Skip to content

[Accuracy diff No.167] Fix accuracy (output type) diff for paddle.cumsum API #74625

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

Merged
merged 4 commits into from
Aug 20, 2025
Merged

[Accuracy diff No.167] Fix accuracy (output type) diff for paddle.cumsum API #74625

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
162f341
fix(math.py, unary.cc): fix output type diff for cumsum kernel
hushenwei2000 Aug 14, 2025
b7ad9c5
fix(math.py): fix output type diff for cumsum kernel
hushenwei2000 Aug 14, 2025
a0165e6
fix(math.py): fix `cumsum` documentation
hushenwei2000 Aug 15, 2025
013def0
fix(cum/cum_grad.cc/cu, test_cumsum_op.py): fix output type diff for …
hushenwei2000 Aug 18, 2025
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
10 changes: 8 additions & 2 deletions paddle/phi/infermeta/unary.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -635,9 +635,15 @@ void CumInferMeta(const MetaTensor& x,
bool reverse,
MetaTensor* out) {
auto x_dims = x.dims();
auto x_dtype = x.dtype();
auto out_dtype =
(x_dtype == phi::DataType::UINT8 || x_dtype == phi::DataType::INT8 ||
x_dtype == phi::DataType::INT16 || x_dtype == phi::DataType::INT32)
? phi::DataType::INT64
: x_dtype;
if (flatten) {
out->set_dims(common::make_ddim({common::product(x_dims)}));
out->set_dtype(x.dtype());
out->set_dtype(out_dtype);
} else {
if (x_dims.size() > 0) {
PADDLE_ENFORCE_GE(
Expand Down Expand Up @@ -667,7 +673,7 @@ void CumInferMeta(const MetaTensor& x,
axis));
}
out->set_dims(x_dims);
out->set_dtype(x.dtype());
out->set_dtype(out_dtype);
}

out->share_lod(x);
Expand Down
9 changes: 8 additions & 1 deletion python/paddle/tensor/math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4308,7 +4308,7 @@ def cumsum(
Args:
x (Tensor): The input tensor needed to be cumsumed.
axis (int, optional): The dimension to accumulate along. -1 means the last dimension. The default (None) is to compute the cumsum over the flattened array.
dtype (str|paddle.dtype|np.dtype|None, optional): The data type of the output tensor, can be bfloat16, float16, float32, float64, int32, int64, complex64, complex128. If specified, the input tensor is casted to dtype before the operation is performed. This is useful for preventing data type overflows. The default value is None.
dtype (str|paddle.dtype|np.dtype|None, optional): The data type of the output tensor, can be bfloat16, float16, float32, float64, int32, int64, complex64, complex128. By default, it is int64 if the input x is int8/int16/int32; otherwise, it is None. If it is not None, the input tensor is casted to dtype before the operation is performed. This is useful for preventing data type overflows.
name (str|None, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

Returns:
Expand Down Expand Up @@ -4350,6 +4350,13 @@ def cumsum(
flatten = False
if dtype is not None and x.dtype != convert_np_dtype_to_dtype_(dtype):
x = cast(x, dtype)
elif isinstance(x, paddle.Tensor) and x.dtype in [
paddle.uint8,
paddle.int8,
paddle.int16,
paddle.int32,
]:
x = cast(x, "int64")

if in_dynamic_or_pir_mode():
if axis is None:
Expand Down
Loading