Arm backend: Add SDPA decomposition to annotation pipeline (pytorch#10657)

Adds SDPA decomposition and since SDPA is decomposed to safe_softmax
instead of regular softmax, softmax decomposition is updated to handle
safe_softmax.

Co-authored-by: Måns Nilsson <[email protected]>
Co-authored-by: Per Åstrand <[email protected]>

Author: mansnils

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/test/models/test_conformer.py

@@ -83,7 +83,6 @@ def test_conformer_tosa_BI(self):
             )
         )
 
-    @unittest.expectedFailure  # TODO(MLETORCH-635)
     def test_conformer_u55_BI(self):
         tester = (
             ArmTester(
@@ -97,13 +96,20 @@ def test_conformer_u55_BI(self):
             .to_executorch()
             .serialize()
         )
+
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1.0,
-                rtol=1.0,
-                atol=5.0,
-                inputs=get_test_inputs(self.dim, self.lengths, self.num_examples),
-            )
+            try:
+                tester.run_method_and_compare_outputs(
+                    qtol=1.0,
+                    rtol=1.0,
+                    atol=5.0,
+                    inputs=get_test_inputs(self.dim, self.lengths, self.num_examples),
+                )
+                self.fail(
+                    "TODO(MLETORCH-635): Expected failure under FVP option, but test passed."
+                )
+            except Exception:
+                pass
 
     @unittest.expectedFailure  # TODO(MLETORCH-635)
     def test_conformer_u85_BI(self):

backends/arm/test/ops/test_sdpa.py

@@ -0,0 +1,45 @@
+# 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 Tuple
+
+import torch
+
+from executorch.backends.arm.test.tester.test_pipeline import (
+    TosaPipelineBI,
+    TosaPipelineMI,
+)
+
+
+class SDPA(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, query, key, value):
+        return torch.nn.functional.scaled_dot_product_attention(
+            query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+
+input_t = Tuple[torch.Tensor, torch.Tensor, torch.Tensor]
+
+
+def test_sdpa_MI():
+    test_input = tuple(torch.randn(1, 3, 197, 64) for x in range(3))
+    pipeline = TosaPipelineMI[input_t](SDPA(), test_input, [], [])
+    pipeline.pop_stage("check_count.exir")
+    pipeline.run()
+
+
+def test_sdpa_BI():
+    test_input = tuple(torch.randn(1, 3, 197, 64) for x in range(3))
+    pipeline = TosaPipelineBI[input_t](SDPA(), test_input, [], [])
+    pipeline.pop_stage("check.quant_nodes")
+    pipeline.pop_stage("check_count.exir")
+    pipeline.pop_stage(
+        "run_method_and_compare_outputs"
+    )  # TODO: reference is not quantized
+    pipeline.run()