Skip to content

[ET-VK] Do not partition batch norm on a non-4d input #22509

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
msluszniak wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[ET-VK] Do not partition batch norm on a non-4d input #22509

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
17 changes: 17 additions & 0 deletions backends/vulkan/op_registry.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1735,13 +1735,30 @@ def register_embedding_q4gsw():
# =============================================================================


def _check_batch_norm_is_4d(node: torch.fx.Node) -> bool:
"""Only support batch norm on a 4d input.

add_native_batch_norm_node() asserts
VK_CHECK_COND(in_sizes.size() == 4, "BatchNorm only support 4d tensor") on
both the input and the output, so partitioning a batch norm whose input is
not 4d yields a .pte that lowers cleanly and then aborts at execute time.
Any conv1d model reaches here with rank-3 activations.
"""
input_node = node.args[0]
if not isinstance(input_node, torch.fx.Node):
return False
val = input_node.meta.get("val")
return val is not None and val.dim() == 4


@update_features(exir_ops.edge.aten._native_batch_norm_legit_no_training.default)
def register_native_batch_norm_legit_no_training():
return OpFeatures(
inputs_storage=utils.CHANNELS_PACKED_TEXTURE,
inputs_dtypes=utils.FP_T,
supports_prepacking=True,
supports_resize=True,
are_node_inputs_supported_fn=_check_batch_norm_is_4d,
)


Expand Down
Loading