Update on "stop double-installing ExecuTorch in one-off linux jobs"

setup-linux.sh already installs ExecuTorch with XNNPACK (and it passes use-pt-pinned-commit as it should).

Differential Revision: [D67996460](https://our.internmc.facebook.com/intern/diff/D67996460/)

[ghstack-poisoned]

.github/pytorch-probot.yml

@@ -1,4 +1,5 @@
 # The schema is from https://github.com/pytorch/pytorch/blob/main/.github/pytorch-probot.yml
+tracking_issue: 7679
 ciflow_push_tags:
 - ciflow/android
 - ciflow/apple

.github/workflows/android-perf.yml

@@ -260,7 +260,7 @@ jobs:
                       --output_name="${OUT_ET_MODEL_NAME}.pte"
                     ls -lh "${OUT_ET_MODEL_NAME}.pte"
                 elif [[ ${{ matrix.config }} == "llama3_qnn_htp" ]]; then
-                    export QNN_SDK_ROOT=/tmp/qnn/2.25.0.240728
+                    export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
                     export LD_LIBRARY_PATH=$QNN_SDK_ROOT/lib/x86_64-linux-clang/
                     export PYTHONPATH=$(pwd)/..
 
@@ -347,7 +347,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
 
         export ANDROID_ABIS="arm64-v8a"
-        PYTHON_EXECUTABLE=python EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.25.0.240728 bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+        PYTHON_EXECUTABLE=python EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029 bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
 
   # Let's see how expensive this job is, we might want to tone it down by running it periodically
   benchmark-on-device:

.github/workflows/pull.yml

@@ -330,6 +330,9 @@ jobs:
 
   unittest-arm:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-arm-sdk
@@ -392,6 +395,25 @@ jobs:
         # Test llama2
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh -model stories110M -build_tool "${BUILD_TOOL}" -mode "${MODE}" -dtype "${DTYPE}" -pt2e_quantize "${PT2E_QUANTIZE}"
 
+  test-qnn-models-linux:
+    name: test-qnn-models-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-qnn-sdk
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 180
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        # placeholder for running test_qnn_delegate.py, can use matrix such that we can trigger different jobs, refers to test-llama-runner-qnn-linux
+        # reminder: make sure each job runs fast
+
   test-phi-3-mini-runner-linux:
     name: test-phi-3-mini-runner-linux
     uses: pytorch/test-infra/.github/workflows/linux_job.yml@main

.github/workflows/trunk.yml

@@ -132,6 +132,9 @@ jobs:
   test-arm-backend-delegation:
     name: test-arm-backend-delegation
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-arm-sdk
@@ -159,6 +162,9 @@ jobs:
   test-arm-reference-delegation:
     name: test-arm-reference-delegation
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
     with:
       runner: linux.2xlarge
       docker-image: executorch-ubuntu-22.04-arm-sdk

.lintrunner.toml

@@ -78,8 +78,6 @@ exclude_patterns = [
     # File contains @generated
     'extension/llm/custom_ops/spinquant/fast_hadamard_transform_special.h',
     'extension/llm/custom_ops/spinquant/test/fast_hadamard_transform_special_unstrided_cpu.h',
-    # Want to be able to keep c10 in sync with PyTorch core.
-    'runtime/core/portable_type/c10/**',
 ]
 command = [
     'python',
@@ -263,8 +261,6 @@ exclude_patterns = [
     'extension/**',
     'kernels/optimized/**',
     'runtime/core/exec_aten/**',
-    # Want to be able to keep c10 in sync with PyTorch core.
-    'runtime/core/portable_type/c10/**',
     'runtime/executor/tensor_parser_aten.cpp',
     'scripts/**',
     'test/**',
@@ -298,6 +294,7 @@ include_patterns = [
     'build/**/*.py',
     'codegen/**/*.py',
     # 'devtools/**/*.py',
+    'devtools/visualization/**/*.py',
     'docs/**/*.py',
     # 'examples/**/*.py',
     # 'exir/**/*.py',

CMakeLists.txt

@@ -361,7 +361,7 @@ if(NOT "${_repo_dir_name}" STREQUAL "executorch")
       "fix for this restriction."
   )
 endif()
-set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/.. ${CMAKE_CURRENT_SOURCE_DIR}/runtime/core/portable_type)
+set(_common_include_directories ${CMAKE_CURRENT_SOURCE_DIR}/..)
 
 #
 # The `_<target>_srcs` lists are defined by including ${EXECUTORCH_SRCS_FILE}.
@@ -544,7 +544,6 @@ endif()
 target_include_directories(
   executorch_core PUBLIC ${_common_include_directories}
 )
-target_compile_definitions(executorch_core PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch_core PUBLIC ${_common_compile_options})
 if(MAX_KERNEL_NUM)
   target_compile_definitions(
@@ -565,7 +564,6 @@ if(EXECUTORCH_BUILD_PYBIND AND APPLE)
   target_include_directories(
     executorch_core_shared PUBLIC ${_common_include_directories}
   )
-  target_compile_definitions(executorch_core_shared PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
   target_compile_options(
     executorch_core_shared PUBLIC ${_common_compile_options}
   )
@@ -586,7 +584,6 @@ endif()
 add_library(executorch ${_executorch__srcs})
 target_link_libraries(executorch PRIVATE executorch_core)
 target_include_directories(executorch PUBLIC ${_common_include_directories})
-target_compile_definitions(executorch PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
 target_compile_options(executorch PUBLIC ${_common_compile_options})
 target_link_options_shared_lib(executorch)
 
@@ -620,12 +617,6 @@ endif()
 
 # Install `executorch` library as well as `executorch-config.cmake` under
 # ${CMAKE_INSTALL_PREFIX}/
-install(DIRECTORY runtime/core/  DESTINATION include/executorch/runtime/core FILES_MATCHING PATTERN "*.h")
-install(DIRECTORY runtime/kernel/  DESTINATION include/executorch/runtime/kernel FILES_MATCHING PATTERN "*.h")
-install(DIRECTORY runtime/platform/  DESTINATION include/executorch/runtime/platform FILES_MATCHING PATTERN "*.h")
-install(DIRECTORY extension/kernel_util/  DESTINATION include/executorch/extension/kernel_util FILES_MATCHING PATTERN "*.h")
-install(DIRECTORY extension/tensor/  DESTINATION include/executorch/extension/tensor FILES_MATCHING PATTERN "*.h")
-install(DIRECTORY extension/threadpool/  DESTINATION include/executorch/extension/threadpool FILES_MATCHING PATTERN "*.h")
 install(
   TARGETS executorch executorch_core
   DESTINATION lib
@@ -784,8 +775,6 @@ if(EXECUTORCH_BUILD_PYBIND)
   target_include_directories(
     util PUBLIC ${_common_include_directories} ${TORCH_INCLUDE_DIRS}
   )
-  target_compile_definitions(util PUBLIC C10_USING_CUSTOM_GENERATED_MACROS)
-
   target_compile_options(util PUBLIC ${_pybind_compile_options})
   target_link_libraries(util PRIVATE torch c10 executorch extension_tensor)
 

backends/apple/coreml/CMakeLists.txt

@@ -134,8 +134,6 @@ target_include_directories(
   coremldelegate PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/runtime/util
 )
 target_include_directories(coremldelegate PRIVATE ${EXECUTORCH_ROOT}/..)
-target_include_directories(coremldelegate PRIVATE ${EXECUTORCH_ROOT}/runtime/core/portable_type)
-target_compile_definitions(coremldelegate PRIVATE C10_USING_CUSTOM_GENERATED_MACROS)
 target_link_libraries(coremldelegate PRIVATE executorch_core)
 
 if(EXECUTORCH_BUILD_DEVTOOLS)

backends/apple/coreml/README.md

@@ -93,14 +93,14 @@ class Model(torch.nn.Module):
 source_model = Model()
 example_inputs = (torch.randn((1, 3, 256, 256)), )
 
-pre_autograd_aten_dialect = export_for_training(model, example_inputs).module()
+pre_autograd_aten_dialect = export_for_training(source_model, example_inputs).module()
 
 quantization_config = LinearQuantizerConfig.from_dict(
     {
         "global_config": {
             "quantization_scheme": QuantizationScheme.symmetric,
-            "activation_dtype": torch.uint8,
-            "weight_dtype": torch.int8,
+            "activation_dtype": torch.quint8,
+            "weight_dtype": torch.qint8,
             "weight_per_channel": True,
         }
     }

backends/apple/coreml/runtime/workspace/executorchcoreml.xcodeproj/project.pbxproj

@@ -830,7 +830,6 @@
 				GCC_OPTIMIZATION_LEVEL = 0;
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
-                                        "C10_USING_CUSTOM_GENERATED_MACROS",
 					"$(inherited)",
 				);
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
@@ -912,7 +911,6 @@
 				DEVELOPMENT_TEAM = "";
 				GCC_PREPROCESSOR_DEFINITIONS = (
 					"DEBUG=1",
-                                        "C10_USING_CUSTOM_GENERATED_MACROS",
 					"ET_EVENT_TRACER_ENABLED=1",
 					"$(inherited)",
 				);
@@ -922,7 +920,6 @@
 					"$(SRCROOT)/../kvstore",
 					"$(SRCROOT)/../inmemoryfs",
 					"$(SRCROOT)/../include",
-					"$(SRCROOT)/../include/executorch/runtime/core/portable_type",
 					"$(SRCROOT)/../sdk",
 					"$(SRCROOT)/../util",
 					"$(SRCROOT)/../../third-party/nlohmann_json/single_include",
@@ -954,7 +951,6 @@
 					"$(SRCROOT)/../kvstore",
 					"$(SRCROOT)/../inmemoryfs",
 					"$(SRCROOT)/../include",
-					"$(SRCROOT)/../include/executorch/runtime/core/portable_type",
 					"$(SRCROOT)/../sdk",
 					"$(SRCROOT)/../util",
 					"$(SRCROOT)/../../third-party/nlohmann_json/single_include",

backends/arm/CMakeLists.txt

@@ -14,8 +14,7 @@ endif()
 
 include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 
-set(_common_include_directories ${EXECUTORCH_ROOT}/.. ${EXECUTORCH_ROOT}/runtime/core/portable_type)
-add_compile_definitions(C10_USING_CUSTOM_GENERATED_MACROS)
+set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 
 # Third-party folder and Ethos-U driver inclued
 set(THIRD_PARTY_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/third-party")

backends/arm/_passes/arm_pass_manager.py

@@ -7,7 +7,6 @@
 
 # pyre-unsafe
 
-import torch
 from executorch.backends.arm._passes.annotate_channels_last_dim_order_pass import (
     AnnotateChannelsLastDimOrder,
 )
@@ -47,7 +46,7 @@
 )
 from executorch.backends.arm._passes.match_arg_ranks_pass import MatchArgRanksPass
 from executorch.backends.arm._passes.meandim_to_averagepool_pass import (
-    ConvertMeanDimToAveragePool,
+    ConvertMeanDimToAveragePoolPass,
 )
 from executorch.backends.arm._passes.mm_to_bmm_pass import ConvertMmToBmmPass
 from executorch.backends.arm._passes.remove_clone_pass import RemoveClonePass
@@ -61,86 +60,98 @@
 from executorch.backends.arm._passes.unsqueeze_scalar_placeholders_pass import (
     UnsqueezeScalarPlaceholdersPass,
 )
+from executorch.backends.arm.tosa_specification import TosaSpecification
 from executorch.backends.xnnpack._passes.remove_getitem_op import RemoveGetItemPass
 from executorch.exir import ExportedProgram
-from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_manager import PassManager
+from torch.fx import GraphModule
 
 
 class ArmPassManager(PassManager):
 
-    def _transform(self, graph_module: torch.fx.GraphModule):
+    def __init__(self, tosa_spec: TosaSpecification) -> None:
+        self.tosa_spec = tosa_spec
+        super().__init__()
+
+    def _transform(self, graph_module: GraphModule):
         return self(graph_module).graph_module
 
-    def transform_to_backend_pipeline(self, exported_program: ExportedProgram):
-        """Apply passes before transforming program to backend"""
+    def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(FuseQuantizedActivationPass())
+        self.add_pass(RemoveGetItemPass())
+        self.add_pass(ConvertSplitToSlicePass())
+        self.add_pass(ConvertMmToBmmPass())
         self.add_pass(DecomposeLinearPass())
+        self.add_pass(ConvertMeanDimToAveragePoolPass())
+
+        self.add_pass(AnnotateDecomposedMatmulPass())
+        self.add_pass(QuantizeFullArgument())
+        self.add_pass(FoldAndAnnotateQParamsPass())
+        self.add_pass(RetraceFoldedDtypesPass())
+        self.add_pass(InsertTableOpsPass(exported_program))
+
+        self.add_pass(RemoveClonePass())
+        self.add_pass(SizeAdjustConv2DPass())
+        self.add_pass(ConvertExpandCopyToRepeatPass())
+        self.add_pass(UnsqueezeBeforeRepeatPass())
+        self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
+        self.add_pass(CastInt64ToInt32Pass(exported_program))
+        self.add_pass(MatchArgRanksPass(exported_program))
+        self.add_pass(KeepDimsFalseToSqueezePass())
+        self.add_pass(Conv1dUnsqueezePass(exported_program))
+        self.add_pass(DecomposeSelectPass())
+
+        self.add_pass(AnnotateChannelsLastDimOrder())
+
+        return self._transform(exported_program.graph_module)
+
+    def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
+
+        self.add_pass(FuseQuantizedActivationPass())
         self.add_pass(RemoveGetItemPass())
+        self.add_pass(ConvertSplitToSlicePass())
+        self.add_pass(ConvertMmToBmmPass())
+        self.add_pass(DecomposeLinearPass())
         self.add_pass(DecomposeLayerNormPass())
         self.add_pass(DecomposeVarPass())
-        self.add_pass(ConvertMeanDimToAveragePool())
         self.add_pass(DecomposeMeanDimPass())
-        self.add_pass(ConvertSplitToSlicePass())
-        self.add_pass(ConvertMmToBmmPass())
-        # TODO MLETORCH-558
+        self.add_pass(ConvertMeanDimToAveragePoolPass())
+        self.add_pass(DecomposeDivPass())
+        self.add_pass(DecomposeSoftmaxesPass())
+
         self.add_pass(AnnotateDecomposedMatmulPass())
         self.add_pass(QuantizeFullArgument())
-        self.add_pass(
-            FoldAndAnnotateQParamsPass(
-                [
-                    exir_ops.edge.aten.minimum.default,
-                    exir_ops.edge.aten.maximum.default,
-                    exir_ops.edge.aten.add.Tensor,
-                    exir_ops.edge.aten.avg_pool2d.default,
-                    exir_ops.edge.aten.bmm.default,
-                    exir_ops.edge.aten.cat.default,
-                    exir_ops.edge.aten.convolution.default,
-                    exir_ops.edge.aten.clone.default,
-                    exir_ops.edge.aten.exp.default,
-                    exir_ops.edge.aten.expand_copy.default,
-                    exir_ops.edge.aten.full.default,
-                    exir_ops.edge.aten.hardtanh.default,
-                    exir_ops.edge.aten.log.default,
-                    exir_ops.edge.aten.max_pool2d.default,
-                    exir_ops.edge.aten.mul.Tensor,
-                    exir_ops.edge.aten.permute_copy.default,
-                    exir_ops.edge.aten.reciprocal.default,
-                    exir_ops.edge.aten.relu.default,
-                    exir_ops.edge.aten.repeat.default,
-                    exir_ops.edge.aten.rsqrt.default,
-                    exir_ops.edge.aten.select_copy.int,
-                    exir_ops.edge.aten.sigmoid.default,
-                    exir_ops.edge.aten.slice_copy.Tensor,
-                    exir_ops.edge.aten.squeeze_copy.dims,
-                    exir_ops.edge.aten.sub.Tensor,
-                    exir_ops.edge.aten.sum.dim_IntList,
-                    exir_ops.edge.aten.tanh.default,
-                    exir_ops.edge.aten.unsqueeze_copy.default,
-                    exir_ops.edge.aten.upsample_nearest2d.vec,
-                    exir_ops.edge.aten.view_copy.default,
-                ]
-            )
-        )
+        self.add_pass(FoldAndAnnotateQParamsPass())
         self.add_pass(RetraceFoldedDtypesPass())
         self.add_pass(InsertTableOpsPass(exported_program))
+
+        self.add_pass(RemoveClonePass())
+        self.add_pass(SizeAdjustConv2DPass())
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(UnsqueezeBeforeRepeatPass())
-        self.add_pass(CastInt64ToInt32Pass(exported_program))
         self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
-        self.add_pass(SizeAdjustConv2DPass())
-        self.add_pass(RemoveClonePass())
+        self.add_pass(CastInt64ToInt32Pass(exported_program))
         self.add_pass(MatchArgRanksPass(exported_program))
-        self.add_pass(DecomposeDivPass())
         self.add_pass(KeepDimsFalseToSqueezePass())
         self.add_pass(Conv1dUnsqueezePass(exported_program))
-        self.add_pass(DecomposeSoftmaxesPass())
         self.add_pass(DecomposeSelectPass())
+
         self.add_pass(AnnotateChannelsLastDimOrder())
 
         return self._transform(exported_program.graph_module)
 
-    def transform_for_annotation_pipeline(self, graph_module: torch.fx.GraphModule):
+    def transform_to_backend_pipeline(self, exported_program: ExportedProgram):
+        """Apply passes before transforming program to backend"""
+        if self.tosa_spec == TosaSpecification.create_from_string("TOSA-0.80.0+BI"):
+            return self._tosa_080_BI_pipeline(exported_program)
+        elif self.tosa_spec == TosaSpecification.create_from_string("TOSA-0.80.0+MI"):
+            return self._tosa_080_MI_pipeline(exported_program)
+        else:
+            raise NotImplementedError(
+                f"No pass pipeline implemented for {self.tosa_spec=}"
+            )
+
+    def transform_for_annotation_pipeline(self, graph_module: GraphModule):
         self.add_pass(ScalarsToAttributePass())
         self.add_pass(DecomposeLayerNormPass())
         self.add_pass(DecomposeVarPass())