Merge branch 'main' into export-D85704977

Author: Gasoonjia

.ci/docker/ci_commit_pins/optimum-executorch.txt

@@ -1 +1 @@
-4361747abfc55e40e929396ed986efe775d745f9
+d03e90c2cd9048e6d9a75285c0355f033cd016fc

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-556fc09a9f67f24ca5591ec049c5d0c347c5f62a
+b31bad1b8f1331bf43d47f46602cf6141db56844

.ci/docker/requirements-ci.txt

@@ -30,7 +30,6 @@ sphinx-reredirects==0.1.4
 matplotlib>=3.9.4
 sphinx-copybutton==0.5.2
 # PyTorch Theme
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@pytorch_sphinx_theme2#egg=pytorch_sphinx_theme2
-
+pytorch_sphinx_theme2==0.2.0
 # script unit test requirements
 yaspin==3.1.0

.ci/scripts/test_qnn_static_llm.sh

@@ -81,7 +81,7 @@ elif [[ "${TASK_NAME}" == "stories_260k_bc" ]]; then
     fi
 
 elif [[ "${TASK_NAME}" == "smollm2_135m" ]]; then
-    $PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_static_smollm2 --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir ./static_smollm2 --enable_x86_64
+    $PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_static_llm_model --model_name smollm2_135m --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir ./static_smollm2 --enable_x86_64
     exit_code1=$?
     if [ $exit_code1 -ne 0 ]; then
         exit 1

.github/workflows/doc-build.yml

@@ -46,14 +46,15 @@ jobs:
           export CHANNEL=nightly
         fi
 
-        # Get the version of ExecuTorch from REF_NAME and save as ET_VERSION_DOCS
-        # ET_VERSION_DOCS will be pulled during the doc build to add to the version dropdown
-        # on the website. See docs/source/conf.py for details
-
+        # Set RELEASE environment variable for tagged releases
         GITHUB_REF=${{ github.ref }}
-        echo "$GITHUB_REF"
-        export ET_VERSION_DOCS="${GITHUB_REF}"
-        echo "$ET_VERSION_DOCS"
+        if [[ "${GITHUB_REF}" =~ ^refs/tags/v[0-9]+\.[0-9]+ ]]; then
+          export RELEASE=true
+          echo "Building release docs (RELEASE=true)"
+        else
+          export RELEASE=false
+          echo "Building main docs (RELEASE=false)"
+        fi
 
         set -eux
 

.lintrunner.toml

@@ -449,3 +449,24 @@ command = [
   "--",
   "@{{PATHSFILE}}",
 ]
+
+[[linter]]
+code = 'ETVKNODEBUG'
+include_patterns = [
+    "backends/vulkan/**/*.glsl",
+]
+command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'grep_linter',
+    '--pattern=((DEBUG_MODE)|(GL_EXT_debug_printf))',
+    '--linter-name=ETVKNODEBUG',
+    '--error-name=Using DEBUG_MODE or GL_EXT_debug_printf in Vulkan shader',
+    """--error-description=\
+        #define DEBUG_MODE or #extension GL_EXT_debug_printf should only be used during development!
+    """,
+    '--',
+    '@{{PATHSFILE}}',
+]

.mypy.ini

@@ -74,6 +74,12 @@ ignore_missing_imports = True
 [mypy-pytorch_sphinx_theme]
 ignore_missing_imports = True
 
+[mypy-pytorch_sphinx_theme2]
+ignore_missing_imports = True
+
+[mypy-executorch.version]
+ignore_missing_imports = True
+
 [mypy-ruamel]
 ignore_missing_imports = True
 

backends/arm/CMakeLists.txt

@@ -48,17 +48,44 @@ endif()
 
 # VGF backend builds
 if(EXECUTORCH_BUILD_VGF)
-
-  # include libvgf
-  set(LIBVGF_PATH
-      "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/sw/vgf-lib/"
-  )
-
   set(VULKAN_THIRD_PARTY_PATH ${EXECUTORCH_ROOT}/backends/vulkan/third-party)
   set(VULKAN_HEADERS_PATH ${VULKAN_THIRD_PARTY_PATH}/Vulkan-Headers/include)
   set(VOLK_HEADERS_PATH ${VULKAN_THIRD_PARTY_PATH}/volk)
 
-  set(LIBVGF_STATIC "${LIBVGF_PATH}/build/src/libvgf.a")
+  if(APPLE
+     OR CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)$"
+     OR EXISTS
+        "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/"
+  )
+    message(STATUS "libvgf sourced from local scratch tree")
+
+    # Legacy layout: libvgf sourced from local scratch tree
+    set(LIBVGF_PATH
+        "${EXECUTORCH_ROOT}/examples/arm/ethos-u-scratch/ml-sdk-for-vulkan-manifest/sw/vgf-lib/"
+    )
+    set(LIBVGF_STATIC "${LIBVGF_PATH}/build/src/libvgf.a")
+  else()
+    message(STATUS "libvgf installed from pip package")
+
+    set(Python3_FIND_VIRTUALENV FIRST)
+    if(EXECUTORCH_ROOT AND EXISTS "${EXECUTORCH_ROOT}/env")
+      set(Python3_EXECUTABLE "${EXECUTORCH_ROOT}/env/bin/python3")
+    endif()
+
+    find_package(Python3 REQUIRED COMPONENTS Interpreter)
+
+    # Prefer arch-specific site-packages if present, else pure
+    set(_vgf_site_arch "${Python3_SITEARCH}/vgf_lib/binaries")
+    set(_vgf_site_pure "${Python3_SITELIB}/vgf_lib/binaries")
+    if(EXISTS "${_vgf_site_arch}")
+      set(LIBVGF_PATH "${_vgf_site_arch}")
+    else()
+      set(LIBVGF_PATH "${_vgf_site_pure}")
+    endif()
+
+    set(LIBVGF_STATIC "${LIBVGF_PATH}/lib/libvgf.a")
+  endif()
+
   set(LIBVGF_INCLUDE "${LIBVGF_PATH}/include/")
 
   add_library(vgf STATIC IMPORTED)

backends/arm/_passes/__init__.py

@@ -21,6 +21,9 @@
 from .convert_int64_output_ops_to_int32 import ConvertInt64OutputOpsToInt32Pass  # noqa
 from .convert_int_pow_to_mul import ConvertIntPowToMuls  # noqa
 from .convert_minmax_pass import ConvertMinMaxPass  # noqa
+from .convert_permute_singleton_to_view_pass import (  # noqa
+    ConvertPermuteSingletonToViewPass,
+)
 from .convert_split_to_slice import ConvertSplitToSlicePass  # noqa
 from .convert_squeezes_to_view import ConvertSqueezesToViewPass  # noqa
 from .convert_to_clamp import ConvertToClampPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -27,6 +27,7 @@
     ConvertIntPowToMuls,
     ConvertMinMaxPass,
     ConvertMmToBmmPass,
+    ConvertPermuteSingletonToViewPass,
     ConvertSplitToSlicePass,
     ConvertSqueezesToViewPass,
     ConvertToClampPass,
@@ -158,7 +159,12 @@ def _transform(self, graph_module: GraphModule):
     def _tosa_pipeline(
         self, exported_program: ExportedProgram, graph_module: GraphModule
     ) -> GraphModule:
+        # Preprocessing passes
+
         self.add_pass(AnnotateOutputDimOrderPass())
+
+        # Node transformation passes (pre q/dq folding)
+
         self.add_pass(FuseQuantizedActivationPass())
         self.add_pass(RemoveGetItemPass())
         self.add_pass(ConvertToClampPass())
@@ -173,8 +179,19 @@ def _tosa_pipeline(
         self.add_pass(ConvertELUParamsPass())
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(QuantizeOperatorArguments())
+
+        # Fold Q/DQ nodes, insert INT8/INT32 rescales.
+
         self.add_pass(FoldAndAnnotateQParamsPass(exported_program))  # type: ignore[call-arg]
         self.add_pass(FuseDuplicateUsersPass())
+        # TODO: DecomposeLinearPass should run after InsertRescaleInt32Pass or
+        # before FoldAndAnnotateQParamsPass but is unable to at the moment.
+        # Ticket: MLETORCH-1539
+        self.add_pass(DecomposeLinearPass())
+        self.add_pass(InsertRescaleInt32Pass())
+
+        # Node transformation passes (post q/dq folding)
+
         self.add_pass(DecomposeExpm1Pass())
         self.add_pass(DecomposeLogitPass())
         self.add_pass(DecomposeMaskedFill())
@@ -195,56 +212,67 @@ def _tosa_pipeline(
         self.add_pass(DecomposeSignPass())
         self.add_pass(DecomposeFloorDividePass())
         self.add_pass(DecomposeDivTensorModePass())
+        self.add_pass(DecomposeGeluPass())
+        self.add_pass(DecomposeAddSubAlphaPass())
+        self.add_pass(DecomposeGroupedConv())
+        self.add_pass(Conv1dUnsqueezePass())
+
+        # Scalars -> tensors, match tensor dtypes and ranks.
+
         self.add_pass(ReplaceScalarWithTensorByProfilePass())
+        self.add_pass(ConvertFullLikeToFullPass())
+        self.add_pass(MatchArgDtypePass())
+        self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
+        # TODO: Move DecomposeNotEqualPass to before or after this block of
+        # passes. Ticket: MLETORCH-1540
+        self.add_pass(DecomposeNotEqualPass())
+        self.add_pass(MatchArgRanksPass(exported_program))
+        self.add_pass(FuseConstantArgsPass(exported_program))
+
+        # Node transformation passes (post scalar-removal)
+
         self.add_pass(DecomposeRemainderPass())
         self.add_pass(DecomposeDivTensorModePass())
         self.add_pass(DecomposeEmbeddingPass())
         self.add_pass(FuseBatchnorm2DPass(exported_program))
         self.add_pass(ConvertMmToBmmPass())
         self.add_pass(DecomposeGluPass())
-        self.add_pass(DecomposeLinearPass())
         self.add_pass(DecomposeLeakyReLUPass())
-        self.add_pass(DecomposeNotEqualPass())
         self.add_pass(DecomposeDivPass())
-        self.add_pass(DecomposeAddSubAlphaPass())
         self.add_pass(DecomposeSoftmaxPass())
-        self.add_pass(DecomposeGeluPass())
-        self.add_pass(ConvertFullLikeToFullPass())
         self.add_pass(ConvertMinMaxPass())
         self.add_pass(ConvertAnyDefaultDimDimsPass())
-        self.add_pass(MatchArgDtypePass())
-        self.add_pass(UnsqueezeScalarPlaceholdersPass(exported_program))
-        self.add_pass(MatchArgRanksPass(exported_program))
         self.add_pass(DecomposeAdaptiveAvgPool2dPass())
         self.add_pass(DecomposeAvgPool2d())
         self.add_pass(
             DecorateFp32toInt32CastingPass()
         )  # Require that no new fp32->int32 is introduced after this pass
         self.add_pass(ComputeConstantOpsAOT(exported_program))
-
-        self.add_pass(DecomposeGroupedConv())
         self.add_pass(ConvertExpandCopyToRepeatPass())
         self.add_pass(UnsqueezeBeforeRepeatPass())
         self.add_pass(DecomposeCumsumPass(exported_program))
-        self.add_pass(Conv1dUnsqueezePass())
         self.add_pass(DecomposeMaxPool2DPass())
         self.add_pass(SizeAdjustInputPass())
         self.add_pass(DecomposeSelectPass())
         self.add_pass(ConvertSqueezesToViewPass())
         self.add_pass(CastToInt32Pass())
         self.add_pass(BroadcastArgsPass())
-
+        self.add_pass(ConvertPermuteSingletonToViewPass())
         self.add_pass(FuseViewCopyTransform())
-        self.add_pass(FuseConstantArgsPass(exported_program))
         self.add_pass(DecomposeConv2dWithInt16ActivationPass())
-        self.add_pass(CastInt64BuffersToInt32Pass(exported_program))
+        self.add_pass(DecomposeSumPass())
         self.add_pass(InsertTableOpsPass(exported_program))
+
+        # Aten -> TOSA transformation passes
+
         self.add_pass(RewriteUpsamplePass())
         self.add_pass(RewriteConv2dPass(exported_program))
         self.add_pass(RewriteMatmulPass())
+
+        # Postprocessing/cleanup passes
+
+        self.add_pass(CastInt64BuffersToInt32Pass(exported_program))
         self.add_pass(FuseEqualPlaceholdersPass(exported_program))
-        self.add_pass(InsertRescaleInt32Pass())
-        self.add_pass(DecomposeSumPass())
         self.add_pass(ToTosaMemoryFormatPass(exported_program))
         self.add_pass(RemoveNoopPass())
         self.add_pass(InsertRescalePass())