Merge branch 'main' into export-D103467782

Author: rascani

.github/workflows/pull.yml

@@ -534,8 +534,8 @@ jobs:
         output=$(ls -la cmake-out/test/size_test)
         arr=($output)
         size=${arr[4]}
-        # Current CI size: 48008 (gcc9-nopytorch, 2026-03-06)
-        threshold="48500"
+        # Current CI size: 52168 (gcc9-nopytorch, 2026-07-06)
+        threshold="52168"
         if [[ "$size" -le "$threshold" ]]; then
           echo "Success $size <= $threshold"
         else

CMakeLists.txt

@@ -157,6 +157,13 @@ if(NOT PYTHON_EXECUTABLE)
 endif()
 announce_configured_options(PYTHON_EXECUTABLE)
 
+# pybind11 >=3.0 uses find_package(Python) which looks for Python_EXECUTABLE,
+# not the legacy PYTHON_EXECUTABLE variable. Bridge the two so pybind11 picks up
+# the interpreter that setup.py (or the user) specified.
+if(PYTHON_EXECUTABLE AND NOT Python_EXECUTABLE)
+  set(Python_EXECUTABLE "${PYTHON_EXECUTABLE}")
+endif()
+
 announce_configured_options(CMAKE_CXX_COMPILER_ID)
 announce_configured_options(CMAKE_TOOLCHAIN_FILE)
 announce_configured_options(BUILD_TESTING)
@@ -260,6 +267,27 @@ endif()
 
 add_subdirectory(third-party)
 
+# On Apple, pybind11 SHARED extension modules must not link libpython directly.
+# When python_add_library creates a SHARED library, it links Python::Python (via
+# pybind11::embed) which pulls in libpython. This conflicts with pybind11
+# >=3.0's multi-phase module init — the duplicate Python runtime causes a GIL
+# state crash at import time. This function replaces pybind11::embed with
+# pybind11::module (which links Python::Module instead of Python::Python —
+# headers and ABI only, no libpython) and adds -undefined dynamic_lookup for
+# symbol resolution. No-op on non-Apple platforms.
+function(strip_python_lib target)
+  if(NOT APPLE)
+    return()
+  endif()
+  get_target_property(_libs ${target} LINK_LIBRARIES)
+  if(_libs)
+    list(REMOVE_ITEM _libs Python::Python pybind11::embed)
+    list(APPEND _libs pybind11::module)
+    set_target_properties(${target} PROPERTIES LINK_LIBRARIES "${_libs}")
+  endif()
+  target_link_options(${target} PRIVATE "LINKER:-undefined,dynamic_lookup")
+endfunction()
+
 # Size-optimized builds disable exceptions, RTTI, and unwind tables.
 # ExecuTorch's runtime uses Result<T>/Error instead of exceptions, so
 # -fno-exceptions is safe. Suppressing .eh_frame via -fno-unwind-tables is safe
@@ -1114,6 +1142,7 @@ if(EXECUTORCH_BUILD_PYBIND)
 
   # pybind portable_lib
   pybind11_add_module(portable_lib SHARED extension/pybindings/pybindings.cpp)
+  strip_python_lib(portable_lib)
   # The actual output file needs a leading underscore so it can coexist with
   # portable_lib.py in the same python package. PyTorch requires C++20, so
   # pybindings must be compiled with C++20.
@@ -1155,6 +1184,7 @@ if(EXECUTORCH_BUILD_PYBIND)
   pybind11_add_module(
     data_loader SHARED extension/pybindings/pybindings_data_loader.cpp
   )
+  strip_python_lib(data_loader)
   target_include_directories(data_loader PRIVATE ${_common_include_directories})
   target_compile_options(data_loader PUBLIC ${_pybind_compile_options})
   target_link_libraries(data_loader PRIVATE executorch)

backends/apple/coreml/CMakeLists.txt

@@ -175,6 +175,7 @@ if(EXECUTORCH_BUILD_PYBIND)
     executorchcoreml SHARED runtime/inmemoryfs/inmemory_filesystem_py.cpp
     runtime/inmemoryfs/inmemory_filesystem_utils.cpp
   )
+  strip_python_lib(executorchcoreml)
   target_link_libraries(
     executorchcoreml PRIVATE coreml_util coreml_inmemoryfs
                              nlohmann_json::nlohmann_json

codegen/tools/CMakeLists.txt

@@ -9,6 +9,7 @@
 
 # Create the selective_build pybind11 module
 pybind11_add_module(selective_build SHARED selective_build.cpp)
+strip_python_lib(selective_build)
 
 # Set the output name to match the module name
 set_target_properties(selective_build PROPERTIES OUTPUT_NAME "selective_build")

examples/arm/image_classification_example_vgf/requirements.txt

@@ -4,5 +4,5 @@
 # LICENSE file in the root directory of this source tree.
 
 evaluate == 0.4.6
-transformers[torch] == 4.56.1 
+transformers[torch] == 5.3.0 
 tqdm == 4.67.1

extension/llm/runner/CMakeLists.txt

@@ -112,6 +112,7 @@ if(EXECUTORCH_BUILD_PYBIND)
   pybind11_add_module(
     _llm_runner SHARED ${CMAKE_CURRENT_SOURCE_DIR}/pybindings.cpp
   )
+  strip_python_lib(_llm_runner)
 
   find_package_torch()
   find_library(

extension/training/CMakeLists.txt

@@ -66,6 +66,7 @@ if(EXECUTORCH_BUILD_PYBIND)
     _training_lib SHARED
     ${CMAKE_CURRENT_SOURCE_DIR}/pybindings/_training_lib.cpp
   )
+  strip_python_lib(_training_lib)
   set_target_properties(_training_lib PROPERTIES CXX_STANDARD 20)
 
   target_include_directories(_training_lib PRIVATE ${TORCH_INCLUDE_DIRS})