Give std/var correction overloads proper defaults

aten/src/ATen/native/ReduceOps.cpp

@@ -1744,6 +1744,9 @@ static Tensor& std_var_out(
   const auto correction = correction_opt.value_or(1);
   ScalarType dtype = get_dtype_from_result(result, {});
   auto iter = make_reduction(fname, result, self, dim, keepdim, dtype);
+  TORCH_CHECK(at::canCast(self.scalar_type(), result.scalar_type()),
+              "result type ", self.scalar_type(), " can't be cast to the "
+              "desired output type ", result.scalar_type());
 
   if (iter.numel() == 0) {
     // Trivial reduction

aten/src/ATen/native/native_functions.yaml

@@ -5176,12 +5176,14 @@
 - func: std(Tensor self, bool unbiased=True) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  cpp_no_default_args: ["unbiased"]
 
 - func: std.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  cpp_no_default_args: ["unbiased"]
 
-- func: std.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> Tensor
+- func: std.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
@@ -5192,12 +5194,14 @@
 - func: std_mean(Tensor self, bool unbiased=True) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
+  cpp_no_default_args: ["unbiased"]
 
 - func: std_mean.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
+  cpp_no_default_args: ["unbiased"]
 
-- func: std_mean.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> (Tensor, Tensor)
+- func: std_mean.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
@@ -5207,15 +5211,17 @@
 - func: std_mean.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
+  cpp_no_default_args: ["unbiased"]
 
-- func: std_mean.correction_names(Tensor self, Dimname[1] dim, *, int? correction, bool keepdim=False) -> (Tensor, Tensor)
+- func: std_mean.correction_names(Tensor self, Dimname[1] dim, *, int? correction=None, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
 
 - func: std.out(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  cpp_no_default_args: ["unbiased"]
 
-- func: std.correction_out(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
+- func: std.correction_out(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: std_out
@@ -5224,15 +5230,17 @@
 - func: std.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  cpp_no_default_args: ["unbiased"]
 
 - func: std.names_out(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  cpp_no_default_args: ["unbiased"]
 
-- func: std.correction_names(Tensor self, Dimname[1] dim, *, int? correction, bool keepdim=False) -> Tensor
+- func: std.correction_names(Tensor self, Dimname[1] dim, *, int? correction=None, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
 
-- func: std.correction_names_out(Tensor self, Dimname[1] dim, *, int? correction, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
+- func: std.correction_names_out(Tensor self, Dimname[1] dim, *, int? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function
 
@@ -5645,13 +5653,15 @@
 - func: var(Tensor self, bool unbiased=True) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  cpp_no_default_args: ["unbiased"]
 
 - func: var.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
   tags: canonical
+  cpp_no_default_args: ["unbiased"]
 
-- func: var.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> Tensor
+- func: var.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
@@ -5660,36 +5670,41 @@
 
 - func: var.out(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  cpp_no_default_args: ["unbiased"]
 
-- func: var.correction_out(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
+- func: var.correction_out(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: var_out
 
 - func: var.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  cpp_no_default_args: ["unbiased"]
 
 - func: var.names_out(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  cpp_no_default_args: ["unbiased"]
 
-- func: var.correction_names(Tensor self, Dimname[1] dim, *, int? correction, bool keepdim=False) -> Tensor
+- func: var.correction_names(Tensor self, Dimname[1] dim, *, int? correction=None, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
 
-- func: var.correction_names_out(Tensor self, Dimname[1] dim, *, int? correction, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
+- func: var.correction_names_out(Tensor self, Dimname[1] dim, *, int? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function
 
 - func: var_mean(Tensor self, bool unbiased=True) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
+  cpp_no_default_args: ["unbiased"]
 
 - func: var_mean.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
+  cpp_no_default_args: ["unbiased"]
 
-- func: var_mean.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> (Tensor, Tensor)
+- func: var_mean.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
@@ -5699,8 +5714,9 @@
 - func: var_mean.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
+  cpp_no_default_args: ["unbiased"]
 
-- func: var_mean.correction_names(Tensor self, Dimname[1] dim, *, int? correction, bool keepdim=False) -> (Tensor, Tensor)
+- func: var_mean.correction_names(Tensor self, Dimname[1] dim, *, int? correction=None, bool keepdim=False) -> (Tensor, Tensor)
   device_check: NoCheck   # TensorIterator
   variants: function
 

test/functorch/test_aotdispatch.py

@@ -1979,7 +1979,9 @@ def forward(self, x):
     xfail('special.i1', ''),  # aten.i0.default - couldn't find symbolic meta function/decomposition
     xfail('special.polygamma', 'special_polygamma_n_0'),  # aten.polygamma.default - couldn't find symbolic ...
     xfail('std', ''),  # Cannot call numel() on tensor with symbolic sizes/strides
+    xfail('std', 'unbiased'),  # Cannot call numel() on tensor with symbolic sizes/strides
     xfail('std_mean', ''),  # Cannot call numel() on tensor with symbolic sizes/strides
+    xfail('std_mean', 'unbiased'),  # Cannot call numel() on tensor with symbolic sizes/strides
     xfail('stft', ''),  # Cannot call sizes() on tensor with symbolic sizes/strides
     xfail('sum_to_size', ''),  # Cannot call sizes() on tensor with symbolic sizes/strides
     xfail('svd', ''),  # aten._linalg_svd.default - couldn't find symbolic meta function/decomposition
@@ -1994,7 +1996,9 @@ def forward(self, x):
     xfail('triangular_solve', ''),  # aten.triangular_solve.default - couldn't find symbolic meta function/de...
     xfail('unflatten', ''),  # Cannot call sizes() on tensor with symbolic sizes/strides
     xfail('var', ''),  # Cannot call numel() on tensor with symbolic sizes/strides
+    xfail('var', 'unbiased'),  # Cannot call numel() on tensor with symbolic sizes/strides
     xfail('var_mean', ''),  # Cannot call numel() on tensor with symbolic sizes/strides
+    xfail('var_mean', 'unbiased'),  # Cannot call numel() on tensor with symbolic sizes/strides
     xfail('view_as', ''),  # Cannot call sizes() on tensor with symbolic sizes/strides
     xfail('vsplit', ''),  # Cannot call sizes() on tensor with symbolic sizes/strides
 }

test/inductor/test_torchinductor_opinfo.py

@@ -332,6 +332,7 @@ def process(device_type):
     "scatter_reduce.prod": {f16, f32, f64},
     "segment_reduce.lengths": {f16, f32, f64},
     "sparse.sampled_addmm": {f32, f64},
+    "std_mean.unbiased": {f16},
     "stft": {f32, f64},
     "svd_lowrank": {f32, f64},
     "tensor_split": {b8, f16, f32, f64, i32, i64},

tools/autograd/derivatives.yaml

@@ -1514,12 +1514,12 @@
   self: unsqueeze_to(grad, dim, self.sym_sizes())
   result: auto_linear
 
-- name: std.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> Tensor
+- name: std.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> Tensor
   self: std_backward(result, grad, self, dim, correction, keepdim)
   # pointwise (variance) + sum + sqrt
   result: (at::real(var_backward(self_t.conj(), self_p, dim, correction, true).sum(dim.value_or(IntArrayRef({})), keepdim)) / (2. * result)).masked_fill_(result == 0, 0)
 
-- name: std_mean.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> (Tensor, Tensor)
+- name: std_mean.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> (Tensor, Tensor)
   self: std_mean_backward(grads[0], grads[1], self, result0, dim, correction, keepdim)
   result0: (at::real(var_backward(self_t.conj(), self_p, dim, correction, true).sum(dim.value_or(IntArrayRef({})), keepdim)) / (2. * result0)).masked_fill_(result0 == 0, 0)
   # linear
@@ -1723,12 +1723,12 @@
   self: grad.squeeze(dim)
   result: auto_linear
 
-- name: var.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> Tensor
+- name: var.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> Tensor
   self: var_backward(grad, self, dim, correction, keepdim)
   # pointwise + sum
   result: at::real(var_backward(self_t.conj(), self_p, dim, correction, true).sum(dim.value_or(IntArrayRef({})), keepdim))
 
-- name: var_mean.correction(Tensor self, int[1]? dim, *, int? correction, bool keepdim=False) -> (Tensor, Tensor)
+- name: var_mean.correction(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False) -> (Tensor, Tensor)
   self: var_mean_backward(grads[0], grads[1], self, dim, correction, keepdim)
   result0: at::real(var_backward(self_t.conj(), self_p, dim, correction, true).sum(dim.value_or(IntArrayRef({})), keepdim))
   # linear

torch/_decomp/decompositions.py

@@ -1556,7 +1556,7 @@ def xlogy(self: Tensor, other: Tensor) -> Tensor:
 @reduction_complex_to_real
 def std_decomposition(
     x: Tensor,
-    dim: Optional[List[int]],
+    dim: Optional[List[int]] = None,
     correction: Optional[int] = None,
     keepdim: bool = False,
 ):

torch/_inductor/lowering.py

@@ -3292,7 +3292,7 @@ def mean(x, axis=None, keepdim=False, *, dtype=None):
 
 
 @register_lowering([aten.var, prims.var])
-def var_(x, axis, correction=1, keepdim=False):
+def var_(x, axis=None, correction=1, keepdim=False):
     size = x.get_size()
     axis = _validate_reduction_axis(x, axis)
     diffs = square(sub(x, mean(x, axis, keepdim=True)))
@@ -3307,7 +3307,7 @@ def var_(x, axis, correction=1, keepdim=False):
 
 
 @register_lowering(aten.var_mean)
-def var_mean(x, dim, unbiased=True, keepdim=False, correction=None):
+def var_mean(x, dim=None, unbiased=True, keepdim=False, correction=None):
     if correction is None:
         correction = int(unbiased)
     return [
@@ -3317,7 +3317,7 @@ def var_mean(x, dim, unbiased=True, keepdim=False, correction=None):
 
 
 @register_lowering(aten.std)
-def std(x, axis, correction=1, keepdim=False):
+def std(x, axis=None, correction=1, keepdim=False):
     return sqrt(var_(x, axis, correction, keepdim=keepdim))
 
 

torch/_tensor_docs.py

@@ -4785,12 +4785,7 @@ def callable(a, b) -> number
 add_docstr_all(
     "std",
     r"""
-std(dim, unbiased=True, keepdim=False) -> Tensor
-
-See :func:`torch.std`
-
-.. function:: std(unbiased=True) -> Tensor
-   :noindex:
+std(dim=None, *, correction=1, keepdim=False) -> Tensor
 
 See :func:`torch.std`
 """,
@@ -5738,12 +5733,7 @@ def callable(a, b) -> number
 add_docstr_all(
     "var",
     r"""
-var(dim, unbiased=True, keepdim=False) -> Tensor
-
-See :func:`torch.var`
-
-.. function:: var(unbiased=True) -> Tensor
-   :noindex:
+var(dim=None, *, correction=1, keepdim=False) -> Tensor
 
 See :func:`torch.var`
 """,