Update on "Dtype selective build: enable in fbcode"

^
By removing `header_namespace` and depending on #include "selected_op_variants.h" instead of the full path #include <executorch/kernels/portable/cpu/selected_op_variants.h>

Note:
expose_operator_symbols=False only works in xplat, so add a failure message for that too. I don't think we should recommend for users to set it to true, as it prevents a library from linking multiple executorch_generated_libs (symbols will clash).

Differential Revision: [D75082395](https://our.internmc.facebook.com/intern/diff/D75082395/)

[ghstack-poisoned]

Author: lucylq

.ci/scripts/tests/test_gather_benchmark_configs.py

@@ -53,11 +53,11 @@ def test_skip_disabled_configs(self):
                 "mv3": [
                     self.gather_benchmark_configs.DisabledConfig(
                         config_name="disabled_config1",
-                        github_issue="https://github.com/org/repo/issues/123",
+                        github_issue="https://github.com/org/repo/issues/123",  # @lint-ignore
                     ),
                     self.gather_benchmark_configs.DisabledConfig(
                         config_name="disabled_config2",
-                        github_issue="https://github.com/org/repo/issues/124",
+                        github_issue="https://github.com/org/repo/issues/124",  # @lint-ignore
                     ),
                 ]
             },
@@ -84,7 +84,9 @@ def test_skip_disabled_configs(self):
             self.assertIn("enabled_config2", result)
 
     def test_disabled_configs_have_github_links(self):
-        github_issue_regex = re.compile(r"https://github\.com/.+/.+/issues/\d+")
+        github_issue_regex = re.compile(
+            r"https://github\.com/.+/.+/issues/\d+"  # @lint-ignore
+        )
 
         for (
             model_name,

.ci/scripts/unittest-linux.sh

@@ -20,8 +20,9 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
     source .ci/scripts/setup-vulkan-linux-deps.sh
 
+    # We need the runner to test the built library.
     PYTHON_EXECUTABLE=python \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON -DEXECUTORCH_BUILD_TESTS=ON" \
     .ci/scripts/setup-linux.sh "$@"
 
     .ci/scripts/unittest-linux-cmake.sh

.ci/scripts/unittest-macos.sh

@@ -20,8 +20,9 @@ export PATH="${TMP_DIR}:$PATH"
 trap 'rm -rfv ${TMP_DIR}' EXIT
 
 # Setup MacOS dependencies as there is no Docker support on MacOS atm
+# We need the runner to test the built library.
 PYTHON_EXECUTABLE=python \
-CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+CMAKE_ARGS="-DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON -DEXECUTORCH_BUILD_TESTS=ON" \
 ${CONDA_RUN} --no-capture-output \
 .ci/scripts/setup-macos.sh "$@"
 

.ci/scripts/wheel/envvar_base.sh

@@ -11,7 +11,3 @@
 # Ensure that CMAKE_ARGS is defined before referencing it. Defaults to empty
 # if not defined.
 export CMAKE_ARGS="${CMAKE_ARGS:-}"
-
-# Link the XNNPACK backend into the pybindings runtime so that users can execute
-# ExecuTorch programs that delegate to it.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON"

.ci/scripts/wheel/envvar_macos.sh

@@ -9,7 +9,3 @@
 # any variables so that subprocesses will see them.
 
 source "${GITHUB_WORKSPACE}/${REPOSITORY}/.ci/scripts/wheel/envvar_base.sh"
-
-# When building for macOS, link additional backends into the pybindings runtime.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_COREML=ON"
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_MPS=ON"

.github/scripts/label_utils.py

@@ -51,7 +51,6 @@ def get_last_page_num_from_header(header: Any) -> int:
     # rel="next", <https://api.github.com/repositories/65600975/labels?per_page=100&page=3>; rel="last"
     link_info = header["link"]
     # Docs does not specify that it should be present for projects with just few labels
-    # And https://github.com/malfet/deleteme/actions/runs/7334565243/job/19971396887 it's not the case
     if link_info is None:
         return 1
     prefix = "&page="

.github/scripts/test_extract_benchmark_results.py

@@ -29,8 +29,8 @@ def get_mock_happy_flow_content(app_type: str = "IOS_APP"):
                 "name": "Syslog",
                 "type": "DEVICE_LOG",
                 "extension": "syslog",
-                "url": "https://job_arn_1_device_log",
-                "s3_url": "https://job_arn_1/test-workflow1/1/syslog.syslog",
+                "url": "https://job_arn_1_device_log",  # @lint-ignore
+                "s3_url": "https://job_arn_1/test-workflow1/1/syslog.syslog",  # @lint-ignore
                 "app_type": app_type,
                 "job_name": "job_arn_1_name",
                 "os": "14",
@@ -55,8 +55,8 @@ def get_mock_happy_flow_content(app_type: str = "IOS_APP"):
                 "name": "Customer Artifacts",
                 "type": "CUSTOMER_ARTIFACT",
                 "extension": "zip",
-                "url": "https://job_arn_1_customer_artifact",
-                "s3_url": "https://job_arn_1_customer_artifact1",
+                "url": "https://job_arn_1_customer_artifact",  # @lint-ignore
+                "s3_url": "https://job_arn_1_customer_artifact1",  # @lint-ignore
                 "app_type": app_type,
                 "job_name": "job_arn_1_device_name",
                 "os": "14",
@@ -68,8 +68,8 @@ def get_mock_happy_flow_content(app_type: str = "IOS_APP"):
                 "name": "Syslog",
                 "type": "DEVICE_LOG",
                 "extension": "syslog",
-                "url": "https://job_arn_1_device_log",
-                "s3_url": "https://job_arn_1/test-workflow1/1/syslog.syslog",
+                "url": "https://job_arn_1_device_log",  # @lint-ignore
+                "s3_url": "https://job_arn_1/test-workflow1/1/syslog.syslog",  # @lint-ignore
                 "app_type": app_type,
                 "job_name": "job_arn_2_name",
                 "os": "14",
@@ -94,8 +94,8 @@ def get_mock_happy_flow_content(app_type: str = "IOS_APP"):
                 "name": "Customer Artifacts",
                 "type": "CUSTOMER_ARTIFACT",
                 "extension": "zip",
-                "url": "https://job_arn_1_customer_artifact",
-                "s3_url": "https://job_arn_1_customer_artifact1",
+                "url": "https://job_arn_1_customer_artifact",  # @lint-ignore
+                "s3_url": "https://job_arn_1_customer_artifact1",  # @lint-ignore
                 "app_type": app_type,
                 "job_name": "job_arn_2_name",
                 "os": "14",

.github/scripts/trymerge.py

@@ -1964,7 +1964,7 @@ def check_for_sev(org: str, project: str, skip_mandatory_checks: bool) -> None:
     response = cast(
         Dict[str, Any],
         gh_fetch_json_list(
-            "https://api.github.com/search/issues",
+            "https://api.github.com/search/issues",  # @lint-ignore
             params={"q": f'repo:{org}/{project} is:open is:issue label:"ci: sev"'},
         ),
     )

.gitignore

@@ -40,3 +40,19 @@ xcuserdata/
 .swiftpm/
 *.xcworkspace/
 *.xcframework/
+
+# misc
+/.vscode/
+*.so
+*.dylib
+/cmake_wrapper.sh
+/data/
+/devtools/bundled_program/serialize/
+/exir/_serialize/program.fbs
+/exir/_serialize/scalar_type.fbs
+/include/
+/share/
+/version.py
+
+# Android
+*.aar

.lintrunner.toml

@@ -378,3 +378,48 @@ command = [
     '--',
     '@{{PATHSFILE}}',
 ]
+
+[[linter]]
+code = "TORCH_AO_IMPORT"
+include_patterns = ["**/*.py"]
+exclude_patterns = [
+    "third-party/**",
+    # TODO: remove exceptions as we migrate
+    # backends
+    "backends/arm/quantizer/**",
+    "backends/arm/test/ops/**",
+    "backends/vulkan/quantizer/**",
+    "backends/vulkan/test/**",
+    "backends/cadence/aot/quantizer/**",
+    "backends/qualcomm/quantizer/**",
+    "examples/qualcomm/**",
+    "backends/xnnpack/quantizer/**",
+    "backends/xnnpack/test/**",
+    "exir/tests/test_passes.py",
+    "extension/llm/export/builder.py",
+    "extension/llm/export/quantizer_lib.py",
+    "exir/tests/test_memory_planning.py",
+    "backends/transforms/duplicate_dynamic_quant_chain.py",
+    "exir/backend/test/demos/test_xnnpack_qnnpack.py",
+]
+
+command = [
+  "python3",
+  "-m",
+  "lintrunner_adapters",
+  "run",
+  "grep_linter",
+  "--pattern=\\bfrom torch\\.ao\\.quantization\\.(?:quantizer|observer|quantize_pt2e|pt2e)(?:\\.[A-Za-z0-9_]+)*\\b",
+  "--linter-name=TorchAOImport",
+  "--error-name=Prohibited torch.ao.quantization import",
+  """--error-description=\
+  Imports from torch.ao.quantization are not allowed. \
+  Please import from torchao.quantization.pt2e instead.\n \
+  * torchao.quantization.pt2e (includes all the utils, including observers, fake quants etc.) \n \
+  * torchao.quantization.pt2e.quantizer (quantizer related objects and utils) \n \
+  * torchao.quantization.pt2e.quantize_pt2e (prepare_pt2e, prepare_qat_pt2e, convert_pt2e) \n\n \
+  If you need something from torch.ao.quantization, you can add your file to an exclude_patterns for TORCH_AO_IMPORT in .lintrunner.toml. \
+  """,
+  "--",
+  "@{{PATHSFILE}}",
+]

.mypy.ini

@@ -97,3 +97,6 @@ ignore_missing_imports = True
 
 [mypy-zstd]
 ignore_missing_imports = True
+
+[mypy-torchao.*]
+follow_untyped_imports = True

CMakeLists.txt

@@ -362,6 +362,9 @@ add_library(executorch_core ${_executorch_core__srcs})
 add_library(executorch_no_prim_ops ALIAS executorch_core)
 
 target_link_libraries(executorch_core PRIVATE program_schema)
+if(ANDROID)
+  target_link_libraries(executorch_core PUBLIC log)
+endif()
 if(EXECUTORCH_USE_DL)
   # Check if dl exists for this toolchain and only then link it.
   find_library(DL_LIBRARY_EXISTS NAMES dl)
@@ -653,10 +656,6 @@ if(EXECUTORCH_BUILD_PYBIND)
   install(TARGETS portable_lib
           LIBRARY DESTINATION executorch/extension/pybindings
   )
-
-  if(EXECUTORCH_BUILD_EXTENSION_TRAINING)
-
-  endif()
 endif()
 
 if(EXECUTORCH_BUILD_EXTENSION_TRAINING)

CMakePresets.json

@@ -1,17 +1,10 @@
 {
-  "version": 10,
-  "cmakeMinimumRequired": {
-    "major": 3,
-    "minor": 31,
-    "patch": 0
-  },
-  "$comment": "On-device AI across mobile, embedded and edge for PyTorch.",
+  "version": 6,
   "configurePresets": [
     {
       "name": "common",
       "hidden": true,
-      "binaryDir": "${sourceDir}/cmake-out",
-      "generator": "Unix Makefiles"
+      "binaryDir": "${sourceDir}/cmake-out"
     },
     {
       "name": "macos-arm64",

CONTRIBUTING.md

@@ -420,7 +420,7 @@ for basics.
 - Use [this](docs/source/backend-delegates-integration.md) guide when
   integrating your delegate with ExecuTorch.
 - Refer to [this](docs/source/backend-delegates-dependencies.md) set of
-  guidelines when including a third-party depenency for your delegate.
+  guidelines when including a third-party dependency for your delegate.
 
  
 

backends/apple/coreml/test/test_coreml_quantizer.py

@@ -15,12 +15,12 @@
 )
 
 from executorch.backends.apple.coreml.quantizer import CoreMLQuantizer
-from torch.ao.quantization.quantize_pt2e import (
+from torch.export import export_for_training
+from torchao.quantization.pt2e.quantize_pt2e import (
     convert_pt2e,
     prepare_pt2e,
     prepare_qat_pt2e,
 )
-from torch.export import export_for_training
 
 
 class TestCoreMLQuantizer:

backends/arm/_passes/__init__.py

@@ -15,6 +15,7 @@
 from .convert_any_default_dim_dims_pass import ConvertAnyDefaultDimDimsPass  # noqa
 from .convert_expand_copy_to_repeat import ConvertExpandCopyToRepeatPass  # noqa
 from .convert_full_like_to_full_pass import ConvertFullLikeToFullPass  # noqa
+from .convert_int_pow_to_mul import ConvertIntPowToMuls  # noqa
 from .convert_minmax_pass import ConvertMinMaxPass  # noqa
 from .convert_split_to_slice import ConvertSplitToSlicePass  # noqa
 from .convert_squeezes_to_view import ConvertSqueezesToViewPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -18,6 +18,7 @@
     ConvertAnyDefaultDimDimsPass,
     ConvertExpandCopyToRepeatPass,
     ConvertFullLikeToFullPass,
+    ConvertIntPowToMuls,
     ConvertMinMaxPass,
     ConvertMmToBmmPass,
     ConvertSplitToSlicePass,
@@ -131,14 +132,14 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
 
     def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(DecomposeSqrtPass())
+        self.add_pass(ConvertIntPowToMuls())
         self.add_pass(ReplaceScalarWithTensorArgPassTOSAMI())
         self.add_pass(FuseQuantizedActivationPass())
         self.add_pass(RemoveGetItemPass())
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(FuseBatchnorm2DPass(exported_program))
         self.add_pass(ConvertMmToBmmPass())
         self.add_pass(DecomposeLinearPass())
-        self.add_pass(DecomposeLinearVectorNormPass())
         self.add_pass(DecomposeLeakyReLUPass())
         self.add_pass(DecomposeBatchNormPass())
         self.add_pass(DecomposeLayerNormPass())

backends/arm/_passes/convert_int_pow_to_mul.py

@@ -0,0 +1,52 @@
+# 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.
+
+# pyre-unsafe
+
+from executorch.backends.arm._passes import ArmPass
+from executorch.exir.dialects._ops import ops as exir_ops
+
+
+class ConvertIntPowToMuls(ArmPass):
+    """
+    Replaces pow with integer exponent with a series of multiplications.
+    Only handles pow.Tensor_Scalar and not pow.Tensor_Tensor.
+    Needs to be run before doing scalar to tensor conversion.
+    """
+
+    def call_operator(self, op, args, kwargs, meta):
+        if op != exir_ops.edge.aten.pow.Tensor_Scalar:
+            return super().call_operator(op, args, kwargs, meta)
+
+        x = args[0]
+        exp = args[1]
+
+        # Handle zero first and return early
+        if exp == 0:
+            # return a tensor of ones with the same shape as x
+            return super().call_operator(
+                exir_ops.edge.aten.full_like.default, (x, 1), {}, meta, True
+            )
+
+        if not isinstance(exp, int):
+            return super().call_operator(op, args, kwargs, meta)
+
+        # Handle negative exponent
+        if exp < 0:
+            x = super().call_operator(
+                exir_ops.edge.aten.reciprocal.default, (x,), {}, meta, True
+            )
+            exp = -exp
+
+        res = x
+
+        # Consider exponentiation by squaring, if exp turns out to be large.
+        # Now we just roll out the multiplications.
+        for _ in range(exp - 1):
+            res = super().call_operator(
+                exir_ops.edge.aten.mul.Tensor, (res, x), {}, meta, True
+            )
+
+        return res