Update base for Update on "Integrate torchgen exception boundary with ExecuTorch"

As of #7746, we build with exceptions by default, so we just need to use them.

Differential Revision: [D67904052](https://our.internmc.facebook.com/intern/diff/D67904052/)

[ghstack-poisoned]

Author: Github Executorch

.buckconfig

@@ -33,3 +33,6 @@
       **/.git, \
       cmake-out, \
       pip-out
+
+[buck2]
+restarter=true

.ci/scripts/gather_benchmark_configs.py

@@ -238,7 +238,7 @@ def set_output(name: str, val: Any) -> None:
     try:
         with open(github_output, "a") as env:
             env.write(f"{name}={val}\n")
-    except PermissionError:
+    except (PermissionError, FileNotFoundError):
         # Fall back to printing in case of permission error in unit tests
         print(f"::set-output name={name}::{val}")
 

.ci/scripts/setup-linux.sh

@@ -23,4 +23,7 @@ fi
 # of nightly. This allows CI to test against latest commits from PyTorch
 install_executorch "use-pt-pinned-commit"
 build_executorch_runner "${BUILD_TOOL}"
-do_not_use_nightly_on_ci
+
+if [[ "${GITHUB_BASE_REF:-}" == *main* || "${GITHUB_BASE_REF:-}" == *gh* ]]; then
+  do_not_use_nightly_on_ci
+fi

.ci/scripts/setup-macos.sh

@@ -137,4 +137,7 @@ install_pytorch_and_domains
 # the pinned commit from PyTorch
 install_executorch "use-pt-pinned-commit"
 build_executorch_runner "${BUILD_TOOL}"
-do_not_use_nightly_on_ci
+
+if [[ "${GITHUB_BASE_REF:-}" == *main* || "${GITHUB_BASE_REF:-}" == *gh* ]]; then
+  do_not_use_nightly_on_ci
+fi

.gitmodules

@@ -64,3 +64,6 @@
 [submodule "third-party/ao"]
 	path = third-party/ao
 	url = https://github.com/pytorch/ao.git
+[submodule "backends/cadence/utils/FACTO"]
+	path = backends/cadence/utils/FACTO
+	url = https://github.com/pytorch-labs/FACTO.git

backends/arm/_passes/_debug_passes.py

@@ -0,0 +1,23 @@
+# 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.
+
+import torch
+from executorch.devtools.visualization.visualization_utils import visualize_graph
+from executorch.exir import ExportedProgram
+from executorch.exir.pass_base import ExportPass, PassResult
+
+
+class VisualizePass(ExportPass):
+    """
+    This pass visualizes the graph at the point of insertion in the pass manager
+    """
+
+    def __init__(self, exported_program: ExportedProgram) -> None:
+        super().__init__()
+        self.exported_program = exported_program
+
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        visualize_graph(graph_module, self.exported_program)
+        return PassResult(graph_module, False)

backends/arm/_passes/annotate_decomposed_matmul.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import itertools
 
 from typing import List

backends/arm/_passes/arm_pass_manager.py

@@ -123,6 +123,7 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(FuseQuantizedActivationPass())
         self.add_pass(RemoveGetItemPass())
         self.add_pass(ConvertSplitToSlicePass())
+        self.add_pass(FuseBatchnorm2DPass(exported_program))
         self.add_pass(ConvertMmToBmmPass())
         self.add_pass(DecomposeLinearPass())
         self.add_pass(DecomposeBatchNormPass())
@@ -132,7 +133,6 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(ConvertMeanDimToAveragePoolPass())
         self.add_pass(DecomposeDivPass())
         self.add_pass(DecomposeSoftmaxesPass())
-        self.add_pass(FuseBatchnorm2DPass(exported_program))
 
         self.add_pass(AnnotateDecomposedMatmulPass())
         self.add_pass(QuantizeOperatorArguments())

backends/arm/_passes/fold_qdq_with_annotated_qparams_pass.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import copy
 
 from typing import cast, Dict, Set, Tuple

backends/arm/_passes/fuse_batchnorm2d_pass.py

@@ -3,6 +3,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import torch
 from executorch.exir import ExportedProgram
 from executorch.exir.dialects._ops import ops as exir_ops

backends/arm/_passes/fuse_quantized_activation_pass.py

@@ -3,6 +3,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import torch
 from executorch.backends.arm.tosa_quant_utils import q_op
 from executorch.exir.dialects._ops import ops as exir_ops

backends/arm/_passes/insert_table_ops.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 from typing import Callable, Dict
 
 import torch

backends/arm/_passes/mm_to_bmm_pass.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 import torch
 from executorch.backends.arm._passes.arm_pass_utils import (
     create_node,

backends/arm/operator_support/__init__.py

@@ -1,8 +1,15 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 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.
 
 # pyre-unsafe
 
-from . import right_shift_support, to_copy_support, tosa_supported_operators  # noqa
+from . import (  # noqa
+    convolution_support,
+    pool_2d_support,
+    reduce_sum_support,
+    right_shift_support,
+    to_copy_support,
+    tosa_supported_operators,
+)

backends/arm/operator_support/convolution_support.py

@@ -0,0 +1,99 @@
+# 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.
+
+from typing import cast
+
+import torch
+import torch.fx as fx
+from executorch.backends.arm.operator_support.tosa_supported_operators import (
+    register_tosa_support_check,
+    SupportedTOSAOperatorCheck,
+)
+from executorch.backends.arm.tosa_specification import Tosa_0_80, TosaSpecification
+from executorch.exir.dialects._ops import ops as exir_ops
+
+
+@register_tosa_support_check
+class ConvolutionSupported(SupportedTOSAOperatorCheck):
+    targets = [exir_ops.edge.aten.convolution.default]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+
+        # Not implemented
+        transposed = cast(bool, node.args[6])
+        output_padding = cast(list[int], node.args[7])
+        if transposed:
+            return False
+
+        for pad in output_padding:
+            if pad != 0:
+                return False
+
+        # Hardware specific constraints
+        if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
+            return True
+        else:
+            return self._is_node_supported_u55(node)
+
+    def _is_node_supported_u55(self, node: fx.Node):
+        """Hardware constraints for Ethos-U-55 case, Vela 4.2.0 (25.02 release)"""
+
+        shape_in = cast(torch.Tensor, node.all_input_nodes[0].meta["val"]).shape
+        shape_out = node.meta["val"].shape
+        kernel = cast(fx.Node, node.args[1]).meta["val"].shape
+        group = cast(int, node.args[8])
+
+        C_in = shape_in[1]
+        C_out = shape_out[1]
+        if (C_in == group) and (C_out % C_in) == 0:
+            # Depthwise convolution
+            for dim in shape_in[1:]:
+                if not 1 <= dim <= 65536:
+                    return False
+        else:
+            # Convolution
+            if not 1 <= C_in <= 65536:
+                return False
+
+        kernel_w = kernel[2]
+        kernel_h = kernel[3] if len(kernel) > 3 else 1
+        # Kernel condition misses constraint on sum of absolute weights
+        if not 1 <= kernel_h <= 64 or not 1 <= kernel_w * kernel_h <= 4096:
+            return False
+
+        if not self._stride_condition(node):
+            return False
+
+        return True
+
+    def _stride_condition(self, node: fx.Node) -> bool:
+        """This condition is somewhat complex but boils down
+        to not supporting stride > 3, unless we have some special conditions.
+        This condition is a simplified, relaxed version of the hardware constraint,
+        since the actual constraint requires information not available
+        here (without a lot of work).
+
+        This means that we might accept ops that are not actually supported.
+        """
+        strides = cast(list[int], node.args[3])
+        has_padding = any(pad > 0 for pad in cast(list[int], node.args[4]))
+        dilations = cast(list[int], node.args[5])
+        if len(dilations) == 1:
+            dilations = [dilations[0]] * 2
+        if len(strides) == 1:
+            strides = [strides[0]] * 2
+
+        for stride, dilation in zip(strides, dilations):
+            stride_condition = 1 <= stride <= 3
+            dilation_condition = (not has_padding) and (dilation == 1)
+            if (not stride_condition) and (not dilation_condition):
+                return False
+
+        return True

backends/arm/operator_support/pool_2d_support.py

@@ -0,0 +1,85 @@
+# 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.
+
+from typing import cast
+
+import torch
+import torch.fx as fx
+from executorch.backends.arm.operator_support.tosa_supported_operators import (
+    register_tosa_support_check,
+    SupportedTOSAOperatorCheck,
+)
+from executorch.backends.arm.tosa_specification import Tosa_0_80, TosaSpecification
+from executorch.exir.dialects._ops import ops as exir_ops
+
+
+def kernel_check(kernel: tuple[int, int]) -> bool:
+    if not (1 <= kernel[0] * kernel[1] <= 65536):
+        return False
+    return 1 <= kernel[1] <= 256
+
+
+def stride_check(strides: tuple[int, int]) -> bool:
+    return all(1 <= stride <= 3 for stride in strides)
+
+
+def dim_check(shape=torch.Size) -> bool:
+    check = shape[0] == 1
+    for dim in shape:
+        check &= 1 <= dim <= 65536
+    return check
+
+
+@register_tosa_support_check
+class AvgPool2dSupported(SupportedTOSAOperatorCheck):
+    targets = [
+        exir_ops.edge.aten.avg_pool2d.default,
+    ]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+        if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
+            return True
+
+        # U55 case, Vela 4.2.0 (25.02 release)
+        shape = cast(torch.Tensor, node.all_input_nodes[0].meta["val"]).shape
+        kernel = cast(tuple[int, int], node.args[1])
+        stride = cast(tuple[int, int], node.args[2])
+        if len(node.args) > 3:
+            # Padding case
+            if not all(1 <= k <= 8 for k in kernel):
+                return False
+        else:
+            if not kernel_check(kernel):
+                return False
+
+        return dim_check(shape) and stride_check(stride)
+
+
+@register_tosa_support_check
+class MaxPool2dSupported(SupportedTOSAOperatorCheck):
+    targets = [
+        exir_ops.edge.aten.max_pool2d_with_indices.default,
+    ]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+        if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
+            return True
+
+        # U55 case, Vela 4.2.0 (25.02 release)
+        shape = cast(torch.Tensor, node.all_input_nodes[0].meta["val"]).shape
+        kernel = cast(tuple[int, int], node.args[1])
+        stride = cast(tuple[int, int], node.args[2])
+
+        return kernel_check(kernel) and dim_check(shape) and stride_check(stride)

backends/arm/operator_support/reduce_sum_support.py

@@ -0,0 +1,51 @@
+# 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.
+
+from typing import cast
+
+import torch.fx as fx
+from executorch.backends.arm.operator_support.tosa_supported_operators import (
+    register_tosa_support_check,
+    SupportedTOSAOperatorCheck,
+)
+from executorch.backends.arm.tosa_specification import Tosa_0_80, TosaSpecification
+from executorch.exir.dialects._ops import ops as exir_ops
+
+
+@register_tosa_support_check
+class SumSupported(SupportedTOSAOperatorCheck):
+    targets = [exir_ops.edge.aten.sum.dim_IntList]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification):
+        if not (isinstance(tosa_spec, Tosa_0_80) and tosa_spec.is_U55_subset):
+            return True
+
+        # U55 case, Vela 4.2.0 (25.02 release)
+        input_shape = node.all_input_nodes[0].meta["val"].shape
+        dim_list = cast(list[int], node.args[1])
+        dim_list = [dim % len(input_shape) for dim in dim_list]
+
+        for dim in dim_list:
+            if not 1 <= input_shape[dim] <= 65536:
+                return False
+
+            # We can't be certain of which dim is the last in memory yet,
+            # Always go for stricter condition.
+            pre_R_product = 1.0
+            for length in input_shape[:dim]:
+                pre_R_product *= length
+            post_R_product = 1.0
+            for length in input_shape[dim + 1 :]:
+                post_R_product *= length
+            if not 1 <= pre_R_product <= 65536:
+                return False
+            if not 1 <= post_R_product <= 65536:
+                return False
+        return True

backends/arm/operator_support/right_shift_support.py

@@ -3,6 +3,8 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+# pyre-unsafe
+
 
 import logging