Skip to content
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

Jump to bottom

Fix ONNX export for MobileBERT #1539

Merged
merged 3 commits into from
Apr 26, 2023
Merged

Fix ONNX export for MobileBERT #1539

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
be8a6d5
Account for the possibility of the quantized embedddings to be in int…
anmarques Apr 26, 2023
1bfcada
Set the padding value to match to the zero point accordingly.
anmarques Apr 26, 2023
c4326b1
Style and quality fixes
anmarques Apr 26, 2023
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
9 changes: 6 additions & 3 deletions src/sparseml/pytorch/sparsification/quantization/quantize_qat_export.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1761,7 +1761,7 @@ def _propagate_mobilebert_embedding_quantization(model: ModelProto):
continue

embedding_array = numpy_helper.to_array(embedding_initializer)
if embedding_array.dtype != numpy.uint8:
if embedding_array.dtype not in [numpy.uint8, numpy.int8]:
continue

dequant_node = graph.get_node_single_child(gather_node)
Expand Down Expand Up @@ -1805,12 +1805,15 @@ def _propagate_mobilebert_embedding_quantization(model: ModelProto):
# switch position of dequantize node
for branch_node in graph.get_node_children(dequant_node):
if branch_node.op_type == "Slice":
zero_point = graph.get_init_by_name(dequant_node.input[2])
zero_point_array = numpy_helper.to_array(zero_point)
branch_node.input[0] = gather_node.output[0]
pad_node = graph.get_node_single_child(branch_node)
pad_value = graph.get_init_by_name(pad_node.input[2])
pad_value_array = numpy_helper.to_array(pad_value)
pad_value_array = pad_value_array + 128
pad_value_array = pad_value_array.astype(numpy.uint8)
pad_value_array = (
pad_value_array.astype(zero_point_array.dtype) + zero_point_array
)
model.graph.initializer.remove(pad_value)
pad_value = numpy_helper.from_array(
pad_value_array, name=pad_value.name
Expand Down