Skip to content

Prevent _safe_softmax decomposition in traceand rewire replaceSafeSoftmaxWithSoftmax #19619

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
meta-codesync merged 1 commit into from
May 16, 2026
+4 −1
Merged

Prevent _safe_softmax decomposition in traceand rewire replaceSafeSoftmaxWithSoftmax #19619

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions backends/cadence/aot/compiler_funcs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@ def trace(
model.eval()

decomp_table = torch.export.default_decompositions()
ops_to_keep = [*(ops_to_keep or []), torch.ops.aten._safe_softmax.default]
# pyre-fixme[6]: For 1st argument expected `Dict[typing.Callable[..., typing.Any
remove_decompositions(decomp_table, ops_to_keep)
program = torch.export.export(model, inputs, strict=strict).run_decompositions(
Expand Down
4 changes: 3 additions & 1 deletion backends/cadence/aot/passes.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@
from executorch.backends.cadence.aot.replace_ops import (
CadenceReplaceOpsInGraph,
ReplaceMulTensorWithMulAndFullOpsPass,
ReplaceSafeSoftmaxWithSoftmax,
)
from executorch.backends.cadence.aot.simplify_ops import CadenceSimplifyOpsInGraph
from executorch.backends.cadence.aot.type_dispatch import CompileTimeTypeDispatchPass
Expand Down Expand Up @@ -131,7 +132,8 @@ def apply_torch_ops_passes(expo_program: ExportedProgram) -> ExportedProgram:
"""

aten_passes: List[Callable[[torch.fx.GraphModule], Optional[PassResult]]] = [
ReplaceMulTensorWithMulAndFullOpsPass()
ReplaceSafeSoftmaxWithSoftmax(),
ReplaceMulTensorWithMulAndFullOpsPass(),
]
# TODO(T230417247): Use PassResult which is currently ignored.
PassManager(aten_passes)(expo_program.graph_module)
Expand Down
Loading