Skip to content

Rewriter: fix cast_constant_of_shape for values outside the target range #3153

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
Lim Jooyoung (mmporong) wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Rewriter: fix cast_constant_of_shape for values outside the target range #3153

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
11 changes: 8 additions & 3 deletions onnxscript/rewriter/rules/common/_cast_constant_of_shape.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@

import logging

import numpy as np
import onnx_ir as ir

from onnxscript.rewriter._rewrite_rule import RewriteRule, RewriteRuleSet
Expand All @@ -17,9 +18,13 @@ def cast_constant_of_shape(op, shape, scalar, dtype):


def fused_cast_constant_of_shape(op, shape: ir.Value, scalar: ir.Attr, dtype: ir.Attr, **_):
# Cast scalar (a TensorProto attribute) to the specified dtype
scalar_value = scalar.value.numpy().item()
cast_value = ir.tensor([scalar_value], dtype=ir.DataType(dtype.as_int()))
# Cast scalar (a TensorProto attribute) to the specified dtype. Convert the array, not a
# Python scalar, so that values outside the target range wrap like the Cast op instead of
# raising OverflowError.
to_dtype = ir.DataType(dtype.as_int())
with np.errstate(invalid="ignore"):
cast_array = scalar.value.numpy().astype(to_dtype.numpy()).reshape(1)
cast_value = ir.tensor(cast_array, dtype=to_dtype)
return op.ConstantOfShape(shape, value=cast_value)


Expand Down
30 changes: 30 additions & 0 deletions onnxscript/rewriter/rules/common/_cast_constant_of_shape_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,9 +2,11 @@
# Licensed under the MIT License.
import unittest

import numpy as np
import onnx.checker
import onnx.parser
import onnx_ir as ir
import parameterized

from onnxscript.rewriter.rules.common import _cast_constant_of_shape

Expand Down Expand Up @@ -49,6 +51,34 @@ def test_cast_after_constant_of_shape_without_value_is_fused(self):
output_model_proto = ir.serde.serialize_model(model)
onnx.checker.check_model(output_model_proto, True)

@parameterized.parameterized.expand(
[
("int_above_uint8_range", "int32", "{300}", 2, np.array([300], np.int32)),
("negative_int_to_uint8", "int32", "{-1}", 2, np.array([-1], np.int32)),
("int64_to_int8", "int64", "{200}", 3, np.array([200], np.int64)),
]
)
def test_fused_value_matches_cast_of_out_of_range_constant(
self, _, value_type, value, to, value_array
):
model_proto = onnx.parser.parse_model(
f"""
<ir_version: 7, opset_import: [ "" : 17]>
agraph (int64[2] input_x) => (output)
{{
constant = ConstantOfShape <value: tensor = {value_type}[1] {value}>(input_x)
output = Cast <to = {to}> (constant)
}}
"""
)
model = ir.serde.deserialize_model(model_proto)
count = _cast_constant_of_shape.rules.apply_to_model(model)
self.assertEqual(count, 1)
fused_value = model.graph[0].attributes["value"].value
self.assertEqual(fused_value.dtype, to)
expected = value_array.astype(ir.DataType(to).numpy())
np.testing.assert_array_equal(fused_value.numpy(), expected)


if __name__ == "__main__":
unittest.main()