jenkinsfile
index 4350816013..96af656c16 100644
--- a/Jenkinsfile
+++ b/Jenkinsfile
@@ -855,13 +855,20 @@ def run_aiter_tests(Map conf=[:]){
     }
 
     withDockerContainer(image: image, args: dockerOpts) {
-        timeout(time: 45, unit: 'MINUTES'){
+        timeout(time: 2, unit: 'HOURS'){
             try{
                 sh "rocminfo"
                 sh "python3 --version"
                 sh "python3 /home/jenkins/workspace/aiter/op_tests/test_gemm_a8w8.py"
                 sh "python3 /home/jenkins/workspace/aiter/op_tests/test_gemm_a8w8_blockscale.py"
                 sh "python3 /home/jenkins/workspace/aiter/op_tests/test_mha.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe_2stage.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe_blockscale.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe_ep.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe_sorting.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe_sorting_mxfp4.py"
+                sh "python3 /home/jenkins/workspace/aiter/op_tests/test_moe_tkw1.py"
             }
             catch(e){
                 echo "Throwing error exception while running AITER tests"
@@ -906,7 +913,7 @@ def run_pytorch_tests(Map conf=[:]){
     }
 
     withDockerContainer(image: image, args: dockerOpts) {
-        timeout(time: 45, unit: 'MINUTES'){
+        timeout(time: 2, unit: 'HOURS'){
             try{
                 sh "rocminfo"
                 sh "python3 --version"
diff --git a/example/CMakeLists.txt b/example/CMakeLists.txt
index 7dc2f92bf9..7bd628edf2 100644
--- a/example/CMakeLists.txt
+++ b/example/CMakeLists.txt
@@ -1,7 +1,6 @@
 include_directories(BEFORE
     ${PROJECT_SOURCE_DIR}/include
     ${PROJECT_SOURCE_DIR}/library/include
-    ${PROJECT_SOURCE_DIR}/example/include
 )
 
 add_custom_target(examples)
diff --git a/example/ck_tile/01_fmha/fmha_bwd.cpp b/example/ck_tile/01_fmha/fmha_bwd.cpp
index 8ad2a3de04..cc2663e751 100644
--- 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>
diff --git a/example/ck_tile/01_fmha/fmha_fwd.cpp b/example/ck_tile/01_fmha/fmha_fwd.cpp
index f6b5b879bd..dd7c444557 100644
--- 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>
diff --git a/example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp b/example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
index 94e4734fb4..54f4e66336 100644
--- 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>
 
diff --git a/example/ck_tile/03_gemm/gemm_utils.hpp b/example/ck_tile/03_gemm/gemm_utils.hpp
index bf98eb655f..0a6e65c345 100644
--- 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
diff --git a/example/ck_tile/05_reduce/reduce.cpp b/example/ck_tile/05_reduce/reduce.cpp
index 7bae39b9d5..ea3253b629 100644
--- 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>
diff --git a/example/ck_tile/09_topk_softmax/topk_softmax.cpp b/example/ck_tile/09_topk_softmax/topk_softmax.cpp
index 8c1da293a6..0487bd05d2 100644
--- 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>
diff --git a/example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp b/example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp
index bf6c4fc68e..6e2664e9ba 100644
--- 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>
diff --git 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
index 919767a129..ae14be026b 100644
--- 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>
diff --git a/example/ck_tile/12_smoothquant/smoothquant.cpp b/example/ck_tile/12_smoothquant/smoothquant.cpp
index b54babdce3..50da06f204 100644
--- 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
diff --git a/example/ck_tile/13_moe_sorting/moe_sorting.cpp b/example/ck_tile/13_moe_sorting/moe_sorting.cpp
index ef1edadf7a..b2ad4eb98c 100644
--- 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[])
 {
diff --git a/example/ck_tile/14_moe_smoothquant/moe_smoothquant.cpp b/example/ck_tile/14_moe_smoothquant/moe_smoothquant.cpp
index 864ff31798..9cd77ca5fa 100644
--- 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>
 
diff --git a/example/ck_tile/15_fused_moe/main.cpp b/example/ck_tile/15_fused_moe/main.cpp
index 3c459c6b95..b6cc3b6543 100644
--- 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
diff --git a/example/ck_tile/16_batched_gemm/batched_gemm.hpp b/example/ck_tile/16_batched_gemm/batched_gemm.hpp
index c11453b0d1..0da6501568 100644
--- 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
diff --git a/example/ck_tile/17_grouped_gemm/grouped_gemm.hpp b/example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
index 7a8b670123..39af33ebab 100644
--- 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
diff --git a/example/ck_tile/18_flatmm/run_flatmm_example.inc b/example/ck_tile/18_flatmm/run_flatmm_example.inc
index e526ddc3f5..b6b92b5801 100644
--- 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()
 {
diff --git 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
index 2b388768b0..366eb8c81a 100644
--- 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,
diff --git a/example/ck_tile/21_elementwise/elementwise_example.cpp b/example/ck_tile/21_elementwise/elementwise_example.cpp
index 545617a212..511449c0de 100644
--- 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[])
diff --git a/example/ck_tile/21_elementwise/elementwise_example_add_4d.cpp b/example/ck_tile/21_elementwise/elementwise_example_add_4d.cpp
index a9f2ea422c..a7b9fb494d 100644
--- 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[])
diff --git a/example/ck_tile/21_elementwise/elementwise_example_transpose.cpp b/example/ck_tile/21_elementwise/elementwise_example_transpose.cpp
index ad31ee0c62..abf2435221 100644
--- 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[])
diff --git a/example/ck_tile/21_elementwise/elementwise_example_unary.cpp b/example/ck_tile/21_elementwise/elementwise_example_unary.cpp
index f39446d351..a1804ea52e 100644
--- 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[])
diff --git a/example/ck_tile/35_batched_transpose/batched_transpose_example.cpp b/example/ck_tile/35_batched_transpose/batched_transpose_example.cpp
index 74441a149a..79013594ec 100644
--- 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)
diff --git a/example/include/json_dump.hpp b/include/ck_tile/utility/json_dump.hpp
similarity index 100%
rename from example/include/json_dump.hpp
rename to include/ck_tile/utility/json_dump.hpp