Skip to content

[0-size Tensor No.133、318、277、136、320、239、325、279] Add 0-size Tensor support for matmul #73094

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 1 commit into from
Jun 13, 2025
Merged

[0-size Tensor No.133、318、277、136、320、239、325、279] Add 0-size Tensor support for matmul #73094

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
7 changes: 6 additions & 1 deletion paddle/phi/infermeta/binary.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2992,7 +2992,12 @@ void MatmulInferMeta(const MetaTensor& x,
} else {
new_dims.reserve(ndims_x);
for (size_t i = 0; i < ndims_x - 2; ++i) {
new_dims.push_back(std::max(dims_x[i], dims_y[i]));
// If one of them is 0, choose 0.
if (dims_x[i] == 0 || dims_y[i] == 0) {
new_dims.push_back(0);
Copy link
Contributor

@DanielSun11 DanielSun11 Jun 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个地方torch的实验结果如下：

case1
Tensor x shape: torch.Size([1, 1, 4, 5])
Tensor y shape: torch.Size([1, 0, 5, 6])

Result shape: torch.Size([1, 0, 4, 6])

case2
Tensor x shape: torch.Size([1, 3, 4, 5])
Tensor y shape: torch.Size([1, 0, 5, 6])
Traceback (most recent call last):
  File "/home/code/test/test_matmul.py", line 20, in <module>
    result = torch.matmul(tensor1, tensor2)
RuntimeError: The size of tensor a (3) must match the size of tensor b (0) at non-singleton dimension 1

针对于case2 的情况paddle是报错还是产生一个Result shape: torch.Size([1, 0, 4, 6])的矩阵？从代码中看可能是Result shape: torch.Size([1, 0, 4, 6])?
从目前PR尚未合入的版本测试发现，paddle的一个不合常理的case如下：

Tensor 1 shape: [1, 3, 2, 3]
Tensor 2 shape: [1, 0, 3, 2]

paddle Result shape: [1, 3, 2, 0, 3, 2]

Copy link
Contributor

@DanielSun11 DanielSun11 Jun 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

请在单测中也添加类似的case，以验证broadcast是否广播正确

Tensor 1 shape: [1, 3, 2, 3]
Tensor 2 shape: [1, 0, 3, 2]

Copy link
Contributor Author

@co63oc co63oc Jun 5, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

paddle Result shape: [1, 3, 2, 0, 3, 2] paddle中是这样，这部分代码已经删除修改

case2 paddle是不报错，返回[1, 0, 4, 6]，修改的代码就是 if (dims_x[i] == 0 || dims_y[i] == 0) {

Tensor 1 shape: [1, 3, 2, 3]
Tensor 2 shape: [1, 0, 3, 2] 这个现在是会返回[1, 0, 2, 2]不报错，那再增加单测

} else {
new_dims.push_back(std::max(dims_x[i], dims_y[i]));
}
}
}
if (!x_broadcasted) {
Expand Down
14 changes: 9 additions & 5 deletions paddle/phi/kernels/impl/matmul_grad_kernel_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,11 +230,15 @@ void MatmulGradKernel(const Context& dev_ctx,
DenseTensor* dx,
DenseTensor* dy) {
if (x.numel() == 0) {
if (dy != nullptr) {
dev_ctx.template Alloc<T>(dx);
phi::FullKernel<T>(
dev_ctx, common::vectorize(y.dims()), 0.0, y.dtype(), dy);
}
dev_ctx.template Alloc<T>(dx);
phi::Full<T, Context>(
dev_ctx, phi::IntArray(common::vectorize(y.dims())), 0, dy);
return;
}
if (y.numel() == 0) {
dev_ctx.template Alloc<T>(dy);
phi::Full<T, Context>(
dev_ctx, phi::IntArray(common::vectorize(x.dims())), 0, dx);
return;
}
// get dims
Expand Down
21 changes: 4 additions & 17 deletions paddle/phi/kernels/impl/matmul_kernel_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,7 @@ limitations under the License. */
#if defined(PADDLE_WITH_CUDA) && CUDA_VERSION >= 11060
#include "paddle/phi/kernels/autotune/auto_tune_base.h"
#endif
#include "paddle/phi/kernels/full_kernel.h"

COMMON_DECLARE_bool(cuda_core_int8_gemm);

Expand Down Expand Up @@ -2007,23 +2008,9 @@ void MatmulKernel(const Context& dev_ctx,
bool transpose_y,
DenseTensor* out) {
if (x.numel() == 0 || y.numel() == 0) {
auto x_dims = x.dims();
auto y_dims = y.dims();
if (transpose_x) {
std::swap(x_dims[x_dims.size() - 1], x_dims[x_dims.size() - 2]);
}
if (transpose_y) {
std::swap(y_dims[y_dims.size() - 1], y_dims[y_dims.size() - 2]);
}
std::vector<std::int64_t> out_dims(x_dims.size() - 1 + y_dims.size() - 1);
for (int64_t i = 0; i < x_dims.size() - 1; ++i) {
out_dims[i] = x_dims[i];
}
for (int64_t i = 1; i < y_dims.size(); ++i) {
out_dims[x_dims.size() - 1 + i - 1] = y_dims[i];
}
out->Resize(phi::make_ddim(out_dims));
dev_ctx.template Alloc<T>(out);
// input shape [1, 1, 5, 0], [1, 1, 0, 5], result shape is [1, 1, 5, 5]
phi::Full<T, Context>(
dev_ctx, phi::IntArray(common::vectorize(out->dims())), 0, out);
return;
}
PADDLE_ENFORCE_GE(
Expand Down
12 changes: 10 additions & 2 deletions python/paddle/tensor/math.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2942,8 +2942,16 @@ def outer(x: Tensor, y: Tensor, name: str | None = None) -> Tensor:


"""
nx = x.reshape((-1, 1))
ny = y.reshape((1, -1))
xshape = x.shape
yshape = y.shape
if math.prod(xshape) == 0: # If the size is 0
nx = x.reshape((0, 0))
else:
nx = x.reshape((-1, 1))
if math.prod(yshape) == 0: # If the size is 0
ny = y.reshape((0, 0))
else:
ny = y.reshape((1, -1))

if in_dynamic_mode():
return _C_ops.matmul(nx, ny, False, False)
Expand Down
35 changes: 35 additions & 0 deletions test/legacy_test/test_matmul_v2_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -943,6 +943,41 @@ def func_dygraph_matmul(self):
paddle.enable_static()


class TestMatMulOp_ZeroSize(OpTest):
def setUp(self):
self.op_type = "matmul_v2"
self.python_api = paddle.matmul
self.init_input_output()

self.inputs = {
'X': OpTest.np_dtype_to_base_dtype(self.x),
'Y': OpTest.np_dtype_to_base_dtype(self.y),
}
self.out = np.matmul(self.x, self.y)
self.attrs = {'axis': -1, 'use_mkldnn': False}
self.outputs = {'Out': self.out}

def init_input_output(self):
self.x = np.random.random((1, 1, 2, 3))
self.y = np.random.random((1, 0, 3, 2))

def test_check_output(self):
self.check_output(check_pir=True)

def test_check_grad(self):
self.check_grad(
['X', 'Y'],
'Out',
check_pir=True,
)


class TestMatMulOp_ZeroSize2(TestMatMulOp_ZeroSize):
def init_input_output(self):
self.x = np.random.random((0, 3, 2, 3))
self.y = np.random.random((1, 3, 3, 2))


if __name__ == "__main__":
paddle.enable_static()
unittest.main()
18 changes: 18 additions & 0 deletions test/legacy_test/test_outer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,5 +172,23 @@ def test_errors_dynamic(self):
self.assertRaises(Exception, paddle.outer, x_data, y_data)


class TestMultiplyApi_ZeroSize(unittest.TestCase):
def test_multiply_dynamic(self):
x_data = np.random.rand(5, 10, 0).astype(np.float64)
y_data = np.random.rand(0, 10).astype(np.float64)
paddle.disable_static()
x = paddle.to_tensor(x_data)
y = paddle.to_tensor(y_data)
x.stop_gradient = False
y.stop_gradient = False
res = paddle.outer(x, y)
np.testing.assert_allclose(
res.numpy(), np.outer(x_data, y_data), rtol=1e-05
)
loss = paddle.sum(res)
loss.backward()
np.testing.assert_allclose(x.grad.shape, x.shape)


if __name__ == '__main__':
unittest.main()
Loading