Update on "[ExecuTorch] Separate extension.Module Namespaces from Aten and non-Aten"

# Context

Separate `extension.Module` Namespaces to be `executorch::extension::module` and `executorch::extension::module::aten`, otherwise if a package relies on both aten and non-aten of the same implementation and the namespace is the same, there will be duplicate symbol issue like:

```
ld.lld: error: duplicate symbol: vtable for executorch::extension::Module
>>> defined at {redacted}/executorch/extension/module/__module__/__stripped__/module.cpp.pic.stripped.o:(vtable for executorch::extension::Module)
>>> defined at {redacted}/executorch/extension/module/__module_aten__/__stripped__/module.cpp.pic.stripped.o:
```

# Proposal

Doing something similar to what we already did for `bundled_program` in #10307

Since `extension.Module` is a public API, we introduce a namespace alias, so that existing use cases won't get affected. Since namespace alias doesn't create additional symbols, there won't be duplicate symbol issue.

Differential Revision: [D73903870](https://our.internmc.facebook.com/intern/diff/D73903870/)

[ghstack-poisoned]

Author: zhenyan-zhang-meta

.ci/scripts/build-qnn-sdk.sh

@@ -33,6 +33,7 @@ set_up_aot() {
   cmake .. \
       -DCMAKE_INSTALL_PREFIX=$PWD \
       -DEXECUTORCH_BUILD_QNN=ON \
+      -DANDROID_NATIVE_API_LEVEL=30 \
       -DQNN_SDK_ROOT=${QNN_SDK_ROOT} \
       -DEXECUTORCH_BUILD_DEVTOOLS=ON \
       -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \

.ci/scripts/test_model.sh

@@ -222,7 +222,7 @@ test_model_with_coreml() {
 
   DTYPE=float16
 
-  "${PYTHON_EXECUTABLE}" -m examples.apple.coreml.scripts.export --model_name="${MODEL_NAME}" --compute_precision "${DTYPE}"
+  "${PYTHON_EXECUTABLE}" -m examples.apple.coreml.scripts.export --model_name="${MODEL_NAME}" --compute_precision "${DTYPE}" --use_partitioner
   EXPORTED_MODEL=$(find "." -type f -name "${MODEL_NAME}*.pte" -print -quit)
 
   if [ -n "$EXPORTED_MODEL" ]; then

Package.swift

@@ -77,7 +77,9 @@ let package = Package(
         name: "\(key)_dependencies",
         dependencies: [.target(name: key)],
         path: ".Package.swift/\(key)",
-        linkerSettings:
+        linkerSettings: [
+          .linkedLibrary("c++")
+        ] +
           (value["frameworks"] as? [String] ?? []).map { .linkedFramework($0) } +
           (value["libraries"] as? [String] ?? []).map { .linkedLibrary($0) }
       ),

backends/apple/coreml/runtime/delegate/coreml_backend_delegate.mm

@@ -88,9 +88,17 @@
         ET_LOG(Error, "%s: DataType=%d is not supported", ETCoreMLStrings.delegateIdentifier.UTF8String, (int)tensor.scalar_type());
         return std::nullopt;
     }
-
+    
     std::vector<ssize_t> strides(tensor.strides().begin(), tensor.strides().end());
     std::vector<size_t> shape(tensor.sizes().begin(), tensor.sizes().end());
+    
+    // If tensor is rank 0, wrap in rank 1
+    // See https://github.com/apple/coremltools/blob/8.2/coremltools/converters/mil/frontend/torch/exir_utils.py#L73
+    if (shape.size() == 0) {
+        shape.push_back(1);
+        strides.push_back(1);
+    }
+    
     MultiArray::MemoryLayout layout(dataType.value(), std::move(shape), std::move(strides));
     switch (argType) {
         case ArgType::Input: {
@@ -233,6 +241,12 @@ ModelLoggingOptions get_logging_options(BackendExecutionContext& context) {
     std::array<SizesType, kTensorDimensionLimit> new_shape;
     for (size_t i = nInputs; i < nInputs + nOutputs; i++) {
         Tensor& t = args[i]->toTensor();
+        // If t has rank 0, do not resize.  delegate_args[i] will have rank 1
+        // because we resized it in get_multi_array
+        if (t.dim() == 0) {
+            continue;
+        }
+
         int rank = delegate_args[i].layout().rank();
         assert (rank <= new_shape.size());
         for (int d = 0; d < rank; d++) {

backends/apple/coreml/runtime/test/ETCoreMLModelManagerTests.mm

@@ -113,7 +113,7 @@ - (void)testAddModelExecution {
     XCTAssertNotNil(inputs);
     MLMultiArray *output = [ETCoreMLTestUtils filledMultiArrayWithShape:inputs[0].shape dataType:inputs[0].dataType repeatedValue:@(0) error:&localError];
     NSArray<MLMultiArray *> *args = [inputs arrayByAddingObject:output];
-    XCTAssertTrue([self.modelManager executeModelWithHandle:handle 
+    XCTAssertTrue([self.modelManager executeModelWithHandle:handle
                                                        args:args
                                              loggingOptions:executorchcoreml::ModelLoggingOptions()
                                                 eventLogger:nullptr

backends/apple/coreml/scripts/install_requirements.sh

@@ -12,7 +12,7 @@ SCRIPT_DIR_PATH="$(
 
 # TODO(jathu): remove the need to fetch coremltools to build deps for coreml_executor_runner.
 # Keep this version in sync with: pyproject.toml
-COREMLTOOLS_VERSION="8.2"
+COREMLTOOLS_VERSION="8.3"
 
 red=`tput setaf 1`
 green=`tput setaf 2`

backends/apple/mps/setup.md

@@ -76,12 +76,12 @@ cd executorch
 ## Run the mv3 generated model using the mps_executor_runner
 
 ```bash
-./cmake-out/examples/apple/mps/mps_executor_runner --model_path mv3_mps_bundled_fp16.pte --bundled_program
+./cmake-out/examples/apple/mps/mps_executor_runner --model_path mv3_mps_float16_bundled.pte --bundled_program
 ```
 
 - You should see the following results. Note that no output file will be generated in this example:
 ```
-I 00:00:00.003290 executorch:mps_executor_runner.mm:286] Model file mv3_mps_bundled_fp16.pte is loaded.
+I 00:00:00.003290 executorch:mps_executor_runner.mm:286] Model file mv3_mps_float16_bundled.pte is loaded.
 I 00:00:00.003306 executorch:mps_executor_runner.mm:292] Program methods: 1
 I 00:00:00.003308 executorch:mps_executor_runner.mm:294] Running method forward
 I 00:00:00.003311 executorch:mps_executor_runner.mm:349] Setting up non-const buffer 1, size 606112.
@@ -118,7 +118,7 @@ python3 -m examples.apple.mps.scripts.mps_example --model_name="mv3" --generate_
 ```
 2. Run your Program on the ExecuTorch runtime and generate an [ETDump](../../../docs/source/etdump.md).
 ```
-./cmake-out/examples/apple/mps/mps_executor_runner --model_path mv3_mps_bundled_fp16.pte --bundled_program --dump-outputs
+./cmake-out/examples/apple/mps/mps_executor_runner --model_path mv3_mps_float16_bundled.pte --bundled_program --dump-outputs
 ```
 3. Create an instance of the Inspector API by passing in the ETDump you have sourced from the runtime along with the optionally generated ETRecord from step 1.
 ```bash

backends/arm/_passes/arm_pass_manager.py

@@ -59,6 +59,9 @@
 )
 
 from executorch.backends.arm.tosa_specification import Tosa_0_80, TosaSpecification
+from executorch.backends.transforms.decompose_sdpa import (
+    DecomposeScaledDotProductAttention,
+)
 from executorch.backends.transforms.fuse_view_copy import FuseViewCopyTransform
 from executorch.backends.xnnpack._passes.remove_getitem_op import RemoveGetItemPass
 from executorch.exir import ExportedProgram
@@ -194,6 +197,7 @@ def transform_to_backend_pipeline(self, exported_program: ExportedProgram):
             )
 
     def transform_for_annotation_pipeline(self, graph_module: GraphModule):
+        self.add_pass(DecomposeScaledDotProductAttention())
         self.add_pass(ReplaceScalarWithTensorArgPassTOSABI())
         self.add_pass(ScalarsToAttributePass())
         self.add_pass(DecomposeLayerNormPass())

backends/arm/_passes/decompose_softmax_pass.py

@@ -8,7 +8,11 @@
 from executorch.exir.pass_base import ExportPass
 
 # For BI case
-torch_softmax = (torch.ops.aten.softmax.int, torch.ops.aten.log_softmax.int)
+torch_softmax = (
+    torch.ops.aten.softmax.int,
+    torch.ops.aten._safe_softmax.default,
+    torch.ops.aten.log_softmax.int,
+)
 # For MI case
 edge_softmax = (
     exir_ops.edge.aten._softmax.default,

backends/arm/operator_support/convolution_support.py

@@ -11,7 +11,11 @@
     register_tosa_support_check,
     SupportedTOSAOperatorCheck,
 )
-from executorch.backends.arm.tosa_specification import Tosa_0_80, TosaSpecification
+from executorch.backends.arm.tosa_specification import (
+    Tosa_0_80,
+    Tosa_1_00,
+    TosaSpecification,
+)
 from executorch.exir.dialects._ops import ops as exir_ops
 
 
@@ -43,6 +47,9 @@ def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
 
         # Hardware specific constraints
         if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
+            # TODO remove this once TOSA 1.0 support for u55 is added.
+            if isinstance(tosa_spec, Tosa_1_00) and "u55" in tosa_spec.extensions:
+                return False
             return True
         else:
             return self._is_node_supported_u55(node)