Update base for Update on "Introduce public MergedDataMap"

Add public merged data map. Module can use this to resolve multiple named data maps.

Differential Revision: [D83527299](https://our.internmc.facebook.com/intern/diff/D83527299/)

[ghstack-poisoned]

Author: lucylq

.ci/scripts/setup-openvino.sh

@@ -10,19 +10,17 @@ set -ex
 # shellcheck source=/dev/null
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
-git clone https://github.com/openvinotoolkit/openvino.git
-cd openvino && git checkout releases/2025/1
-git submodule update --init --recursive
-sudo ./install_build_dependencies.sh
-mkdir build && cd build
-cmake .. -DCMAKE_BUILD_TYPE=Release -DENABLE_PYTHON=ON
-make -j$(nproc)
+# Download and install OpenVINO from release packages
+OPENVINO_VERSION="2025.3"
+OPENVINO_BUILD="2025.3.0.19807.44526285f24"
+OPENVINO_URL="https://storage.openvinotoolkit.org/repositories/openvino/packages/${OPENVINO_VERSION}/linux/openvino_toolkit_ubuntu22_${OPENVINO_BUILD}_x86_64.tgz"
 
-cd ..
-cmake --install build --prefix dist
+curl -Lo /tmp/openvino_toolkit.tgz --retry 3 --fail ${OPENVINO_URL}
+tar -xzf /tmp/openvino_toolkit.tgz
+mv openvino_toolkit_ubuntu22_${OPENVINO_BUILD}_x86_64 openvino
 
-source dist/setupvars.sh
-cd ../backends/openvino
+source openvino/setupvars.sh
+cd backends/openvino
 pip install -r requirements.txt
 cd scripts
 ./openvino_build.sh --enable_python

.ci/scripts/test_openvino.sh

@@ -10,7 +10,7 @@ set -ex
 # shellcheck source=/dev/null
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
-source openvino/dist/setupvars.sh
+source openvino/setupvars.sh
 cd backends/openvino/tests
 python test_runner.py --test_type ops
 python test_runner.py --test_type models

.github/workflows/pull.yml

@@ -787,7 +787,6 @@ jobs:
       contents: read
     strategy:
       fail-fast: false
-    if: false # TODO Re-enable after fixing timeouts (#14314)
     with:
       runner: linux.2xlarge
       docker-image: ci-image:executorch-ubuntu-22.04-gcc9

.github/workflows/trunk.yml

@@ -823,11 +823,26 @@ jobs:
         echo "Recipe: $RECIPE"
         echo "Quantize: $QUANTIZE"
 
-        echo "::group::Set up ExecuTorch"
         # 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}"
-        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool cmake
+
+        echo "::group::Setup ExecuTorch"
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool "cmake"
+        echo "::endgroup::"
+
+        echo "::group::Setup Huggingface"
+        pip install -U "huggingface_hub[cli]" accelerate
+        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
+        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
+        pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
+        echo "::endgroup::"
+
+        echo "::group::Test MODEL: $MODEL RECIPE: $RECIPE QUANTIZE: $QUANTIZE"
+        export OUTPUT_DIR="$(pwd)/${MODEL}_${RECIPE}_${QUANTIZE}"
+        python .ci/scripts/test_huggingface_optimum_model.py --model "$MODEL" --recipe "$RECIPE" $QUANTIZE --model_dir "$OUTPUT_DIR"
+        echo "::endgroup::"
+
         # Build executor_runner with ETdump enabled
         PYTHON_EXECUTABLE=python cmake -DPYTHON_EXECUTABLE=python \
           -DCMAKE_INSTALL_PREFIX=cmake-out \
@@ -845,25 +860,6 @@ jobs:
           -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
           -Bcmake-out .
         cmake --build cmake-out -j16 --target install --config Release
-        echo "::endgroup::"
-
-        echo "::group::Set up Hugging Face"
-        pip install -U "huggingface_hub[cli]"
-        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
-        OPTIMUM_ET_COMMIT=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
-        git clone https://github.com/huggingface/optimum-executorch
-        pushd optimum-executorch
-        # There is no release yet, for CI stability, always test from the same commit on main
-        git checkout $OPTIMUM_ET_COMMIT
-        python install_dev.py --skip_override_torch
-        popd
-        pip list
-        echo "::endgroup::"
-
-        echo "::group::Run tests"
-        export OUTPUT_DIR="$(pwd)/${MODEL}_${RECIPE}_${QUANTIZE}"
-        python .ci/scripts/test_huggingface_optimum_model.py --model ${MODEL} --recipe ${RECIPE} ${QUANTIZE} --model_dir ${OUTPUT_DIR}
-        echo "::endgroup::"
 
         echo "::group::Generate artifacts for performance profiling"
         ./cmake-out/executor_runner \
@@ -930,16 +926,11 @@ jobs:
         ${CONDA_RUN} python install_executorch.py
         echo "::endgroup::"
 
-        echo "::group::Set up Hugging Face"
-        pip install -U "huggingface_hub[cli]"
-        huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
-        OPTIMUM_ET_COMMIT=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
-        git clone https://github.com/huggingface/optimum-executorch
-        pushd optimum-executorch
-        # There is no release yet, for CI stability, always test from the same commit on main
-        git checkout $OPTIMUM_ET_COMMIT
-        ${CONDA_RUN} python install_dev.py --skip_override_torch
-        popd
+        echo "::group::Set up Huggingface"
+        ${CONDA_RUN} pip install -U "huggingface_hub[cli]" accelerate
+        ${CONDA_RUN} huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
+        OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt)
+        ${CONDA_RUN} pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION}
         ${CONDA_RUN} pip list
         echo "::endgroup::"
 

backends/arm/_passes/annotate_decomposed_matmul.py

@@ -73,7 +73,10 @@ def call(self, graph_module: GraphModule) -> PassResult:
                 node for node in partition.nodes if node.target in matmul_targets
             ][0]
 
-            if quantized_input:
+            if quantized_input and not all(
+                input_node.target in DQ_OPS
+                for input_node in matmul_node.all_input_nodes
+            ):
                 matmul_args = matmul_node.all_input_nodes
                 for node in matmul_args:
                     # Find the dq-node connected to this mm/bmm arg
@@ -99,7 +102,9 @@ def call(self, graph_module: GraphModule) -> PassResult:
 
             partition_output = list(partition.output_nodes[0].users)[0]
             quantized_output = partition_output.target in Q_OPS
-            if quantized_output:
+            if quantized_output and not all(
+                user.target in Q_OPS for user in matmul_node.users
+            ):
                 with graph_module.graph.inserting_after(matmul_node):
                     # Create q-node after matmul
                     q_node = create_node(

backends/arm/test/ops/test_matmul.py

@@ -22,6 +22,7 @@
 
 class MatMul(torch.nn.Module):
     test_data_generators = {
+        "rand_rand_2d": lambda: (torch.rand(5, 5), torch.rand(5, 2)),
         "rand_rand_3d": lambda: (torch.rand(2, 3, 5), torch.rand(2, 5, 2)),
         "rand_rand_4d": lambda: (torch.rand(1, 2, 3, 5), torch.rand(1, 2, 5, 2)),
     }
@@ -32,6 +33,7 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
 
 class MatMulSingleInput(torch.nn.Module):
     test_data_generators = {
+        "rand_2d": lambda: (torch.rand(5, 5),),
         "rand_3d": lambda: (torch.rand(2, 5, 5),),
         "rand_4d": lambda: (torch.rand(1, 2, 5, 5),),
     }
@@ -42,6 +44,11 @@ def forward(self, x: torch.Tensor):
 
 class MatMulCombo(torch.nn.Module):
     test_data_generators = {
+        "rand_rand_rand_2d": lambda: (
+            torch.rand(5, 5),
+            torch.rand(5, 2),
+            torch.rand(2, 5),
+        ),
         "rand_rand_rand_3d": lambda: (
             torch.rand(2, 5, 5),
             torch.rand(2, 5, 2),

backends/cadence/aot/replace_ops.py

@@ -1590,7 +1590,7 @@ def call_operator(self, op, args, kwargs, meta):
         updated_args = list(args)
         for op_arg_index in args_to_be_replaced:
             arg = args[op_arg_index]
-            if not arg.is_tensor():
+            if not isinstance(arg, ProxyValue) or not arg.is_tensor():
                 return super().call_operator(op, args, kwargs, meta)
 
             if not isinstance(arg.node.target, EdgeOpOverload):

backends/openvino/partitioner.py

@@ -27,6 +27,10 @@
 
 
 class OpenvinoOperatorsSupport(OperatorSupportBase):
+    extended_support_dict = {
+        "torch.ops.dim_order_ops._clone_dim_order.default": None,
+        "torch.ops.dim_order_ops._to_dim_order_copy.default": None,
+    }
 
     def __init__(
         self,
@@ -62,7 +66,9 @@ def is_node_supported(self, _, node: torch.fx.Node) -> bool:
             op_type = node.target.__name__
         else:
             op_type = str(node.target)
-        supported_ops = OperatorSupport(options)._support_dict
+        supported_ops = (
+            OperatorSupport(options)._support_dict | self.extended_support_dict
+        )
         if op_type == "getitem":
             return True
 

backends/openvino/preprocess.py

@@ -14,6 +14,8 @@
     PreprocessResult,
 )
 from executorch.exir.backend.compile_spec_schema import CompileSpec
+
+from executorch.exir.passes.memory_format_ops_pass import DimOrderOpsRevertPass
 from openvino.frontend.pytorch.torchdynamo.compile import (  # type: ignore[import-untyped]
     openvino_compile,
 )
@@ -36,6 +38,12 @@ def preprocess(
         Returns:
             PreprocessResult: The result of preprocessing, including the compiled model bytes.
         """
+        transformed_ep = DimOrderOpsRevertPass()(edge_program.graph_module)
+
+        # Update the edge_program with the transformed graph
+        if transformed_ep and transformed_ep.graph_module:
+            edge_program._graph_module = transformed_ep.graph_module
+
         input_names = edge_program.graph_signature.user_inputs
         args = []
         for node in edge_program.graph.nodes:

export/export.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# 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.
@@ -200,7 +201,9 @@ def _build_stages(self, stages: List[StageType]) -> Dict[StageType, Stage]:
                     aten_transform_passes = list(
                         self._export_recipe.aten_transform_passes
                     )
-                stage = TorchExportStage(aten_transform_passes)
+                stage = TorchExportStage(
+                    aten_transform_passes, strict=self._export_recipe.strict
+                )
             elif stage_type == StageType.TO_EDGE_TRANSFORM_AND_LOWER:
                 stage = EdgeTransformAndLowerStage.from_recipe(self._lowering_recipe)
             elif stage_type == StageType.TO_EDGE: