Update on "Add U55 and U85 16A8W linear tests"

Add U55 and U85 16A8W quantization tests for linear operations following the pattern established in D80510628 for mul operations.

This adds test coverage for 16A8W (16-bit activations, 8-bit weights) quantization on both Ethos-U55 and Ethos-U85 platforms for linear operations. Both tests are marked with xfail annotations due to the known Vela compilation issue where int16 linear operations fail with "Conv2D Invalid arguments" error.

The tests follow the same structure as the existing 16A8W TOSA test but target specific hardware platforms:
- test_linear_16a8w_u55_INT16: Tests on U55 with rank1 data
- test_linear_16a8w_u85_INT16: Tests on U85 with both rank1 and rank4 data

Both tests use the same symmetric A16W8 quantization configuration and are properly decorated with platform-specific skip conditions (XfailIfNoCorstone300/320) and pytest xfail markers referencing GitHub issue 13947.

Addresses: #13729

Differential Revision: [D82618039](https://our.internmc.facebook.com/intern/diff/D82618039/)

[ghstack-poisoned]

Author: Ninja91

.ci/scripts/test_model.sh

@@ -131,13 +131,13 @@ test_model_with_xnnpack() {
     return 0
   fi
 
-  # Delegation
+  # Delegation and test with pybindings
   if [[ ${WITH_QUANTIZATION} == true ]]; then
     SUFFIX="q8"
-    "${PYTHON_EXECUTABLE}" -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate --quantize
+    "${PYTHON_EXECUTABLE}" -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate --quantize --test_after_export
   else
     SUFFIX="fp32"
-    "${PYTHON_EXECUTABLE}" -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate
+    "${PYTHON_EXECUTABLE}" -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate --test_after_export
   fi
 
   OUTPUT_MODEL_PATH="${MODEL_NAME}_xnnpack_${SUFFIX}.pte"

.ci/scripts/test_wheel_package_qnn.sh

@@ -145,6 +145,7 @@ run_core_tests () {
   echo "=== [$LABEL] Import smoke tests ==="
   "$PYBIN" -c "import executorch; print('executorch imported successfully')"
   "$PYBIN" -c "import executorch.backends.qualcomm; print('executorch.backends.qualcomm imported successfully')"
+  "$PYBIN" -c "from executorch.export.target_recipes import get_android_recipe; recipe = get_android_recipe('android-arm64-snapdragon-fp16'); print(f'executorch.export.target_recipes imported successfully: {recipe}')"
 
   echo "=== [$LABEL] List installed executorch/backends/qualcomm/python ==="
   local SITE_DIR

.ci/scripts/wheel/test_base.py

@@ -41,6 +41,18 @@ class ModelTest:
 
 
 def run_tests(model_tests: List[ModelTest]) -> None:
+    # Test that we can import the portable_lib module - verifies RPATH is correct
+    print("Testing portable_lib import...")
+    try:
+        from executorch.extension.pybindings._portable_lib import (  # noqa: F401
+            _load_for_executorch,
+        )
+
+        print("✓ Successfully imported _load_for_executorch from portable_lib")
+    except ImportError as e:
+        print(f"✗ Failed to import portable_lib: {e}")
+        raise
+
     # Why are we doing this envvar shenanigans? Since we build the testers, which
     # uses buck, we cannot run as root. This is a sneaky of getting around that
     # test.

CMakeLists.txt

@@ -869,6 +869,21 @@ if(EXECUTORCH_BUILD_PYBIND)
   target_compile_options(portable_lib PUBLIC ${_pybind_compile_options})
   target_link_libraries(portable_lib PRIVATE ${_dep_libs})
 
+  # Set RPATH to find PyTorch libraries relative to the installation location
+  # This goes from executorch/extension/pybindings up to site-packages, then to
+  # torch/lib
+  if(APPLE)
+    set_target_properties(
+      portable_lib PROPERTIES BUILD_RPATH "@loader_path/../../../torch/lib"
+                              INSTALL_RPATH "@loader_path/../../../torch/lib"
+    )
+  else()
+    set_target_properties(
+      portable_lib PROPERTIES BUILD_RPATH "$ORIGIN/../../../torch/lib"
+                              INSTALL_RPATH "$ORIGIN/../../../torch/lib"
+    )
+  endif()
+
   install(
     TARGETS portable_lib
     EXPORT ExecuTorchTargets

backends/arm/_passes/add_bias_pass.py

@@ -3,13 +3,15 @@
 # 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 Set, Type
+
 import torch
 from executorch.backends.arm._passes import ArmPass
 from executorch.backends.arm._passes.arm_pass_utils import get_first_fake_tensor
 from executorch.backends.transforms.utils import create_constant_placeholder
 
 from executorch.exir.dialects._ops import ops as exir_ops
-from executorch.exir.pass_base import PassResult
+from executorch.exir.pass_base import ExportPass, PassResult
 from torch.export.graph_signature import InputKind
 
 
@@ -19,6 +21,8 @@ class AddBiasPass(ArmPass):
     The bias is set to zero.
     """
 
+    _passes_required_after: Set[Type[ExportPass]] = set()
+
     targeted_ops = (exir_ops.edge.aten.convolution.default,)
 
     def call(self, graph_module):

backends/arm/_passes/annotate_decomposed_matmul.py

@@ -7,7 +7,7 @@
 
 import itertools
 import operator
-from typing import cast, List
+from typing import cast, List, Set, Type
 
 import torch
 from executorch.backends.arm._passes.arm_pass_utils import create_node
@@ -29,6 +29,8 @@ class AnnotateDecomposedMatmulPass(ExportPass):
     matmul-op (can be mm or bmm).
     """
 
+    _passes_required_after: Set[Type[ExportPass]] = set()
+
     def _match_partition_to_node(
         self, node: torch.fx.Node, partitioned_inputs: List[torch.fx.Node]
     ) -> torch.fx.Node:

backends/arm/_passes/annotate_output_dim_order_pass.py

@@ -3,9 +3,12 @@
 # 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 Set, Type
+
 from executorch.backends.arm._passes import ArmPass
 from executorch.backends.arm._passes.arm_pass_utils import get_output_dim_orders
-from executorch.exir.pass_base import PassResult
+from executorch.exir.pass_base import ExportPass, PassResult
 
 
 class AnnotateOutputDimOrderPass(ArmPass):
@@ -14,6 +17,8 @@ class AnnotateOutputDimOrderPass(ArmPass):
     for verifying that the dim order does not change unexpectedly in later passes.
     """
 
+    _passes_required_after: Set[Type[ExportPass]] = set()
+
     def call(self, graph_module):
         output_node = graph_module.graph.output_node()
         output_node.meta["original_dim_orders"] = get_output_dim_orders(graph_module)

backends/arm/_passes/arm_pass.py

@@ -6,7 +6,8 @@
 # pyre-unsafe
 
 import traceback
-from typing import Optional
+from abc import abstractmethod
+from typing import List, Optional, Set, Type
 
 import torch
 from executorch.exir.pass_base import ExportPass, NodeMetadata
@@ -19,6 +20,36 @@ def __init__(self, exported_program: Optional[torch.export.ExportedProgram] = No
         super(ArmPass, self).__init__()
         self.exported_program = exported_program
 
+    @property
+    @abstractmethod
+    def _passes_required_after(self) -> Set[Type[ExportPass]]:
+        """The subclass defines passes that must run after it"""
+        pass
+
+    @staticmethod
+    def get_required_passes(pass_) -> List[str]:
+        """
+        Returns the list of passes that must be run after this pass, sorted by name.
+        """
+        if hasattr(pass_, "_passes_required_after"):
+            return sorted([ArmPass.get_name(p) for p in pass_._passes_required_after])
+        else:
+            return []
+
+    @staticmethod
+    def get_name(pass_) -> str:
+        """
+        Returns the name of the pass.
+        """
+        if isinstance(pass_, ExportPass):
+            return pass_.__class__.__name__
+        elif hasattr(pass_, "__name__"):
+            return pass_.__name__
+        else:
+            raise ValueError(
+                f"Cannot get name for pass: {pass_}. It must be an instance of ExportPass or have a __name__ attribute."
+            )
+
     def call_operator(self, op, args, kwargs, meta, updated: Optional[bool] = False):
         if not updated:
             return super().call_operator(op, args, kwargs, meta)

backends/arm/_passes/arm_pass_manager.py

@@ -7,6 +7,9 @@
 
 # pyre-unsafe
 
+
+from collections import defaultdict
+
 import executorch.backends.arm.tosa.dialect  # noqa: unused
 from executorch.backends.arm._passes import (
     AddBiasPass,
@@ -94,6 +97,7 @@
     UnsqueezeScalarPlaceholdersPass,
 )
 
+from executorch.backends.arm._passes.arm_pass import ArmPass
 from executorch.backends.arm.tosa.specification import (
     TosaLoweringContext,
     TosaSpecification,
@@ -115,6 +119,32 @@ def __init__(self, tosa_spec: TosaSpecification) -> None:
         self.tosa_spec = tosa_spec
         super().__init__()
 
+    def validate_constraints_mandatory(self):
+        """
+        Validates that necessary passes have run before transforming to backend.
+
+        Note that this differs from the original validate_constraints function, which
+        only checks the order of passes.
+        """
+        passes_to_run = defaultdict(list)
+
+        for current_pass in self.passes:
+            current_pass_name = ArmPass.get_name(current_pass)
+            for required_pass_name in ArmPass.get_required_passes(current_pass):
+                passes_to_run[required_pass_name].append(current_pass_name)
+
+            passes_to_run.pop(current_pass_name, None)
+
+        if len(passes_to_run) > 0:
+            error_msg = "The following constraints for passes are not met:\n"
+            for required_pass, requiring_passes in passes_to_run.items():
+                for requiring_pass in requiring_passes:
+                    error_msg += (
+                        f"  - {required_pass} must run after {requiring_pass}\n"
+                    )
+
+            raise RuntimeError(error_msg)
+
     def _transform(self, graph_module: GraphModule):
         with TosaLoweringContext(self.tosa_spec):
             return self(graph_module).graph_module
@@ -125,7 +155,6 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(RemoveGetItemPass())
         self.add_pass(ConvertSplitToSlicePass())
         self.add_pass(ConvertMmToBmmPass())
-        self.add_pass(DecomposeLinearVectorNormPass())
         self.add_pass(
             DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec)
         )
@@ -175,6 +204,7 @@ def _tosa_INT_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(RemoveNoopPass())
         self.add_pass(InsertRescalePass())
 
+        self.validate_constraints_mandatory()
         return self._transform(exported_program.graph_module)
 
     def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
@@ -258,6 +288,7 @@ def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(RemoveNoopPass())
         self.add_pass(InsertRescalePass())
 
+        self.validate_constraints_mandatory()
         return self._transform(exported_program.graph_module)
 
     def transform_to_backend_pipeline(self, exported_program: ExportedProgram):

backends/arm/_passes/broadcast_args_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.
 
+from typing import Set, Type
+
 from executorch.backends.arm._passes import ArmPass
 
 from executorch.backends.arm._passes.arm_pass_utils import (
@@ -12,7 +14,7 @@
 
 from executorch.exir.dialects._ops import ops as exir_ops
 
-from executorch.exir.pass_base import PassResult
+from executorch.exir.pass_base import ExportPass, PassResult
 from torch.fx import GraphModule, Node
 
 
@@ -22,6 +24,8 @@ class BroadcastArgsPass(ArmPass):
     This is done when more than one arg needs broadcasting.
     """
 
+    _passes_required_after: Set[Type[ExportPass]] = set()
+
     targeted_ops = {
         exir_ops.edge.aten.add.Tensor,
         exir_ops.edge.aten.sub.Tensor,