Merge branch 'main' into update-module-java

Author: kirklandsign

CMakeLists.txt

@@ -48,21 +48,33 @@ project(executorch)
 # MARK: - Start EXECUTORCH_H12025_BUILD_MIGRATION --------------------------------------------------
 
 include(${PROJECT_SOURCE_DIR}/tools/cmake/common/preset.cmake)
+include(${PROJECT_SOURCE_DIR}/tools/cmake/Utils.cmake)
+include(CMakeDependentOption)
+include(ExternalProject)
 
 if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
 endif()
 announce_configured_options(CMAKE_CXX_STANDARD)
 
+if(NOT CMAKE_SYSTEM_PROCESSOR)
+  set(CMAKE_SYSTEM_PROCESSOR ${CMAKE_HOST_SYSTEM_PROCESSOR})
+endif()
+announce_configured_options(CMAKE_SYSTEM_PROCESSOR)
+
 if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Debug)
 endif()
 announce_configured_options(CMAKE_BUILD_TYPE)
 
+if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+endif()
+announce_configured_options(PYTHON_EXECUTABLE)
+
 announce_configured_options(CMAKE_CXX_COMPILER_ID)
 announce_configured_options(CMAKE_TOOLCHAIN_FILE)
 announce_configured_options(BUCK2)
-announce_configured_options(PYTHON_EXECUTABLE)
 
 load_build_preset()
 include(${PROJECT_SOURCE_DIR}/tools/cmake/preset/default.cmake)
@@ -72,10 +84,6 @@ print_configured_options()
 
 # MARK: - End EXECUTORCH_H12025_BUILD_MIGRATION ----------------------------------------------------
 
-include(tools/cmake/Utils.cmake)
-include(CMakeDependentOption)
-include(ExternalProject)
-
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # Setup RPATH.
@@ -251,11 +259,6 @@ if(EXECUTORCH_BUILD_TESTS)
   include(CTest)
 endif()
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-message(STATUS "Using python executable '${PYTHON_EXECUTABLE}'")
-
 # TODO(dbort): Fix these warnings and remove this flag.
 set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 

backends/apple/mps/CMakeLists.txt

@@ -18,10 +18,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 set(_common_compile_options -Wno-deprecated-declarations)
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 

backends/cadence/CMakeLists.txt

@@ -30,10 +30,6 @@ add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
 if(EXECUTORCH_CADENCE_CPU_RUNNER)
   include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-  if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-  endif()
-
   set(_common_compile_options -Wno-deprecated-declarations -fPIC)
 
   # Find prebuilt libraries. executorch package should contain portable_ops_lib,

backends/cadence/fusion_g3/operators/CMakeLists.txt

@@ -14,10 +14,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"

backends/cadence/hifi/operators/CMakeLists.txt

@@ -14,10 +14,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/backends/cadence/hifi/operators/op_add.cpp"

backends/cadence/reference/operators/CMakeLists.txt

@@ -14,10 +14,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # ATen compliant ops that are needed to run this model.
 set(_aten_ops__srcs
     "${CMAKE_CURRENT_SOURCE_DIR}/op_add.cpp"

backends/cortex_m/CMakeLists.txt

@@ -23,10 +23,6 @@ endif()
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # Cortex-M ops kernel sources
 set(_cortex_m_kernels__srcs
     ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_quantize_per_tensor.cpp

backends/mediatek/preprocess.py

@@ -20,6 +20,8 @@
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 
 SKIP_COMPILE_SPEC_KEYS = {"ImportForever"}
+REQUIRED_COMPILE_SPEC_KEYS = {"platform-config"}
+SUPPORTED_PLATFORM_CONFIGS = {"mt6989", "mt6991"}
 
 
 def assert_default_dim_order(edge_graph_module: torch.fx.GraphModule) -> None:
@@ -47,6 +49,28 @@ def preprocess(
         cls, edge_program: ExportedProgram, module_compile_spec: List[CompileSpec]
     ) -> PreprocessResult:
 
+        # Validate CompileSpec settings
+        compile_spec_keys = [spec.key for spec in module_compile_spec]
+        if len(compile_spec_keys) != len(set(compile_spec_keys)):
+            raise RuntimeError(
+                "Unsupported duplicated keys in the CompileSpec settings."
+            )
+        if not REQUIRED_COMPILE_SPEC_KEYS.issubset(set(compile_spec_keys)):
+            raise RuntimeError(
+                "Following keys are required in the CompileSpec settings: {}."
+                "".format(REQUIRED_COMPILE_SPEC_KEYS)
+            )
+        platform = [
+            spec.value.decode("utf-8")
+            for spec in module_compile_spec
+            if spec.key == "platform-config"
+        ][0]
+        if platform not in SUPPORTED_PLATFORM_CONFIGS:
+            raise ValueError(
+                "Unsupported value of platform-config CompileSpec. Given {} but expected to be one "
+                "of {}.".format(platform, SUPPORTED_PLATFORM_CONFIGS)
+            )
+
         # Make sure all inputs are contiguous_format or NCHW or default dim order
         assert_default_dim_order(edge_program.graph_module)
 
@@ -64,8 +88,7 @@ def preprocess(
             if name_to_node_mappings[name].meta["val"].dtype == torch.float32
         ]
 
-        # This default compile options are only for mt6989 SOC
-        compile_options = ["--arch=mdla5.1,edpa1.0", "--relax-fp32", "--opt=3"]
+        compile_options = ["--relax-fp32", "--opt=3"]
         for spec in module_compile_spec:
             if spec.key in SKIP_COMPILE_SPEC_KEYS:
                 continue

backends/vulkan/CMakeLists.txt

@@ -24,10 +24,6 @@ if(NOT RUNTIME_PATH)
   set(RUNTIME_PATH ${CMAKE_CURRENT_SOURCE_DIR}/runtime)
 endif()
 
-if(NOT PYTHON_EXECUTABLE)
-  set(PYTHON_EXECUTABLE python3)
-endif()
-
 # Include this file to access target_link_options_shared_lib This is required to
 # provide access to target_link_options_shared_lib which allows libraries to be
 # linked with the --whole-archive flag. This is required for libraries that

backends/vulkan/test/op_tests/CMakeLists.txt

@@ -45,10 +45,6 @@ find_library(LIB_TORCH torch HINTS ${TORCH_INSTALL_PREFIX}/lib)
 find_library(LIB_TORCH_CPU torch_cpu HINTS ${TORCH_INSTALL_PREFIX}/lib)
 find_library(LIB_C10 c10 HINTS ${TORCH_INSTALL_PREFIX}/lib)
 
-if(NOT PYTHON_EXECUTABLE)
-  set(PYTHON_EXECUTABLE python3)
-endif()
-
 # Third party include paths
 
 set(VULKAN_THIRD_PARTY_PATH ${CMAKE_CURRENT_SOURCE_DIR}/../../third-party)

backends/xnnpack/CMakeLists.txt

@@ -25,10 +25,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # NB: Enabling this will serialize execution of delegate instances Keeping this
 # OFF by default to maintain existing behavior, to be revisited.
 option(EXECUTORCH_XNNPACK_SHARED_WORKSPACE

configurations/CMakeLists.txt

@@ -18,10 +18,6 @@ endif()
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 set(_common_compile_options -Wno-deprecated-declarations)
 
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)

examples/mediatek/aot_utils/oss_utils/utils.py

@@ -14,6 +14,7 @@
     NeuropilotQuantizer,
     Precision,
 )
+from executorch.exir.backend.backend_details import CompileSpec
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
 
@@ -48,7 +49,7 @@ def build_executorch_binary(
     edge_compile_config = exir.EdgeCompileConfig(_check_ir_validity=False)
     # skipped op names are used for deeplabV3 model
     neuro_partitioner = NeuropilotPartitioner(
-        [],
+        [CompileSpec("platform-config", b"mt6989")],
         op_names_to_skip={
             "aten_convolution_default_106",
             "aten_convolution_default_107",

examples/mediatek/model_export_scripts/llama.py

@@ -350,10 +350,11 @@ def export_to_et_ir(
 
         print("Delegating Edge Program to Neuropilot Backend")
         compile_spec = [
-            CompileSpec("gno", struct.pack("3s", b"LTS")),
-            CompileSpec("gno-exp", struct.pack("0s", b"")),
-            CompileSpec("gno-non-4d-tiling", struct.pack("0s", b"")),
+            CompileSpec("gno", b"LTS"),
+            CompileSpec("gno-exp", b""),
+            CompileSpec("gno-non-4d-tiling", b""),
             CompileSpec("ImportForever", struct.pack("?", True)),
+            CompileSpec("platform-config", b"mt6989"),
         ]
         partitioner = NeuropilotPartitioner(compile_spec)
         delegated_program = edge_program.to_backend(partitioner)

examples/models/llama/README.md

@@ -295,6 +295,7 @@ Note for Mac users: There's a known linking issue with Xcode 15.1. Refer to the
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
         -DEXECUTORCH_BUILD_XNNPACK=ON \
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+        -DSUPPORT_REGEX_LOOKAHEAD=ON
         -Bcmake-out/examples/models/llama \
         examples/models/llama
 
@@ -353,6 +354,7 @@ cmake  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
     -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
     -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -DSUPPORT_REGEX_LOOKAHEAD=ON
     -Bcmake-out-android/examples/models/llama \
     examples/models/llama
 

examples/models/llama/export_llama_lib.py

@@ -719,6 +719,7 @@ def _prepare_for_llama_export(args) -> LLMEdgeManager:
             preq_mode=args.preq_mode,
             preq_group_size=args.preq_group_size,
             preq_embedding_quantize=args.preq_embedding_quantize,
+            local_global_attention=args.local_global_attention,
         )
     )
 
@@ -1447,7 +1448,7 @@ def _get_source_transforms(  # noqa
         transforms.append(
             partial(
                 replace_kv_cache_with_ring_kv_cache,
-                layer_sizes=args.local_global_attention,
+                layer_sizes=local_global_attention,
             )
         )
 

examples/models/qwen3/README.md

@@ -86,3 +86,6 @@ cmake-out/examples/models/llama/llama_main
 ```
 
 To run the model on an example iOS or Android app, see the Llama README's [Step 5: Build Mobile apps](../llama/README.md#step-5-build-mobile-apps) section.
+
+### FAQ
+For more help with exporting or running this model, feel free to ask in our [discord channel](https://lnkd.in/gWCM4ViK).

extension/flat_tensor/test/CMakeLists.txt

@@ -22,7 +22,7 @@ add_custom_command(
   OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/ModuleAddMulProgram.pte"
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleAddMulProgram.ptd"
   COMMAND
-    python -m test.models.export_program --modules "ModuleAddMul"
+    ${PYTHON_EXECUTABLE} -m test.models.export_program --modules "ModuleAddMul"
     --external-constants --outdir "${CMAKE_CURRENT_BINARY_DIR}" 2> /dev/null
   WORKING_DIRECTORY ${EXECUTORCH_ROOT}
 )

extension/llm/custom_ops/CMakeLists.txt

@@ -11,10 +11,6 @@ if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
 endif()
 
-if(NOT PYTHON_EXECUTABLE)
-  set(PYTHON_EXECUTABLE python3)
-endif()
-
 # Source root directory for executorch.
 if(NOT EXECUTORCH_ROOT)
   set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)

extension/module/test/CMakeLists.txt

@@ -24,10 +24,10 @@ add_custom_command(
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleAddMulProgram.pte"
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleAddMulProgram.ptd"
   COMMAND
-    python3 -m test.models.export_program --modules "ModuleAdd"
+    ${PYTHON_EXECUTABLE} -m test.models.export_program --modules "ModuleAdd"
     --outdir "${CMAKE_CURRENT_BINARY_DIR}" 2> /dev/null
   COMMAND
-    python3 -m test.models.export_program --modules "ModuleAddMul"
+    ${PYTHON_EXECUTABLE} -m test.models.export_program --modules "ModuleAddMul"
     --external-constants --outdir "${CMAKE_CURRENT_BINARY_DIR}" 2> /dev/null
   WORKING_DIRECTORY ${EXECUTORCH_ROOT}
 )

extension/runner_util/test/CMakeLists.txt

@@ -19,7 +19,7 @@ include(${EXECUTORCH_ROOT}/tools/cmake/Test.cmake)
 
 add_custom_command(
   OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/ModuleAdd.pte"
-  COMMAND python3 -m test.models.export_program --modules "ModuleAdd" --outdir
+  COMMAND ${PYTHON_EXECUTABLE} -m test.models.export_program --modules "ModuleAdd" --outdir
           "${CMAKE_CURRENT_BINARY_DIR}" 2> /dev/null
   WORKING_DIRECTORY ${EXECUTORCH_ROOT}
 )

kernels/optimized/CMakeLists.txt

@@ -36,9 +36,6 @@ list(APPEND _common_compile_options -DET_BUILD_WITH_BLAS)
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
 # Build cpublas.
 list(TRANSFORM _optimized_cpublas__srcs PREPEND "${EXECUTORCH_ROOT}/")
 add_library(cpublas STATIC ${_optimized_cpublas__srcs})

kernels/portable/CMakeLists.txt

@@ -26,10 +26,6 @@ set(_common_compile_options -Wno-deprecated-declarations)
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # Portable kernel sources TODO(larryliu0820): use buck2 to gather the sources
 file(GLOB_RECURSE _portable_kernels__srcs
      "${CMAKE_CURRENT_SOURCE_DIR}/cpu/*.cpp"

kernels/quantized/CMakeLists.txt

@@ -29,10 +29,6 @@ set(_common_compile_options -Wno-deprecated-declarations)
 include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # Quantized ops kernel sources TODO(larryliu0820): use buck2 to gather the
 # sources
 list(TRANSFORM _quantized_kernels__srcs PREPEND "${EXECUTORCH_ROOT}/")

kernels/test/CMakeLists.txt

@@ -42,11 +42,11 @@ foreach(kernel ${_kernels})
            "${_wrapper_dir}/supported_features.h"
     COMMAND mkdir -p ${_wrapper_dir}
     COMMAND
-      python kernels/test/gen_supported_features.py
+      ${PYTHON_EXECUTABLE} kernels/test/gen_supported_features.py
       kernels/${kernel}/test/supported_features_def.yaml >
       ${_wrapper_dir}/supported_features.cpp
     COMMAND
-      python kernels/test/gen_supported_features.py
+      ${PYTHON_EXECUTABLE} kernels/test/gen_supported_features.py
       kernels/test/supported_features.yaml >
       ${_wrapper_dir}/supported_features.h
     WORKING_DIRECTORY "${EXECUTORCH_ROOT}"

runtime/executor/method.cpp

@@ -754,7 +754,7 @@ Result<Method> Method::load(
     temp_allocator = platform_allocator;
   }
   Method method(program, memory_manager, event_tracer, temp_allocator);
-
+  ET_LOG(Debug, "Loading method: %s.", s_plan->name()->c_str());
   Error err = method.init(s_plan, named_data_map);
   if (err != Error::Ok) {
     return err;
@@ -1522,6 +1522,7 @@ Error Method::execute() {
       initialized(),
       NotSupported,
       "Cannot execute until method has been initialized.");
+  ET_LOG(Debug, "Executing method: %s.", method_meta().name());
 
   // Chains are executed sequentially today, but future async designs may
   // branch and run many in parallel or out of order.