Skip to content

Fix tensor name/type disagreement on multi-output nodes #22319

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
psiddh wants to merge 1 commit into
base: main
Choose a base branch
from
+10 −5
Open

Fix tensor name/type disagreement on multi-output nodes #22319

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: 10 additions & 5 deletions backends/qualcomm/builders/node_visitor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -447,6 +447,7 @@ def get_tensor_name(
self,
node: torch.fx.Node,
wrapper_idx: int = 0,
output_index: Optional[int] = None,
):
tensor_name = f"{node.name}@{wrapper_idx}"
# The `input_{id}` is utilized for sorting at runtime. Due to multiple passes in qnn_preprocess,
Expand All @@ -469,7 +470,7 @@ def get_tensor_name(
self.edge_program.graph_signature.buffers_to_mutate.keys()
).index(node.name)
tensor_name = f"output_mutbuf_{position_index}_{tensor_name}"
elif is_graph_output(node, wrapper_idx):
elif is_graph_output(node, output_index):
tensor_name = f"output_{tensor_name}"

# Only add qcom_tensor_name when enable tensor dump.
Expand Down Expand Up @@ -541,14 +542,18 @@ def define_tensor(
if cached := nodes_to_wrappers[node_name].get(wrapper_idx, None):
return cached

tensor_name = self.get_tensor_name(tensor_source_node, wrapper_idx)
dims = torch.Size([1]) if len(tensor.size()) == 0 else tensor.size()
dynamic_dims, nominal_dims = self.get_dynamic_dimension(dims)
# wrapper_idx indexes a node's own outputs only when the tensor being
# defined belongs to that node. Builders also use it to key scratch
# tensors built from some other node (op_scatter_elements), where it
# carries no output meaning.
# carries no output meaning. The naming and the type path must agree on
# this, otherwise a tensor gets an `output_` name without the matching
# APP_READ type (or vice versa) and the QNN graph build fails.
output_index = wrapper_idx if tensor_source_node is target_build_node else None
tensor_name = self.get_tensor_name(
tensor_source_node, wrapper_idx, output_index
)
dims = torch.Size([1]) if len(tensor.size()) == 0 else tensor.size()
dynamic_dims, nominal_dims = self.get_dynamic_dimension(dims)
tensor_type = self.get_tensor_type(
tensor_source_node, tensor_type, output_index
)
Expand Down
Loading