Make _get_program_from_buffer work for bundled programs take 2

CMakeLists.txt

@@ -738,9 +738,6 @@ endif()
 
 if(EXECUTORCH_BUILD_PYBIND)
 
-  # Add codegen tools subdirectory for selective_build pybind module
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/codegen/tools)
-
   if(NOT EXECUTORCH_BUILD_EXTENSION_DATA_LOADER)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/data_loader)
   endif()
@@ -749,6 +746,9 @@ if(EXECUTORCH_BUILD_PYBIND)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/devtools)
   endif()
 
+  # Add codegen tools subdirectory for selective_build pybind module
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/codegen/tools)
+
   # Create bundled_module target only for pybindings when bundled_program exists
   # This target has hard dependencies on devtools generated headers
   if(TARGET bundled_program)

codegen/tools/CMakeLists.txt

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2025 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -28,6 +29,10 @@ target_compile_options(
 )
 
 # Link against required libraries
+if(TARGET bundled_program)
+  target_compile_definitions(selective_build PRIVATE -DET_BUNDLE_IO)
+  target_link_libraries(selective_build PRIVATE bundled_program)
+endif()
 target_link_libraries(selective_build PRIVATE executorch_core program_schema)
 
 # Install the module

codegen/tools/selective_build.cpp

@@ -1,16 +1,21 @@
 /*
  * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
+ * Copyright 2025 Arm Limited and/or its affiliates.
  *
  * This source code is licensed under the BSD-style license found in the
  * LICENSE file in the root directory of this source tree.
  */
 
+#include <executorch/runtime/platform/assert.h>
+#include <executorch/schema/program_generated.h>
 #include <pybind11/pybind11.h>
 #include <pybind11/stl.h>
 
-#include <executorch/runtime/platform/assert.h>
-#include <executorch/schema/program_generated.h>
+#ifdef ET_BUNDLE_IO
+#include <executorch/devtools/bundled_program/bundled_program.h>
+#include <stdexcept>
+#endif
 
 namespace py = pybind11;
 
@@ -186,8 +191,39 @@ get_kernel_tensor_metadatas_from_execution_plan(
 
 const executorch_flatbuffer::Program* _get_program_from_buffer(
     const py::bytes& buffer) {
+  // Access the Python bytes without copying and get raw pointer/size.
+  const std::string_view sv = buffer.cast<std::string_view>();
+#ifdef ET_BUNDLE_IO
+  void* buf_ptr = const_cast<void*>(static_cast<const void*>(sv.data()));
+  const size_t buf_len = sv.size();
+
+  // If this is a bundled program, extract the inner ExecuTorch program bytes.
+  if (executorch::bundled_program::is_bundled_program(buf_ptr, buf_len)) {
+    const void* program_data = nullptr;
+    size_t program_size = 0;
+
+    const auto status = executorch::bundled_program::get_program_data(
+        buf_ptr, // serialized BundledProgram start
+        buf_len, // total size of the BundledProgram blob
+        &program_data, // [out] pointer to inner .pte bytes
+        &program_size // [out] size of inner .pte bytes
+    );
+
+    if (status != ::executorch::runtime::Error::Ok || program_data == nullptr ||
+        program_size == 0) {
+      throw std::runtime_error(
+          "bundled_program::get_program_data() failed or returned empty data");
+    }
+
+    // program_data points directly at the flatbuffer-encoded Program region.
+    return executorch_flatbuffer::GetProgram(
+        reinterpret_cast<const uint8_t*>(program_data));
+  }
+#endif
+  // Otherwise treat the buffer as a raw .pte (flatbuffer Program with optional
+  // extended header).
   return executorch_flatbuffer::GetProgram(
-      buffer.cast<std::string_view>().data());
+      reinterpret_cast<const uint8_t*>(sv.data()));
 }
 
 py::list _get_program_operators(const executorch_flatbuffer::Program* program) {

docs/source/backends-arm-ethos-u.md

@@ -268,8 +268,7 @@ You can see how  this coupling between the memory mode and runtime application i
 
 The arm_executor_runner supports [bundled-io](https://docs.pytorch.org/executorch/0.4/bundled-io.html) and [ETdump](https://docs.pytorch.org/executorch/stable/etdump.html) debugging tools.
 
-To enable bundled-io, set `EXECUTORCH_BUILD_DEVTOOLS` when building Executorch and `DET_BUNDLE_IO` when building the executor_runner. Currently using bundled-io requires specifying your
-non delegated Aten ops manually by setting `EXECUTORCH_SELECT_OPS_LIST`. To enable ETdump, set `EXECUTORCH_BUILD_ARM_ETDUMP` when building Executorch and `DEXECUTORCH_ENABLE_EVENT_TRACER`
+To enable bundled-io, set `EXECUTORCH_BUILD_DEVTOOLS` when building Executorch and `DET_BUNDLE_IO` when building the executor_runner. To enable ETdump, set `EXECUTORCH_BUILD_ARM_ETDUMP` when building Executorch and `DEXECUTORCH_ENABLE_EVENT_TRACER`
 when building the executor_runner.
 
 

examples/arm/executor_runner/CMakeLists.txt

@@ -235,10 +235,10 @@ list(
   -Map=arm_executor_runner.map
 )
 
-# Prefer to generate kernel bindings from model file if possible, which is when
-# 1. Not building for semihosting 2. Not building with bundleio If that is not
-# the case, fallback to select_ops_list If the model file does not contain any
-# aten ops, a workaround is currently needed to avoid crashing.
+# Figure out which ops to include: For semihosting build, use
+# (user-set)SELECT_OPS_MODEL variable. For normal build, use
+# EXECUTORCH_SELECT_OPS_MODEL to include ops automatically. If the pte contains
+# no undelegated ops, use neither.
 execute_process(
   COMMAND
     python "${ET_DIR_PATH}/codegen/tools/gen_oplist.py"
@@ -264,11 +264,6 @@ elseif(${FOUND_OPS_IN_FILE})
   message(
     "gen_oplist:  EXECUTORCH_SELECT_OPS_MODEL=${ET_PTE_FILE_PATH} is used to auto generate ops from"
   )
-elseif(NOT ${FOUND_OPS_IN_FILE} AND ${ET_BUNDLE_IO})
-  set(EXECUTORCH_SELECT_OPS_MODEL "")
-  message(
-    "gen_oplist: Building with ET_BUNDLE_IO and .bpte is not supported to auto generate ops from will use EXECUTORCH_SELECT_OPS_LIST=${EXECUTORCH_SELECT_OPS_LIST}"
-  )
 else()
   set(EXECUTORCH_SELECT_OPS_LIST "")
   set(EXECUTORCH_SELECT_OPS_MODEL "")

examples/arm/run.sh

@@ -53,8 +53,8 @@ function help() {
     echo "  --no_delegate                          Do not delegate the model (can't override builtin models)"
     echo "  --no_quantize                          Do not quantize the model (can't override builtin models)"
     echo "  --portable_kernels=<OPS>               TO BE DEPRECATED: Alias to select_ops_list."
-    echo "  --select_ops_list=<OPS>                Comma separated list of portable (non delegated) kernels to include Default: ${select_ops_list}"
-    echo "                                           NOTE: This is used when select_ops_model is not possible to use, e.g. for semihosting or bundleio."
+    echo "  --select_ops_list=<OPS>                Comma separated list of portable (non delagated) kernels to include Default: ${select_ops_list}"
+    echo "                                           NOTE: This is only used when building for semihosting."
     echo "                                           See https://docs.pytorch.org/executorch/stable/kernel-library-selective-build.html for more information."
     echo "  --target=<TARGET>                      Target to build and run for Default: ${target}"
     echo "  --output=<FOLDER>                      Target build output folder Default: ${output_folder}"