Skip to content

Qualcomm AI Engine Direct - Move QCOM_AXIS_ORDER pop into LayoutTransform pass #21932

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
paul1106 wants to merge 2 commits into
base: main
Choose a base branch
from
+9 −15
Open

Qualcomm AI Engine Direct - Move QCOM_AXIS_ORDER pop into LayoutTransform pass #21932

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b587362
Qualcomm AI Engine Direct - Move QCOM_AXIS_ORDER pop into LayoutTrans…
paul1106 Aug 4, 2026
f701d62
Qualcomm AI Engine Direct - Remove dead QCOM_AXIS_ORDER strip in pass…
paul1106 Aug 19, 2026
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
8 changes: 8 additions & 0 deletions backends/qualcomm/_passes/layout_transform.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -351,6 +351,14 @@ def call(self, graph_module: torch.fx.GraphModule):
self.traverse(node, graph_module)
self.insert_permute, self.transformed_tag = True, QCOM_AXIS_ORDER

for node in graph.nodes:
if hasattr(node, "meta"):
# Pop QCOM_AXIS_ORDER written by the to-edge LayoutTransform pass.
# Without this, the main for-loop below would see is_transformed_node=True
# for every sensitive node (deepcopy carries the tag from to-edge) and
# skip them entirely, so no permute nodes would ever be inserted.
node.meta.pop(QCOM_AXIS_ORDER, "")

for node in sensitive_nodes:
if not self.is_transformed_node(node):
self.mark_as_transformed(node)
Expand Down
9 changes: 1 addition & 8 deletions backends/qualcomm/qnn_preprocess.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,10 +22,7 @@
from executorch.backends.qualcomm.serialization.qc_schema_serialize import (
flatbuffer_to_option,
)
from executorch.backends.qualcomm.utils.constants import (
QCOM_AXIS_ORDER,
QCOM_TENSOR_NAME,
)
from executorch.backends.qualcomm.utils.constants import QCOM_TENSOR_NAME
from executorch.backends.qualcomm.utils.qnn_manager_lifecycle import (
get_current_qnn_manager,
)
Expand Down Expand Up @@ -56,10 +53,6 @@ def _build_op_wrappers(
use_mha2sha: bool,
backend_type: QnnExecuTorchBackendType,
):
for node in edge_program.graph_module.graph.nodes:
if hasattr(node, "meta"):
# pop certain keys in meta for not affecting the passes in compilation
node.meta.pop(QCOM_AXIS_ORDER, "")
# QNN Delegate Specific Passes
graph_module = get_qnn_pass_manager_cls(
backend_type
Expand Down
7 changes: 0 additions & 7 deletions backends/qualcomm/tests/rework/passes/passes_helper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -287,13 +287,6 @@ def lower_preprocess_gm(
quantizer=quantizer,
)
gm = edge_ep.graph_module
# Mirror qnn_preprocess.py: strip QCOM_AXIS_ORDER from all nodes
from executorch.backends.qualcomm.utils.constants import QCOM_AXIS_ORDER

for node in gm.graph.nodes:
if hasattr(node, "meta"):
node.meta.pop(QCOM_AXIS_ORDER, "")

pm_cls = get_qnn_pass_manager_cls(backend_type)
pass_classes = PassPipeline._slice_to_target(
pm_cls.get_preprocess_passes(use_mha2sha=use_mha2sha),
Expand Down
Loading