Merge branch 'main' into gemma4-mlx-install-path

Author: metascroy

backends/arm/_passes/arm_pass_manager.py

@@ -150,7 +150,6 @@
 from executorch.backends.arm.common.arm_compile_spec import ArmCompileSpec
 from executorch.backends.arm.common.pipeline_config import (
     ArmPassPipelineConfig,
-    FuseDuplicateUsersConfig,
     SoftmaxDecompositionConfig,
 )
 from executorch.backends.arm.tosa.specification import (
@@ -238,9 +237,6 @@ def configure_skip_passes(
             case SoftmaxDecompositionConfig.STABLE:
                 skip_set.add(DecomposeMaskedFillPass)
 
-        if config.fuse_duplicate_users is FuseDuplicateUsersConfig.DISABLED:
-            skip_set.add(FuseDuplicateUsersPass)
-
         self._skip_pass_types = tuple(skip_set)
         skip_names = [skipped_pass.__name__ for skipped_pass in self._skip_pass_types]
         logger.debug(f"Passes in skip list: {skip_names}")
@@ -403,9 +399,6 @@ def _tosa_pipeline(
                 ConvertToClampPass(),
                 DecomposeTOSAUnsupportedClampPass(),
                 DecomposeGroupNormPass(),
-                DecomposeGruPass(),
-                DecomposeLstmPass(),
-                DecomposeRnnPass(),
                 DecomposeLayerNormPass(),
                 DecomposeVarPass(),
                 DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec),

backends/arm/_passes/decompose_gru_pass.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 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.
@@ -13,7 +14,6 @@
     create_node,
     get_getitem_users,
 )
-from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.exir.pass_base import ExportPass, PassResult
 
 
@@ -34,7 +34,7 @@ class DecomposeGruPass(ArmPass):
 
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = {InsertTableOpsPass}
+    _passes_required_after: Set[Type[ExportPass]] = set()
 
     _TARGET = torch.ops.aten.gru.input
 

backends/arm/_passes/decompose_lstm_pass.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 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.
@@ -13,7 +14,6 @@
     create_node,
     get_getitem_users,
 )
-from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.exir.pass_base import ExportPass, PassResult
 
 
@@ -36,7 +36,7 @@ class DecomposeLstmPass(ArmPass):
 
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = {InsertTableOpsPass}
+    _passes_required_after: Set[Type[ExportPass]] = set()
 
     _TARGET = torch.ops.aten.lstm.input
 

backends/arm/_passes/decompose_rnn_pass.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 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.
@@ -13,7 +14,6 @@
     create_node,
     get_getitem_users,
 )
-from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
 from executorch.exir.pass_base import ExportPass, PassResult
 
 
@@ -30,7 +30,7 @@ class DecomposeRnnPass(ArmPass):
 
     """
 
-    _passes_required_after: Set[Type[ExportPass]] = {InsertTableOpsPass}
+    _passes_required_after: Set[Type[ExportPass]] = set()
 
     _TARGETS = {
         torch.ops.aten.rnn_tanh.input,

backends/arm/_passes/decompose_sum_pass.py

@@ -78,7 +78,7 @@ def call_operator(self, op, args, kwargs, meta):
         for dim in dims:
             input_node = super().call_operator(
                 sum_op,
-                (input_node, dim, True),
+                (input_node, [dim], True),
                 kwargs,
                 meta,
                 updated=True,

backends/arm/common/pipeline_config.py

@@ -14,24 +14,12 @@ class SoftmaxDecompositionConfig(Enum):
     STABLE = auto()  # Stable softmax, no masked fill decomposition
 
 
-class FuseDuplicateUsersConfig(Enum):
-    ENABLED = auto()
-    DISABLED = auto()
-
-
 @dataclass
 class ArmPassPipelineConfig:
     softmax: SoftmaxDecompositionConfig = SoftmaxDecompositionConfig.MASKED
-    fuse_duplicate_users: FuseDuplicateUsersConfig = FuseDuplicateUsersConfig.ENABLED
-
-    def disable_fuse_duplicate_users(self) -> None:
-        self.fuse_duplicate_users = FuseDuplicateUsersConfig.DISABLED
 
     def is_default(self) -> bool:
-        return (
-            self.softmax is SoftmaxDecompositionConfig.MASKED
-            and self.fuse_duplicate_users is FuseDuplicateUsersConfig.ENABLED
-        )
+        return self.softmax is SoftmaxDecompositionConfig.MASKED
 
     def to_dict(self) -> dict[str, str]:
         return {f.name: getattr(self, f.name).name for f in fields(self)}

backends/arm/operators/op_sum.py

@@ -43,7 +43,7 @@ def define_node(
 
         tensor = inputs[0]
         input_shape = list(tensor.shape)
-        dim = int(inputs[1].number % len(input_shape))
+        dim = int(inputs[1].special[0] % len(input_shape))
 
         attr = ts.TosaSerializerAttribute()
         attr.ReduceSumAttribute(dim)

backends/arm/test/misc/test_compile_spec.py

@@ -5,10 +5,7 @@
 
 import warnings
 
-from executorch.backends.arm.common.pipeline_config import (
-    FuseDuplicateUsersConfig,
-    SoftmaxDecompositionConfig,
-)
+from executorch.backends.arm.common.pipeline_config import SoftmaxDecompositionConfig
 from executorch.backends.arm.ethosu import EthosUCompileSpec
 from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
 from executorch.backends.arm.vgf import VgfCompileSpec
@@ -66,11 +63,11 @@ def test_compile_spec_vgf_no_quant():
         EthosUCompileSpec._from_list(spec_list)
 
 
-def test_compile_spec_vgf_defaults_to_enabled_fuse_duplicate_users():
+def test_compile_spec_vgf_uses_default_pipeline_config():
     compile_spec = VgfCompileSpec()
     pipeline_config = compile_spec._get_pass_pipeline_config()
 
-    assert pipeline_config.fuse_duplicate_users == FuseDuplicateUsersConfig.ENABLED
+    assert pipeline_config.is_default()
 
 
 def test_compile_spec_tosa_INT():

backends/arm/test/misc/test_pass_pipeline_config.py

@@ -29,14 +29,13 @@ def test_pipeline_config_override_outside_compile_spec():
     override_compile_spec = TosaCompileSpec(
         TosaSpecification.create_from_string("TOSA-1.00+INT")
     )
-    override_config = ArmPassPipelineConfig()
-    override_config.disable_fuse_duplicate_users()
+    override_config = ArmPassPipelineConfig(softmax=SoftmaxDecompositionConfig.STABLE)
     override_compile_spec.set_pass_pipeline_config(override_config)
     override_manager = ArmPassManager(override_compile_spec)
     skip_passes = override_manager._skip_pass_types
 
-    assert FuseDuplicateUsersPass in skip_passes
-    assert DecomposeSoftmaxPass not in skip_passes
+    assert FuseDuplicateUsersPass not in skip_passes
+    assert DecomposeMaskedFillPass in skip_passes
 
 
 def test_softmax_config_masked_no_target():

backends/arm/test/misc/test_tosa_dialect_avg_pool2d_adaptive.py

@@ -19,6 +19,27 @@
 from torch._subclasses.fake_tensor import FakeTensorMode
 
 
+def test_avg_pool2d_tosa_non_square_kernel_output_shape():
+    with TosaLoweringContext(
+        TosaSpecification.create_from_string("TOSA-1.0+FP")
+    ), FakeTensorMode() as mode:
+        x = mode.from_tensor(torch.randn((1, 20, 20, 8), dtype=torch.float32))
+        input_zp = mode.from_tensor(torch.zeros((1,), dtype=torch.float32))
+        output_zp = mode.from_tensor(torch.zeros((1,), dtype=torch.float32))
+
+        output = exir_ops.backend.tosa.AVG_POOL2D.default(
+            x,
+            input_zp,
+            output_zp,
+            [2, 3],
+            [2, 1],
+            [1, 1, 0, 0],
+            torch.float32,
+        )
+
+        assert tuple(output.shape) == (1, 11, 18, 8)
+
+
 def test_avg_pool2d_adaptive_tosa_INT():
     sample_inputs = [
         (