Skip to content

Fast-copy cold nodes during ExportPass retracing (#18497) #18497

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

Open
apullin wants to merge 1 commit into
base: main
Choose a base branch
from
+642 −8
Open

Fast-copy cold nodes during ExportPass retracing (#18497) #18497

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
15 changes: 12 additions & 3 deletions backends/arm/_passes/arm_pass.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,15 @@ def _is_quantized_meta(self, meta: NodeMetadata | dict[str, Any]) -> bool:
output_qparams = meta_dict.get("output_qparams", {})
return bool(input_qparams) and bool(output_qparams)

def should_fast_copy_node(self, target: torch.fx.node.Target) -> bool:
ops_without_quantized_fake_kernel = {
exir_ops.edge.aten.bmm.default,
exir_ops.edge.aten.leaky_relu.default,
}
if any(target is op for op in ops_without_quantized_fake_kernel):
return False
return super().should_fast_copy_node(target)

@property
@abstractmethod
def _passes_required_after(self) -> Set[Type[ExportPass]]:
Expand Down Expand Up @@ -142,11 +151,11 @@ def call_submodule(
self, graph_module: GraphModule, inputs: tuple[Any, ...]
) -> PassResult:
self.submodule_depth += 1
if self.submodule_depth == 1:
if self.submodule_depth == 1 or self.should_run_pass(graph_module):
result = super().call_submodule(graph_module, inputs)
else:
# When we trace a submodule, we don't want to apply the calling pass.
# Temporarily replace call_operator to avoid this.
# Nested submodules that do not need this pass still need normal replay.
# Temporarily replace call_operator to avoid applying subclass rewrites.
_call_operator_fn = self.call_operator
self.call_operator = super().call_operator # type: ignore
result = super().call_submodule(graph_module, inputs)
Expand Down
Loading
Loading