Revert "Add linalg.vander"

This reverts commit 1ea49c68d0.

Reverted https://github.com/pytorch/pytorch/pull/76303 on behalf of https://github.com/malfet

aten/src/ATen/native/BatchLinearAlgebra.cpp

@@ -4617,33 +4617,4 @@ Tensor linalg_solve_triangular(
   return out;
 }
 
-Tensor linalg_vander(
-    const Tensor& x,
-    c10::optional<int64_t> N) {
-  const auto x_ = x.dim() == 0 ? x.unsqueeze(-1) : x;
-
-  auto shape = x_.sizes().vec();
-  const auto n = N.value_or(shape.back());
-  TORCH_CHECK(n >= 0, "N must be non-negative.");
-
-  // Get the dtype of cumsum
-  auto dtype = x_.scalar_type();
-  auto options = x_.options()
-                   .dtype(at::isIntegralType(dtype, /*includeBool=*/true) ? kLong : dtype);
-
-  // n = 0 or n = 1 case (Empty or one row case)
-  shape.push_back(std::min<int64_t>(n, 1LL));
-  auto ones =  x_.new_ones(shape, options);
-
-  if (n <= 1) {
-    // new_ones does not propagate requires_grad. UGH
-    ones.requires_grad_(x_.requires_grad());
-    return ones;
-  } else {
-    // Append cumprod of the oher 0...n-1 powers
-    shape.back() = n - 1;
-    auto result = at::cumprod(x_.unsqueeze(-1).expand(shape), -1);
-    return at::cat({ones, result}, /*dim=*/ -1);
-  }
-}
 }}  // namespace at::native

aten/src/ATen/native/native_functions.yaml

@@ -7179,9 +7179,6 @@
   dispatch:
     CPU, CUDA: linalg_solve_triangular
 
-- func: linalg_vander(Tensor x, *, int? N=None) -> Tensor
-  python_module: linalg
-
 - func: symeig.e(Tensor self, bool eigenvectors=False, bool upper=True, *, Tensor(a!) e, Tensor(b!) V) -> (Tensor(a!) eigenvalues, Tensor(b!) eigenvectors)
   dispatch:
     CompositeExplicitAutograd: symeig_out

docs/source/linalg.rst

@@ -95,15 +95,6 @@ Tensor Operations
     tensorinv
     tensorsolve
 
-Misc
-----
-
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    vander
-
 Experimental Functions
 ----------------------
 .. autosummary::

test/allowlist_for_publicAPI.json

@@ -2104,7 +2104,6 @@
     "svdvals",
     "tensorinv",
     "tensorsolve",
-    "vander",
     "vector_norm"
   ],
   "torch.multiprocessing": [

torch/linalg/__init__.py

@@ -2527,53 +2527,3 @@ Examples::
     >>> torch.dist(Q.mT @ Q, torch.eye(4))
     tensor(6.2158e-07)
 """)
-
-vander = _add_docstr(_linalg.linalg_vander, r"""
-vander(x, N=None) -> Tensor
-
-Generates a Vandermonde matrix.
-
-Returns the Vandermonde matrix :math:`V`
-
-.. math::
-
-    V = \begin{pmatrix}
-            1 & x_1 & x_1^2 & \dots & x_1^{N-1}\\
-            1 & x_2 & x_2^2 & \dots & x_2^{N-1}\\
-            1 & x_3 & x_3^2 & \dots & x_3^{N-1}\\
-            \vdots & \vdots & \vdots & \ddots &\vdots \\
-            1 & x_n & x_n^2 & \dots & x_n^{N-1}
-        \end{pmatrix}.
-
-If :attr:`N`\ `= None`, then `N = x.size(-1)` so that the output is a square matrix.
-
-Supports inputs of all dtypes.
-Also supports batches of vectors, and if :attr:`x` is a batch of vectors then
-the output has the same batch dimensions.
-
-Differences with `numpy.vander`:
-
-- Unlike `numpy.vander`, this function returns the powers of :attr:`x` in ascending order.
-  To get them in the reverse order call ``linalg.vander(x, N).flip(-1)``.
-
-Args:
-    x (Tensor): tensor of shape `(*, n)` where `*` is zero or more batch dimensions
-                consisting of vectors.
-
-Keyword args:
-    N (int, optional): Number of columns in the output. Default: `x.size(-1)`
-
-Example::
-
-    >>> x = torch.tensor([1, 2, 3, 5])
-    >>> linalg.vander(x)
-    tensor([[  1,   1,   1,   1],
-            [  1,   2,   4,   8],
-            [  1,   3,   9,  27],
-            [  1,   5,  25, 125]])
-    >>> linalg.vander(x, N=3)
-    tensor([[ 1,  1,  1],
-            [ 1,  2,  4],
-            [ 1,  3,  9],
-            [ 1,  5, 25]])
-""")

torch/overrides.py

@@ -1037,7 +1037,6 @@ def get_testing_overrides() -> Dict[Callable, Callable]:
         torch.unsafe_split: lambda tensor, split_size_or_sections, dim=0: -1,
         torch.unsafe_split_with_sizes: lambda tensor, split_size_or_sections, dim=0: -1,
         torch.unsqueeze: lambda input, dim, out=None: -1,
-        torch.linalg.vander: lambda x, N=None: -1,
         torch.var: lambda input, dim=None: -1,
         torch.var_mean: lambda input, dim=None: -1,
         torch.vsplit: lambda input, indices_or_sections: -1,

torch/testing/_internal/common_methods_invocations.py

@@ -369,7 +369,7 @@ NumericsFilter = collections.namedtuple('NumericsFilter', ['condition', 'safe_va
 #   "name".
 #
 #   All the "sample_inputs" functions are invoked within a `torch.no_grad()`
-#   environment for efficiency and correctness. As such remember to set the
+#   environment for efficiency and correctness. As such remember to set the the
 #   "requires_grad" flag on the inputs **after** performing any transformations
 #   on them.
 #
@@ -402,7 +402,7 @@ NumericsFilter = collections.namedtuple('NumericsFilter', ['condition', 'safe_va
 # Sample inputs are designed to be used with many tests, some
 #   that are very time consuming, so they should be a small
 #   set with small tensors. An elaborated set of sample inputs
-#   can be specified using the "reference_inputs_func" attribute.
+#   can be specified using the the "reference_inputs_func" attribute.
 #   The "reference inputs" for an operation are an extended
 #   set of sample inputs that can more exhausively test an
 #   operator. They are used by only a few tests that are careful
@@ -3550,37 +3550,6 @@ def sample_inputs_linalg_cond(op_info, device, dtype, requires_grad=False, **kwa
     for shape in shapes:
         yield SampleInput(make_arg(shape))
 
-def sample_inputs_linalg_vander(op_info, device, dtype, requires_grad=False, **kwargs):
-    make_arg = partial(make_tensor, dtype=dtype, device=device, requires_grad=requires_grad)
-
-    shapes = ((),
-              (1,),
-              (S,),
-              (2, S),)
-
-    for shape in shapes:
-        yield SampleInput(make_arg(shape))
-        n = shape[-1] if len(shape) > 0 else 1
-        for i in range(3):
-            # n-1, n, n+1
-            N = n + i - 1
-            if N < 0:
-                continue
-            yield SampleInput(make_arg(shape), kwargs=dict(N=N))
-
-def np_vander_batched(x, N=None):
-    # Wrapper around np.vander that supports batches of 1 dimension (enough for the tests)
-    if x.ndim == 0:
-        x = x[np.newaxis]
-    if x.ndim == 1:
-        y = np.vander(x, N=N, increasing=True)
-        return y
-    else:
-        if N is None:
-            N = x.shape[-1]
-        y = np.vander(x.ravel(), N=N, increasing=True).reshape((*x.shape, N))
-        return y
-
 def np_sinc_with_fp16_as_fp32(x):
     # Wraps numpy's sinc function so that fp16 values are promoted to fp32
     # before sinc is invoked. Context: numpy's sinc returns NaN when evaluated
@@ -11292,15 +11261,6 @@ op_db: List[OpInfo] = [
            dtypes=floating_and_complex_types(),
            sample_inputs_func=sample_inputs_linalg_slogdet,
            decorators=[skipCUDAIfNoMagma, skipCPUIfNoLapack],),
-    OpInfo('linalg.vander',
-           aten_name='linalg_vander',
-           ref=np_vander_batched,
-           op=torch.linalg.vander,
-           dtypes=all_types_and_complex_and(torch.bool, torch.float16, torch.bfloat16),
-           supports_forward_ad=True,
-           supports_fwgrad_bwgrad=True,
-           supports_out=False,
-           sample_inputs_func=sample_inputs_linalg_vander),
     OpInfo('linalg.vector_norm',
            op=torch.linalg.vector_norm,
            dtypes=floating_and_complex_types_and(torch.float16, torch.bfloat16),