Merge branch 'main' into bump-tokenizers-pin-again

Author: GregoryComer

.ci/scripts/test_wheel_package_qnn.sh

@@ -98,7 +98,7 @@ PYTHON_VERSION=$1
 # Check wheel does NOT contain qualcomm/sdk
 # ----------------------------
 echo "Checking wheel does not contain qualcomm/sdk..."
-SDK_FILES=$(unzip -l "$WHEEL_FILE" | awk '{print $4}' | grep "executorch/backends/qualcomm/sdk" || true)
+SDK_FILES=$(unzip -l "$WHEEL_FILE" | awk '{print $4}' | grep -E "executorch/backends/qualcomm/sdk" || true)
 if [ -n "$SDK_FILES" ]; then
     echo "ERROR: Wheel package contains unexpected qualcomm/sdk files:"
     echo "$SDK_FILES"
@@ -111,7 +111,7 @@ fi
 # Check .so files in the wheel
 # ----------------------------
 echo "Checking for .so files inside the wheel..."
-WHEEL_SO_FILES=$(unzip -l "$WHEEL_FILE" | awk '{print $4}' | grep "executorch/backends/qualcomm/python" || true)
+WHEEL_SO_FILES=$(unzip -l "$WHEEL_FILE" | awk '{print $4}' | grep -E "executorch/backends/qualcomm/python" || true)
 if [ -z "$WHEEL_SO_FILES" ]; then
     echo "ERROR: No .so files found in wheel under executorch/backends/qualcomm/python"
     exit 1
@@ -139,8 +139,30 @@ run_core_tests () {
   echo "=== [$LABEL] Installing wheel & deps ==="
   "$PIPBIN" install --upgrade pip
   "$PIPBIN" install "$WHEEL_FILE"
-  "$PIPBIN" install torch=="2.9.0.dev20250906" --index-url "https://download.pytorch.org/whl/nightly/cpu"
-  "$PIPBIN" install --pre torchao --index-url "https://download.pytorch.org/whl/nightly/cpu"
+  TORCH_VERSION=$(
+  "$PYBIN" - <<'PY'
+import runpy
+module_vars = runpy.run_path("torch_pin.py")
+print(module_vars["TORCH_VERSION"])
+PY
+)
+
+  NIGHTLY_VERSION=$(
+  "$PYBIN" - <<'PY'
+import runpy
+module_vars = runpy.run_path("torch_pin.py")
+print(module_vars["NIGHTLY_VERSION"])
+PY
+)
+  echo "=== [$LABEL] Install torch==${TORCH_VERSION}.${NIGHTLY_VERSION} ==="
+
+  # Install torchao based on the pinned PyTorch version
+  "$PIPBIN" install torch=="${TORCH_VERSION}.${NIGHTLY_VERSION}" --index-url "https://download.pytorch.org/whl/nightly/cpu"
+
+  # Install torchao based on the pinned commit from third-party/ao submodule
+  pushd "$REPO_ROOT/third-party/ao" > /dev/null
+  USE_CPP=0 "$PYBIN" setup.py develop
+  popd > /dev/null
 
   echo "=== [$LABEL] Import smoke tests ==="
   "$PYBIN" -c "import executorch; print('executorch imported successfully')"

.github/workflows/add-unanswered-to-project.yml

@@ -12,6 +12,7 @@ jobs:
       - name: Add open issues and open, non-draft PRs to org project (excluding certain authors)
         uses: actions/github-script@v7
         with:
+          github-token: ${{ secrets.ET_EXT_CONTRIB }}
           script: |
             const projectId = "PVT_kwDOAUB9vs4A_PUL"; // PyTorch org project 136
             const owner = 'pytorch';

CMakeLists.txt

@@ -729,9 +729,6 @@ 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()
@@ -740,6 +737,9 @@ 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)

backends/aoti/utils.h

@@ -36,6 +36,8 @@ inline executorch::aten::ScalarType dtype_to_scalar_type(int32_t dtype) {
   switch (dtype) {
     case 6: // PyTorch's float32 dtype code
       return executorch::aten::ScalarType::Float;
+    case 15: // PyTorch's bfloat16 dtype code
+      return executorch::aten::ScalarType::BFloat16;
     // Future support for additional dtypes can be added here
     default:
       ET_LOG(Error, "Unsupported dtype: %d for ScalarType conversion", dtype);

backends/apple/coreml/compiler/coreml_preprocess.py

@@ -6,6 +6,7 @@
 import logging
 
 import shutil
+import tempfile
 import uuid
 from dataclasses import asdict, dataclass
 from enum import Enum
@@ -415,7 +416,7 @@ def preprocess_model(
         mlmodel: ct.models.MLModel, model_type: MODEL_TYPE
     ) -> PreprocessResult:
         identifier = "executorch_" + str(uuid.uuid4())
-        dir_path: Path = Path("tmp") / identifier
+        dir_path: Path = Path(tempfile.gettempdir()) / identifier
         model_dir_path: Path = dir_path / "lowered_module"
         model_spec: ct.proto.Model_pb2 = mlmodel.get_spec()
         logger.warning(

backends/arm/_passes/__init__.py

@@ -91,6 +91,7 @@
     ReplaceScalarWithTensorArgPassTOSABI,
     ReplaceScalarWithTensorArgPassTOSAMI,
 )
+from .rewrite_upsample import RewriteUpsamplePass  # noqa
 from .scalars_to_attribute_pass import ScalarsToAttributePass  # noqa
 from .size_adjust_input_pass import SizeAdjustInputPass  # noqa
 from .to_tosa_memory_format_pass import ToTosaMemoryFormatPass  # noqa

backends/arm/_passes/annotate_decomposed_matmul.py

@@ -11,6 +11,9 @@
 
 import torch
 from executorch.backends.arm._passes.arm_pass_utils import create_node
+from executorch.backends.arm._passes.fold_qdq_with_annotated_qparams_pass import (
+    FoldAndAnnotateQParamsPass,
+)
 
 from executorch.backends.arm.constants import DQ_OPS, Q_OPS
 from executorch.exir.dialects._ops import ops as exir_ops
@@ -29,7 +32,7 @@ class AnnotateDecomposedMatmulPass(ExportPass):
     matmul-op (can be mm or bmm).
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = set()
+    _passes_required_after: Set[Type[ExportPass]] = {FoldAndAnnotateQParamsPass}
 
     def _match_partition_to_node(
         self, node: torch.fx.Node, partitioned_inputs: List[torch.fx.Node]

backends/arm/_passes/arm_pass_manager.py

@@ -91,6 +91,7 @@
     ReplaceScalarWithTensorArgPassTOSABI,
     ReplaceScalarWithTensorArgPassTOSAMI,
     RetraceFoldedDtypesPass,
+    RewriteUpsamplePass,
     ScalarsToAttributePass,
     SizeAdjustInputPass,
     ToTosaMemoryFormatPass,
@@ -112,6 +113,8 @@
 from executorch.exir.pass_manager import PassManager
 from executorch.exir.passes.remove_graph_asserts_pass import RemoveGraphAssertsPass
 from torch.fx import GraphModule
+from torch.fx.passes.infra.pass_base import PassResult
+from torch.nn.modules import Module
 
 
 class ArmPassManager(PassManager):
@@ -204,6 +207,7 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         # needs to happen before AddBiasPass, but after the table ops are inserted
         # to be able to validate that conv2d has right dtype arguments.
         self.add_pass(DecomposeConv2dWithInt16ActivationPass())
+        self.add_pass(RewriteUpsamplePass(exported_program))
         self.add_pass(AddBiasPass(exported_program))
 
         self.add_pass(FuseEqualPlaceholdersPass(exported_program))
@@ -288,6 +292,7 @@ def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(FuseViewCopyTransform())
         self.add_pass(FuseConstantArgsPass(exported_program))
         self.add_pass(CastInt64BuffersToInt32Pass(exported_program))
+        self.add_pass(RewriteUpsamplePass(exported_program))
         self.add_pass(AddBiasPass(exported_program))
         self.add_pass(InsertTableOpsPass(exported_program))
         self.add_pass(FuseEqualPlaceholdersPass(exported_program))
@@ -355,3 +360,20 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
             self.add_pass(DecomposeMaskedFill())
 
         return self._transform(graph_module)
+
+    def __call__(self, module: Module) -> PassResult:
+        try:
+            return super().__call__(module)
+        except Exception as e:
+            first_exception = e.__cause__ or e.__context__ or e
+            import re
+
+            message = e.args[0]
+            m = re.search(r"An error occurred when running the '([^']+)' pass", message)
+            if m:
+                pass_name = m.group(1)
+                first_exception.args = (
+                    f"{pass_name}: {first_exception.args[0]}",
+                    *first_exception.args[1:],
+                )
+            raise first_exception

backends/arm/_passes/conv1d_unsqueeze_pass.py

@@ -8,6 +8,9 @@
 
 from typing import Set, Type
 
+from executorch.backends.arm._passes.add_bias_pass import AddBiasPass
+from executorch.backends.arm._passes.size_adjust_input_pass import SizeAdjustInputPass
+
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass
 
@@ -23,7 +26,7 @@ class Conv1dUnsqueezePass(ExportPass):
     3) squeeze the output back down to 3d.
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = set()
+    _passes_required_after: Set[Type[ExportPass]] = {AddBiasPass, SizeAdjustInputPass}
 
     def call_operator(self, op, args, kwargs, meta):
         if op != exir_ops.edge.aten.convolution.default:

backends/arm/_passes/convert_any_default_dim_dims_pass.py

@@ -6,6 +6,9 @@
 from typing import Set, Type
 
 import torch
+from executorch.backends.arm._passes.convert_squeezes_to_view import (
+    ConvertSqueezesToViewPass,
+)
 from executorch.exir.dialects._ops import (  # type: ignore[import-not-found]
     ops as exir_ops,
 )
@@ -46,7 +49,7 @@ class ConvertAnyDefaultDimDimsPass(ExportPass):
         squeeze(dim = [dim1, dim2])
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = set()
+    _passes_required_after: Set[Type[ExportPass]] = {ConvertSqueezesToViewPass}
 
     def call(self, graph_module: torch.fx.GraphModule):
         modified = False