diff --git a/build/cmake_deps.toml b/build/cmake_deps.toml
index 4811563269c..b1ed81b6a7e 100644
--- a/build/cmake_deps.toml
+++ b/build/cmake_deps.toml
@@ -58,6 +58,21 @@ deps = [
   "executorch_core",
 ]
 
+# HACK: prevent reduce_util from also showing up in custom_ops. The
+# actual medium-term fix is to stop using Buck to drive our CMake
+# builds.
+[targets.reduce_util]
+buck_targets = [
+  "//kernels/portable/cpu/util:reduce_util",
+]
+filters = [
+  ".cpp$",
+]
+deps = [
+  "executorch",
+  "executorch_core",
+]
+
 [targets.optimized_kernels]
 buck_targets = [
   "//kernels/optimized:generated_lib",
@@ -414,6 +429,7 @@ deps = [
   "optimized_kernels",
   "extension_parallel",
   "extension_threadpool",
+  "reduce_util",
   "xnnpack_backend",
 ]
 
diff --git a/build/executorch-config.cmake b/build/executorch-config.cmake
index 2c459b66ac8..35fe03467f2 100644
--- a/build/executorch-config.cmake
+++ b/build/executorch-config.cmake
@@ -65,6 +65,7 @@ set(lib_list
     neuron_backend
     qnn_executorch_backend
     portable_ops_lib
+    custom_ops
     extension_module
     extension_module_static
     extension_parallel
@@ -128,3 +129,4 @@ if(TARGET cpublas)
     cpublas PROPERTIES INTERFACE_LINK_LIBRARIES extension_parallel
   )
 endif()
+target_compile_definitions(extension_threadpool INTERFACE ET_USE_THREADPOOL)
diff --git a/examples/models/llama/CMakeLists.txt b/examples/models/llama/CMakeLists.txt
index 5f49581ea25..96ff28d8f49 100644
--- a/examples/models/llama/CMakeLists.txt
+++ b/examples/models/llama/CMakeLists.txt
@@ -15,7 +15,7 @@
 # ~~~
 # It should also be cmake-lint clean.
 #
-cmake_minimum_required(VERSION 3.19)
+cmake_minimum_required(VERSION 3.24)  # 3.24 is required for WHOLE_ARCHIVE
 project(llama_runner)
 
 # Duplicating options as root CMakeLists.txt
@@ -84,14 +84,6 @@ if(CMAKE_TOOLCHAIN_IOS OR ANDROID)
   target_link_options_shared_lib(executorch)
 endif()
 
-# custom ops library
-if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
-  add_subdirectory(
-    ${CMAKE_CURRENT_SOURCE_DIR}/../../../extension/llm/custom_ops
-    ${CMAKE_CURRENT_BINARY_DIR}/../../../extension/llm/custom_ops
-  )
-endif()
-
 # llama_runner library
 add_subdirectory(runner)
 
@@ -119,8 +111,7 @@ target_link_options_shared_lib(quantized_ops_lib)
 list(APPEND link_libraries quantized_kernels quantized_ops_lib)
 
 if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
-  target_link_options_shared_lib(custom_ops)
-  list(APPEND link_libraries custom_ops)
+  list(APPEND link_libraries $<LINK_LIBRARY:WHOLE_ARCHIVE,custom_ops>)
 endif()
 
 if(EXECUTORCH_BUILD_TORCHAO)
@@ -140,7 +131,6 @@ endif()
 set(XNNPACK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../../backends/xnnpack)
 # Extra compile option and include dir for pthreadpool
 if(EXECUTORCH_BUILD_PTHREADPOOL)
-  list(APPEND _common_compile_options -DET_USE_THREADPOOL)
   list(APPEND link_libraries extension_threadpool pthreadpool)
   list(APPEND _common_include_directories
        ${XNNPACK_ROOT}/third-party/pthreadpool/include
diff --git a/examples/models/llama/main.cpp b/examples/models/llama/main.cpp
index 5fe0ce93cf6..e91d71e4dfb 100644
--- a/examples/models/llama/main.cpp
+++ b/examples/models/llama/main.cpp
@@ -13,6 +13,8 @@
 #if defined(ET_USE_THREADPOOL)
 #include <executorch/extension/threadpool/cpuinfo_utils.h>
 #include <executorch/extension/threadpool/threadpool.h>
+#else
+#error "threadpool is required"
 #endif
 
 DEFINE_string(
diff --git a/examples/models/llava/CMakeLists.txt b/examples/models/llava/CMakeLists.txt
index ecd00809fdb..5d5857dd5af 100644
--- a/examples/models/llava/CMakeLists.txt
+++ b/examples/models/llava/CMakeLists.txt
@@ -93,14 +93,6 @@ if(CMAKE_TOOLCHAIN_IOS OR ANDROID)
   target_link_options_shared_lib(executorch)
 endif()
 
-# custom ops library
-if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
-  add_subdirectory(
-    ${EXECUTORCH_ROOT}/extension/llm/custom_ops
-    ${CMAKE_CURRENT_BINARY_DIR}/../../../extension/llm/custom_ops
-  )
-endif()
-
 # llava_runner library
 add_subdirectory(runner)
 
@@ -132,14 +124,12 @@ target_link_options_shared_lib(quantized_ops_lib)
 list(APPEND link_libraries quantized_kernels quantized_ops_lib)
 
 if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
-  target_link_options_shared_lib(custom_ops)
   list(APPEND link_libraries custom_ops)
 endif()
 
 set(XNNPACK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../../backends/xnnpack)
 # Extra compile option and include dir for pthreadpool
 if(EXECUTORCH_BUILD_PTHREADPOOL)
-  list(APPEND _common_compile_options -DET_USE_THREADPOOL)
   list(APPEND link_libraries extension_threadpool pthreadpool)
   list(APPEND _common_include_directories
        ${XNNPACK_ROOT}/third-party/pthreadpool/include
diff --git a/examples/models/llava/main.cpp b/examples/models/llava/main.cpp
index b01b33f5dd8..22e40d64bc8 100644
--- a/examples/models/llava/main.cpp
+++ b/examples/models/llava/main.cpp
@@ -17,6 +17,8 @@
 #if defined(ET_USE_THREADPOOL)
 #include <executorch/extension/threadpool/cpuinfo_utils.h>
 #include <executorch/extension/threadpool/threadpool.h>
+#else
+#error "threadpool is required"
 #endif
 
 DEFINE_string(
diff --git a/examples/models/llava/targets.bzl b/examples/models/llava/targets.bzl
index 5efb099f06f..6f3a370acf4 100644
--- a/examples/models/llava/targets.bzl
+++ b/examples/models/llava/targets.bzl
@@ -7,9 +7,6 @@ def define_common_targets():
             "main.cpp",
         ],
         compiler_flags = ["-Wno-global-constructors"],
-        preprocessor_flags = [
-            "-DET_USE_THREADPOOL",
-        ],
         deps = [
             "//executorch/examples/models/llava/runner:runner",
             "//executorch/extension/evalue_util:print_evalue",
diff --git a/extension/android/CMakeLists.txt b/extension/android/CMakeLists.txt
index 70f21f2751c..ba722d9c791 100644
--- a/extension/android/CMakeLists.txt
+++ b/extension/android/CMakeLists.txt
@@ -4,7 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-cmake_minimum_required(VERSION 3.19)
+cmake_minimum_required(VERSION 3.24) # 3.24 is required for WHOLE_ARCHIVE
 
 project(executorch_jni)
 
@@ -115,16 +115,10 @@ if(TARGET vulkan_backend)
 endif()
 
 if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
-  add_subdirectory(
-    ${EXECUTORCH_ROOT}/extension/llm/custom_ops
-    ${CMAKE_CURRENT_BINARY_DIR}/../../extension/llm/custom_ops
-  )
-  list(APPEND link_libraries custom_ops)
-  target_link_options_shared_lib(custom_ops)
+  list(APPEND link_libraries $<LINK_LIBRARY:WHOLE_ARCHIVE,custom_ops>)
 endif()
 
 if(TARGET pthreadpool)
-  target_compile_definitions(executorch_jni PRIVATE ET_USE_THREADPOOL=1)
   target_include_directories(
     executorch_jni
     PUBLIC
diff --git a/extension/android/jni/jni_layer_llama.cpp b/extension/android/jni/jni_layer_llama.cpp
index 5c58c89ee91..861497f333b 100644
--- a/extension/android/jni/jni_layer_llama.cpp
+++ b/extension/android/jni/jni_layer_llama.cpp
@@ -24,6 +24,8 @@
 #if defined(ET_USE_THREADPOOL)
 #include <executorch/extension/threadpool/cpuinfo_utils.h>
 #include <executorch/extension/threadpool/threadpool.h>
+#else
+#error "threadpool is required"
 #endif
 
 #include <fbjni/ByteBuffer.h>
diff --git a/extension/llm/custom_ops/CMakeLists.txt b/extension/llm/custom_ops/CMakeLists.txt
index c3969e6f9bf..eeb118d4344 100644
--- a/extension/llm/custom_ops/CMakeLists.txt
+++ b/extension/llm/custom_ops/CMakeLists.txt
@@ -78,7 +78,7 @@ target_include_directories(
 target_link_libraries(custom_ops PUBLIC ${custom_ops_libs} executorch_core)
 
 target_compile_options(
-  custom_ops PUBLIC ${_common_compile_options} -DET_USE_THREADPOOL
+  custom_ops PUBLIC ${_common_compile_options}
 )
 
 install(TARGETS custom_ops DESTINATION lib)
@@ -130,7 +130,7 @@ if(EXECUTORCH_BUILD_KERNELS_CUSTOM_AOT)
   target_compile_options(
     custom_ops_aot_lib
     PUBLIC -Wno-deprecated-declarations -fPIC -frtti -fexceptions
-           ${_common_compile_options} -DET_USE_THREADPOOL
+           ${_common_compile_options}
   )
 
   install(TARGETS custom_ops_aot_lib
diff --git a/extension/llm/custom_ops/op_sdpa.cpp b/extension/llm/custom_ops/op_sdpa.cpp
index f0a7775e803..afaa85534fc 100644
--- a/extension/llm/custom_ops/op_sdpa.cpp
+++ b/extension/llm/custom_ops/op_sdpa.cpp
@@ -21,6 +21,8 @@
 #ifdef ET_USE_THREADPOOL
 #include <executorch/extension/parallel/thread_parallel.h>
 #include <executorch/extension/threadpool/threadpool.h>
+#else
+#error "threadpool is required"
 #endif
 #include <executorch/extension/kernel_util/make_boxed_from_unboxed_functor.h>
 
diff --git a/extension/threadpool/CMakeLists.txt b/extension/threadpool/CMakeLists.txt
index 90288656674..c1d86acf75d 100644
--- a/extension/threadpool/CMakeLists.txt
+++ b/extension/threadpool/CMakeLists.txt
@@ -32,6 +32,7 @@ target_include_directories(
   PUBLIC ${EXECUTORCH_ROOT}/backends/xnnpack/third-party/cpuinfo/include
          ${EXECUTORCH_ROOT}/backends/xnnpack/third-party/pthreadpool/include
 )
+target_compile_definitions(extension_threadpool PUBLIC ET_USE_THREADPOOL)
 target_compile_options(extension_threadpool PUBLIC ${_common_compile_options})
 
 # Install libraries