[API] Support 0-size tensor for concat

paddle/phi/kernels/impl/concat_grad_kernel_impl.h

@@ -45,9 +45,8 @@ void ConcatGradKernel(const Context& dev_ctx,
  // get output tensor that the name is not kEmptyVarName
  std::vector<DenseTensor*> outputs;
  for (size_t j = 0; j < outs.size(); ++j) {
- if (outs[j] && outs[j]->numel() != 0UL) {
+ if (outs[j]) {
  dev_ctx.template Alloc<T>(outs[j]);
-
  outputs.push_back(outs[j]);
  } else {
  outputs.push_back(nullptr);

paddle/phi/kernels/xpu/concat_grad_kernel.cc

@@ -47,11 +47,9 @@ void ConcatGradKernel(const Context& dev_ctx,
  // get output tensor that the name is not kEmptyVarName
  std::vector<XPUType*> ptrs(outs.size());
  for (size_t j = 0; j < outs.size(); ++j) {
- if (outs[j] && outs[j]->numel() != 0UL) {
+ if (outs[j]) {
  dev_ctx.template Alloc<T>(outs[j]);
  ptrs[j] = reinterpret_cast<XPUType*>(outs[j]->data<T>());
- } else {
- ptrs[j] = nullptr;
  }
  }
  PADDLE_ENFORCE_GE(axis,

python/paddle/tensor/manipulation.py

@@ -1458,8 +1458,6 @@ def is_in_amp_mode():
  else:
  input = [input]
 
- if not isinstance(input, paddle.pir.Value):
- input = [t for t in input if t.shape.count(0) == 0]
  return _C_ops.concat(input, axis)
  else:
  check_type(input, 'input', (list, tuple, Variable), 'concat')

test/legacy_test/test_concat_op.py

@@ -286,6 +286,76 @@ def init_test_data(self):
  self.axis = 1
 
 
+class TestConcatOp0Size(TestConcatOp):
+ def setUp(self):
+ self.op_type = "concat"
+ self.python_api = paddle.concat
+ self.public_python_api = paddle.concat
+ self.prim_op_type = "prim"
+ self.if_enable_cinn()
+ self.dtype = self.get_dtype()
+ self.init_test_data()
+ self.inputs = {'X': [('x0', self.x0), ('x1', self.x1), ('x2', self.x2)]}
+ self.attrs = {'axis': self.axis}
+ if self.axis < 0:
+ self.actual_axis = self.axis + len(self.x0.shape)
+ self.actual_axis = max(0, self.actual_axis)
+ else:
+ self.actual_axis = self.axis
+
+ self.outputs = {
+ 'Out': np.concatenate(
+ (self.x0, self.x1, self.x2), axis=self.actual_axis
+ )
+ }
+
+ def if_enable_cinn(self):
+ pass
+
+ def get_dtype(self):
+ return "float64"
+
+ def test_check_output(self):
+ self.check_output(check_pir=True)
+
+ def test_check_grad(self):
+ self.check_grad(
+ ['x0'],
+ 'Out',
+ check_prim=True,
+ check_pir=True,
+ check_prim_pir=True,
+ )
+ self.check_grad(
+ ['x1'],
+ 'Out',
+ check_prim=True,
+ check_pir=True,
+ check_prim_pir=True,
+ )
+ self.check_grad(
+ ['x2'],
+ 'Out',
+ check_prim=True,
+ check_pir=True,
+ check_prim_pir=True,
+ )
+
+ def init_test_data(self):
+ if self.dtype == np.uint16:
+ x0 = np.random.random((5, 1, 4, 5)).astype(np.float32)
+ self.x0 = convert_float_to_uint16(x0)
+ x1 = np.random.random((5, 0, 4, 5)).astype(np.float32)
+ self.x1 = convert_float_to_uint16(x1)
+ x2 = np.random.random((5, 3, 4, 5)).astype(np.float32)
+ self.x2 = convert_float_to_uint16(x2)
+ else:
+ self.x0 = np.random.random((5, 1, 4, 5)).astype(self.dtype)
+ self.x1 = np.random.random((5, 0, 4, 5)).astype(self.dtype)
+ self.x2 = np.random.random((5, 3, 4, 5)).astype(self.dtype)
+ self.axis = 1
+
+
 def create_test_AxisTensor(parent):
  class TestConcatAxisTensor(parent):
  def setUp(self):