torch-rb 0.8.0 → 0.9.0

data/codegen/native_functions.yaml

@@ -89,6 +89,10 @@
   manual_cpp_binding: True
   variants: method
 
+- func: retains_grad(Tensor self) -> bool
+  manual_cpp_binding: True
+  variants: method
+
 - func: _fw_primal(Tensor(a) self, int level) -> Tensor(a)
   variants: method
   dispatch:
@@ -278,15 +282,15 @@
 
 - func: sgn(Tensor self) -> Tensor
   variants: function, method
-  dispatch:
-    CompositeExplicitAutograd: sgn
+  structured_delegate: sgn.out
 
 - func: sgn_(Tensor(a!) self) -> Tensor(a!)
   variants: method
-  dispatch:
-    CompositeExplicitAutograd: sgn_
+  structured_delegate: sgn.out
 
 - func: sgn.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sgn_out
 
@@ -298,20 +302,43 @@
   device_check: NoCheck   # TensorIterator
   variants: function
 
+- func: _conj(Tensor(a) self) -> Tensor(a)
+  variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: _conj
+
 - func: conj(Tensor(a) self) -> Tensor(a)
-  device_check: NoCheck   # TensorIterator
   variants: function, method
+  manual_cpp_binding: True
 
-- func: conj.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
-  device_check: NoCheck   # TensorIterator
+- func: _conj_physical(Tensor self) -> Tensor
+  variants: function, method
   dispatch:
-    CPU, CUDA: conj_out
-    SparseCPU, SparseCUDA: conj_out_sparse
+    CompositeExplicitAutograd: _conj_physical
 
-- func: _conj(Tensor self) -> Tensor
-  variants: function
+- func: conj_physical(Tensor self) -> Tensor
+  variants: function, method
+
+- func: conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CompositeExplicitAutograd: _conj
+    CPU, CUDA: conj_physical_out
+    SparseCPU, SparseCUDA: conj_physical_out_sparse
+
+- func: conj_physical_(Tensor(a!) self) -> Tensor(a!)
+  variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: conj_physical_
+
+- func: resolve_conj(Tensor(a) self) -> Tensor(a)
+  variants: function, method
+
+- func: resolve_neg(Tensor(a) self) -> Tensor(a)
+  variants: function, method
+
+- func: _neg_view(Tensor(a) self) -> Tensor(a)
+  variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: _neg_view
 
 - func: acos(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -352,7 +379,7 @@
   variants: function, method
   dispatch:
     SparseCPU, SparseCUDA: add_sparse
-    SparseCsrCPU: add_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA: add_sparse_csr
     MkldnnCPU: mkldnn_add
 
 - func: add_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!)
@@ -361,7 +388,7 @@
   structured_delegate: add.out
   dispatch:
     SparseCPU, SparseCUDA: add_sparse_
-    SparseCsrCPU: add_sparse_csr_
+    SparseCsrCPU, SparseCsrCUDA: add_sparse_csr_
     MkldnnCPU: mkldnn_add_
 
 - func: add.out(Tensor self, Tensor other, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
@@ -373,6 +400,7 @@
     SparseCPU: add_out_sparse_cpu
     SparseCUDA: add_out_sparse_cuda
     SparseCsrCPU: add_out_sparse_csr_cpu
+    SparseCsrCUDA: add_out_sparse_csr_cuda
     MkldnnCPU: mkldnn_add_out
 
 - func: _add_relu.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor
@@ -390,6 +418,16 @@
   dispatch:
     CPU: add_relu_out
 
+- func: _add_relu.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor
+  variants: function
+  dispatch:
+    CPU: add_relu
+
+- func: _add_relu_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!)
+  variants: function
+  dispatch:
+    CPU: add_relu_
+
 # For C++ only, until we have conversion from C++ numbers to Tensor
 - func: add.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -443,12 +481,14 @@
 
 - func: all.dim(Tensor self, int dim, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: all.out
   variants: function, method
-  dispatch:
-    CPU, CUDA: all
 
 - func: all.out(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  precomputed:
+  - dim -> int dim
   dispatch:
     CPU, CUDA: all_out
 
@@ -464,12 +504,14 @@
 
 - func: any.dim(Tensor self, int dim, bool keepdim=False) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: any.out
   variants: function, method
-  dispatch:
-    CPU, CUDA: any
 
 - func: any.out(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  precomputed:
+  - dim -> int dim
   dispatch:
     CPU, CUDA: any_out
 
@@ -501,22 +543,22 @@
 - func: _dim_arange(Tensor like, int dim) -> Tensor
 
 - func: argmax(Tensor self, int? dim=None, bool keepdim=False) -> Tensor
+  structured_delegate: argmax.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CPU, CUDA: argmax
 
 - func: argmax.out(Tensor self, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
     CPU, CUDA: argmax_out
 
 - func: argmin(Tensor self, int? dim=None, bool keepdim=False) -> Tensor
+  structured_delegate: argmin.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CPU, CUDA: argmin
 
 - func: argmin.out(Tensor self, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
     CPU, CUDA: argmin_out
 
@@ -905,11 +947,6 @@
     SparseCPU: bmm_sparse_cpu
     SparseCUDA: bmm_sparse_cuda
 
-- func: _bmm(Tensor self, Tensor mat2, *, bool deterministic=False) -> Tensor
-  variants: function
-  dispatch:
-    SparseCUDA: _bmm_sparse_cuda
-
 - func: bmm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
@@ -918,11 +955,6 @@
     SparseCPU: bmm_out_sparse_cpu
     SparseCUDA: bmm_out_sparse_cuda
 
-- func: _bmm.out(Tensor self, Tensor mat2, *, bool deterministic=False, Tensor(a!) out) -> Tensor(a!)
-  variants: function
-  dispatch:
-    SparseCUDA: _bmm_out_sparse_cuda
-
 - func: broadcast_tensors(Tensor[] tensors) -> Tensor[]
   device_check: NoCheck
   device_guard: False
@@ -942,6 +974,15 @@
 
 - func: cat.names_out(Tensor[] tensors, Dimname dim, *, Tensor(a!) out) -> Tensor(a!)
 
+# alias for torch.cat
+- func: concat(Tensor[] tensors, int dim=0) -> Tensor
+
+- func: concat.out(Tensor[] tensors, int dim=0, *, Tensor(a!) out) -> Tensor(a!)
+
+- func: concat.names(Tensor[] tensors, Dimname dim) -> Tensor
+
+- func: concat.names_out(Tensor[] tensors, Dimname dim, *, Tensor(a!) out) -> Tensor(a!)
+
 - func: block_diag(Tensor[] tensors) -> Tensor
   variants: function
 
@@ -996,8 +1037,8 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   cpp_no_default_args: ['min']
+  structured_delegate: clamp.out
   dispatch:
-    CPU, CUDA: clamp
     QuantizedCPU: clamp_quantized_cpu
 
 - func: clamp.Tensor(Tensor self, Tensor? min=None, Tensor? max=None) -> Tensor
@@ -1009,6 +1050,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   cpp_no_default_args: ['min']
+  structured_delegate: clamp.out
   dispatch:
     CompositeExplicitAutograd: clamp_
 
@@ -1020,6 +1062,8 @@
 - func: clamp.out(Tensor self, Scalar? min=None, Scalar? max=None, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   cpp_no_default_args: ['min']
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: clamp_out
 
@@ -1200,6 +1244,11 @@
 - func: _copy_from(Tensor self, Tensor dst, bool non_blocking=False) -> Tensor
   dispatch: {}
 
+# We need this to be able to properly copy from a CPU to an XLA tensor with different sizes.
+# See https://github.com/pytorch/xla/issues/2881
+- func: _copy_from_and_resize(Tensor self, Tensor dst) -> Tensor
+  dispatch: {}
+
 - func: cos(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
@@ -1239,13 +1288,20 @@
 - func: count_nonzero.dim_IntList(Tensor self, int[] dim) -> Tensor
   variants: function, method
   dispatch:
-    CPU, CUDA: count_nonzero
+    CPU: count_nonzero_cpu
+    CUDA: count_nonzero_cuda
 
 - func: count_nonzero(Tensor self, int? dim=None) -> Tensor
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: count_nonzero
 
+- func: cov(Tensor self, *, int correction=1, Tensor? fweights=None, Tensor? aweights=None) -> Tensor
+  variants: function, method
+
+- func: corrcoef(Tensor self) -> Tensor
+  variants: function, method
+
 - func: cudnn_affine_grid_generator(Tensor theta, int N, int C, int H, int W) -> Tensor grid
   dispatch:
     CUDA: cudnn_affine_grid_generator_forward
@@ -1385,20 +1441,19 @@
   device_guard: False
 
 - func: cumprod(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor
+  structured_delegate: cumprod.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CompositeExplicitAutograd: cumprod
 
 - func: cumprod_(Tensor(a!) self, int dim, *, ScalarType? dtype=None) -> Tensor(a!)
+  structured_delegate: cumprod.out
   variants: method
-  dispatch:
-    CompositeExplicitAutograd: cumprod_
 
 - func: cumprod.out(Tensor self, int dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CompositeExplicitAutograd: cumprod_out
+    CPU, CUDA: cumprod_out
 
 - func: cumprod.dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -1416,20 +1471,19 @@
   device_guard: False
 
 - func: cumsum(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor
+  structured_delegate: cumsum.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CompositeExplicitAutograd: cumsum
 
 - func: cumsum_(Tensor(a!) self, int dim, *, ScalarType? dtype=None) -> Tensor(a!)
+  structured_delegate: cumsum.out
   variants: method
-  dispatch:
-    CompositeExplicitAutograd: cumsum_
 
 - func: cumsum.out(Tensor self, int dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CompositeExplicitAutograd: cumsum_out
+    CPU, CUDA: cumsum_out
 
 - func: cumsum.dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -1441,6 +1495,10 @@
 - func: cumsum.dimname_out(Tensor self, Dimname dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
 
+- func: cumulative_trapezoid.x(Tensor y, Tensor x, *, int dim=-1) -> Tensor
+
+- func: cumulative_trapezoid.dx(Tensor y, *, Scalar dx=1, int dim=-1) -> Tensor
+
 - func: ctc_loss.IntList(Tensor log_probs, Tensor targets, int[] input_lengths, int[] target_lengths, int blank=0, int reduction=Mean, bool zero_infinity=False) -> Tensor
 
 # convenience function that converts to intlists for you
@@ -1470,10 +1528,12 @@
 - func: diagonal.Dimname(Tensor(a) self, *, Dimname outdim, Dimname dim1, Dimname dim2, int offset=0) -> Tensor(a)
   variants: function, method
 
-- func: diagonal_backward(Tensor grad, int[] input_sizes, int offset, int dim1, int dim2) -> Tensor
+- func: diagonal_backward(Tensor grad_output, int[] input_sizes, int offset, int dim1, int dim2) -> Tensor
   variants: function
   device_check: NoCheck
   device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: diagonal_backward
 
 - func: fill_diagonal_(Tensor(a!) self, Scalar fill_value, bool wrap=False) -> Tensor(a!)
   variants: method
@@ -1734,6 +1794,9 @@
 - func: new_zeros(Tensor self, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   variants: method
 
+- func: new_ones(Tensor self, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
+  variants: method
+
 # other overrides are to provide a more helpful error message that dtype is required
 - func: _empty_affine_quantized(int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, float scale=1, int zero_point=0, MemoryFormat? memory_format=contiguous_format) -> Tensor
   dispatch:
@@ -1758,7 +1821,8 @@
     CUDA: resize_cuda_
     QuantizedCPU: quantized_resize_cpu_
 
-- func: empty_quantized(int[] size, Tensor qtensor) -> Tensor
+- func: empty_quantized(int[] size, Tensor qtensor, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
+  category_override: factory
   variants: function
   dispatch:
     QuantizedCPU, QuantizedCUDA: empty_quantized
@@ -2214,6 +2278,36 @@
 - func: isclose(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False) -> Tensor
   variants: function, method
 
+- func: isin.Tensor_Tensor_out(Tensor elements, Tensor test_elements, *, bool assume_unique=False, bool invert=False, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  structured: True
+  dispatch:
+    CPU, CUDA: isin_Tensor_Tensor_out
+
+- func: isin.Tensor_Tensor(Tensor elements, Tensor test_elements, *, bool assume_unique=False, bool invert=False) -> Tensor
+  variants: function
+  structured_delegate: isin.Tensor_Tensor_out
+
+- func: isin.Tensor_Scalar_out(Tensor elements, Scalar test_element, *, bool assume_unique=False, bool invert=False, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  structured: True
+  dispatch:
+    CPU, CUDA: isin_Tensor_Scalar_out
+
+- func: isin.Tensor_Scalar(Tensor elements, Scalar test_element, *, bool assume_unique=False, bool invert=False) -> Tensor
+  variants: function
+  structured_delegate: isin.Tensor_Scalar_out
+
+- func: isin.Scalar_Tensor_out(Scalar element, Tensor test_elements, *, bool assume_unique=False, bool invert=False, Tensor(a!) out) -> Tensor(a!)
+  variants: function
+  structured: True
+  dispatch:
+    CPU, CUDA: isin_Scalar_Tensor_out
+
+- func: isin.Scalar_Tensor(Scalar element, Tensor test_elements, *, bool assume_unique=False, bool invert=False) -> Tensor
+  variants: function
+  structured_delegate: isin.Scalar_Tensor_out
+
 - func: isnan(Tensor self) -> Tensor
   variants: function, method
   device_check: NoCheck
@@ -2239,6 +2333,16 @@
   device_guard: False
   manual_cpp_binding: True
 
+- func: is_conj(Tensor self) -> bool
+  variants: function, method
+  device_guard: False
+  manual_cpp_binding: True
+
+- func: is_neg(Tensor self) -> bool
+  variants: function, method
+  device_guard: False
+  manual_cpp_binding: True
+
 - func: isreal(Tensor self) -> Tensor
   variants: function, method
 
@@ -2258,6 +2362,12 @@
   device_guard: False
   manual_cpp_binding: True
 
+- func: is_inference(Tensor self) -> bool
+  variants: function, method
+  device_check: NoCheck
+  device_guard: False
+  manual_cpp_binding: True
+
 - func: kl_div(Tensor self, Tensor target, int reduction=Mean, *, bool log_target=False) -> Tensor
   dispatch:
     CompositeExplicitAutograd: kl_div
@@ -2317,6 +2427,9 @@
 - func: linear(Tensor input, Tensor weight, Tensor? bias=None) -> Tensor
   python_module: nn
 
+- func: linear.out(Tensor input, Tensor weight, Tensor? bias=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+
 - func: mkldnn_linear(Tensor self, Tensor weight, Tensor? bias=None) -> Tensor
   python_module: nn
   dispatch:
@@ -2464,38 +2577,38 @@
 
 - func: xlogy.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: xlogy.OutTensor
   variants: function, method
-  dispatch:
-    CPU, CUDA: xlogy
 
 - func: xlogy.Scalar_Self(Scalar self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
-    CPU, CUDA: xlogy
+    CompositeExplicitAutograd: xlogy
 
 - func: xlogy.Scalar_Other(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: xlogy
+    CompositeExplicitAutograd: xlogy
 
 # xlogy: inplace variant
 - func: xlogy_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CPU, CUDA: xlogy_
+  structured_delegate: xlogy.OutTensor
 
 - func: xlogy_.Scalar_Other(Tensor(a!) self, Scalar other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: xlogy_
+    CompositeExplicitAutograd: xlogy_
 
 # xlogy: out variant
 - func: xlogy.OutTensor(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   variants: function
   dispatch:
     CPU, CUDA: xlogy_out
@@ -2504,13 +2617,13 @@
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
-    CPU, CUDA: xlogy_out
+    CompositeExplicitAutograd: xlogy_out
 
 - func: xlogy.OutScalar_Other(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
-    CPU, CUDA: xlogy_out
+    CompositeExplicitAutograd: xlogy_out
 
 - func: logdet(Tensor self) -> Tensor
   variants: function, method
@@ -2532,14 +2645,22 @@
   variants: function, method
 
 - func: _log_softmax(Tensor self, int dim, bool half_to_float) -> Tensor
+  structured_delegate: _log_softmax.out
+
+- func: _log_softmax.out(Tensor self, int dim, bool half_to_float, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: log_softmax_cpu
-    CUDA: log_softmax_cuda
+    CPU: log_softmax_cpu_out
+    CUDA: log_softmax_cuda_out
 
 - func: _log_softmax_backward_data(Tensor grad_output, Tensor output, int dim, Tensor self) -> Tensor
+  structured_delegate: _log_softmax_backward_data.out
+
+- func: _log_softmax_backward_data.out(Tensor grad_output, Tensor output, int dim, Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: log_softmax_backward_cpu
-    CUDA: log_softmax_backward_cuda
+    CPU: log_softmax_backward_cpu_out
+    CUDA: log_softmax_backward_cuda_out
 
 - func: _logcumsumexp(Tensor self, int dim) -> Tensor
   dispatch:
@@ -2608,16 +2729,27 @@
 
 - func: matrix_exp_backward(Tensor self, Tensor grad) -> Tensor
 
+# DEPRECATED: Use torch.aminmax instead
 - func: _aminmax(Tensor self) -> (Tensor, Tensor)
-  variants: function
   dispatch:
     CPU, CUDA: _aminmax_all
 
+# DEPRECATED: Use torch.aminmax instead
 - func: _aminmax.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor, Tensor)
-  variants: function
   dispatch:
     CPU, CUDA: _aminmax
 
+- func: aminmax(Tensor self, *, int? dim=None, bool keepdim=False) -> (Tensor min, Tensor max)
+  device_check: NoCheck   # TensorIterator
+  structured_delegate: aminmax.out
+  variants: function, method
+
+- func: aminmax.out(Tensor self, *, int? dim=None, bool keepdim=False, Tensor(a!) min, Tensor(b!) max) -> (Tensor(a!) min, Tensor(b!) max)
+  device_check: NoCheck   # TensorIterator
+  structured: True
+  dispatch:
+    CPU, CUDA: aminmax_out
+
 - func: _compute_linear_combination(Tensor input, Tensor coefficients) -> Tensor
   dispatch:
     CPU, CUDA: _compute_linear_combination
@@ -2697,20 +2829,20 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: mean_cpu_gpu
-    QuantizedCPU: mean_quantized_cpu
+    CompositeExplicitAutograd: mean
 
 - func: mean.dim(Tensor self, int[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
+  structured_delegate: mean.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: mean_cpu_gpu
     QuantizedCPU: mean_quantized_cpu
 
 - func: mean.out(Tensor self, int[1] dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: mean_out_cpu_gpu
+    CPU, CUDA: mean_out
     QuantizedCPU: mean_out_quantized_cpu
 
 - func: mean.names_dim(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
@@ -2720,6 +2852,13 @@
 - func: mean.names_out(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
 
+- func: nanmean(Tensor self, int[1] dim=[], bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
+  device_check: NoCheck   # Composite
+  variants: function, method
+
+- func: nanmean.out(Tensor self, int[1] dim=[], bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # Composite
+
 - func: median(Tensor self) -> Tensor
   variants: function, method
   dispatch:
@@ -2872,18 +3011,18 @@
     CUDA: miopen_rnn_backward
 
 - func: mm(Tensor self, Tensor mat2) -> Tensor
+  structured_delegate: mm.out
   variants: function, method
   dispatch:
-    CPU: mm_cpu
-    CUDA: mm_cuda
-    SparseCPU, SparseCUDA, SparseCsrCPU: _sparse_mm
+    SparseCPU, SparseCUDA, SparseCsrCPU, SparseCsrCUDA: _sparse_mm
 
 - func: mm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: mm_cpu_out
+    CPU: mm_out_cpu
     CUDA: mm_out_cuda
     SparseCPU, SparseCUDA: _sparse_mm_out
-    SparseCsrCPU: _sparse_csr_mm_out
+    SparseCsrCPU, SparseCsrCUDA: _sparse_csr_mm_out
 
 - func: _sparse_mm(Tensor sparse, Tensor dense) -> Tensor
 
@@ -2969,12 +3108,16 @@
   variants: function, method
   dispatch:
     CPU, CUDA: mv
-    SparseCPU, SparseCUDA, SparseCsrCPU: mv_sparse
+    SparseCPU, SparseCUDA, SparseCsrCPU, SparseCsrCUDA: mv_sparse
 
 - func: mv.out(Tensor self, Tensor vec, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CompositeExplicitAutograd: mv_out
 
+- func: mvlgamma.out(Tensor self, int p, *, Tensor(a!) out) -> Tensor(a!)
+  dispatch:
+    CPU, CUDA: mvlgamma_out
+
 - func: mvlgamma(Tensor self, int p) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
@@ -3152,12 +3295,22 @@
     CPU: channel_shuffle
     QuantizedCPU: channel_shuffle_quantized_cpu
 
-- func: is_pinned(Tensor self) -> bool
+- func: is_pinned(Tensor self, Device? device=None) -> bool
   variants: method
+  dispatch:
+    CUDA: is_pinned_cuda
+    CompositeExplicitAutograd: is_pinned_default
 
-- func: pin_memory(Tensor(a) self) -> Tensor(a)
+# TODO: add a copy kwarg that guarantees that the tensor is put into fresh
+# pinned memory
+- func: pin_memory(Tensor(a) self, Device? device=None) -> Tensor(a)
   variants: method
 
+# Unlike pin_memory, this is guaranteed to give a new non-aliasing tensor
+- func: _pin_memory(Tensor self, Device? device=None) -> Tensor
+  dispatch:
+    CUDA: _pin_memory_cuda
+
 - func: pinverse(Tensor self, float rcond=1e-15) -> Tensor
   variants: function, method
 
@@ -3326,16 +3479,16 @@
   dispatch:
     CompositeExplicitAutograd: repeat
 
-- func: repeat_interleave.Tensor(Tensor repeats) -> Tensor
+- func: repeat_interleave.Tensor(Tensor repeats, *, int? output_size=None) -> Tensor
   variants: function
   dispatch:
     CPU: repeat_interleave_cpu
     CUDA: repeat_interleave_cuda
 
-- func: repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None) -> Tensor
+- func: repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None, *, int? output_size=None) -> Tensor
   variants: function, method
 
-- func: repeat_interleave.self_int(Tensor self, int repeats, int? dim=None) -> Tensor
+- func: repeat_interleave.self_int(Tensor self, int repeats, int? dim=None, *, int? output_size=None) -> Tensor
   variants: function, method
 
 - func: reshape(Tensor(a) self, int[] shape) -> Tensor(a)
@@ -3343,6 +3496,17 @@
   device_check: NoCheck
   device_guard: False
 
+# NOTE [ _reshape_alias ] is meant to be used in the implementation of reshape.
+# They are not user-facing, hence the leading underscore. Please don't use it
+# anywhere else.
+- func: _reshape_alias(Tensor(a) self, int[] size, int[] stride) -> Tensor(a)
+  variants: function, method
+  device_check: NoCheck
+  device_guard: False
+  dispatch:
+    CPU, CUDA, Meta, QuantizedCPU, QuantizedCUDA: _reshape_alias
+    # We don't need to support mkldnn since this is handled explicitly by the reshape operator.
+
 - func: _mkldnn_reshape(Tensor self, int[] shape) -> Tensor
   device_check: NoCheck
   device_guard: False
@@ -3412,19 +3576,35 @@
     CPU: prelu_backward_cpu
     CUDA: prelu_backward_cuda
 
+- func: gelu.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  device_check: NoCheck   # TensorIterator
+  python_module: nn
+  dispatch:
+    CPU: gelu_out_cpu
+    CUDA: gelu_out_cuda
+
 - func: gelu(Tensor self) -> Tensor
+  structured_delegate: gelu.out
   device_check: NoCheck   # TensorIterator
   python_module: nn
   dispatch:
     MkldnnCPU: mkldnn_gelu
-    CPU: gelu_cpu
-    CUDA: gelu_cuda
+
+- func: gelu_backward.grad_input(Tensor grad, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  python_module: nn
+  dispatch:
+    CPU: gelu_backward_out_cpu
+    CUDA: gelu_backward_out_cuda
 
 - func: gelu_backward(Tensor grad, Tensor self) -> Tensor
+  structured_delegate: gelu_backward.grad_input
   python_module: nn
   dispatch:
-    CPU: gelu_backward_cpu
-    CUDA: gelu_backward_cuda
+    MkldnnCPU: mkldnn_gelu_backward
 
 - func: infinitely_differentiable_gelu_backward(Tensor grad, Tensor self) -> Tensor
   variants: function
@@ -3432,16 +3612,27 @@
   device_check: NoCheck
   device_guard: False
 
+- func: hardshrink.out(Tensor self, Scalar lambd=0.5, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  device_check: NoCheck   # TensorIterator
+  dispatch:
+    CPU, CUDA: hardshrink_out
+
 - func: hardshrink(Tensor self, Scalar lambd=0.5) -> Tensor
+  structured_delegate: hardshrink.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
+
+- func: hardshrink_backward.grad_input(Tensor grad_out, Tensor self, Scalar lambd, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: hardshrink
+    CPU, CUDA: hardshrink_backward_out
 
 - func: hardshrink_backward(Tensor grad_out, Tensor self, Scalar lambd) -> Tensor
+  structured_delegate: hardshrink_backward.grad_input
   variants: function, method
-  dispatch:
-    CPU, CUDA: hardshrink_backward
 
 - func: rsqrt(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -3472,10 +3663,12 @@
   dispatch:
     CompositeExplicitAutograd: select
 
-- func: select_backward(Tensor grad, int[] input_sizes, int dim, int index) -> Tensor
+- func: select_backward(Tensor grad_output, int[] input_sizes, int dim, int index) -> Tensor
   variants: function
   device_check: NoCheck
   device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: select_backward
 
 - func: selu(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -3512,10 +3705,17 @@
   dispatch:
     CPU, CUDA: silu_out
 
+- func: silu_backward.grad_input(Tensor grad_output, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  python_module: nn
+  dispatch:
+    CPU, CUDA: silu_backward_out
+
 - func: silu_backward(Tensor grad_output, Tensor self) -> Tensor
+  structured_delegate: silu_backward.grad_input
   python_module: nn
   dispatch:
-    CPU, CUDA: silu_backward
     CompositeImplicitAutograd: math_silu_backward
 
 - func: mish(Tensor self) -> Tensor
@@ -3669,10 +3869,12 @@
   dispatch:
     CompositeExplicitAutograd: slice
 
-- func: slice_backward(Tensor grad, int[] input_sizes, int dim, int start, int end, int step) -> Tensor
+- func: slice_backward(Tensor grad_output, int[] input_sizes, int dim, int start, int end, int step) -> Tensor
   variants: function
   device_check: NoCheck
   device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: slice_backward
 
 - func: slogdet(Tensor self) -> (Tensor sign, Tensor logabsdet)
   variants: function, method
@@ -3690,15 +3892,24 @@
   variants: function, method
 
 - func: _softmax(Tensor self, int dim, bool half_to_float) -> Tensor
+  structured_delegate: _softmax.out
   dispatch:
-    CPU: softmax_cpu
-    CUDA: softmax_cuda
     MkldnnCPU: mkldnn_softmax
 
+- func: _softmax.out(Tensor self, int dim, bool half_to_float, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  dispatch:
+    CPU: softmax_cpu_out
+    CUDA: softmax_cuda_out
+
 - func: _softmax_backward_data(Tensor grad_output, Tensor output, int dim, Tensor self) -> Tensor
+  structured_delegate: _softmax_backward_data.out
+
+- func: _softmax_backward_data.out(Tensor grad_output, Tensor output, int dim, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: softmax_backward_cpu
-    CUDA: softmax_backward_cuda
+    CPU: softmax_backward_cpu_out
+    CUDA: softmax_backward_cuda_out
 
 - func: unsafe_split.Tensor(Tensor self, int split_size, int dim=0) -> Tensor[]
   variants: function, method
@@ -3849,19 +4060,19 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: sum
+    CompositeExplicitAutograd: sum
 
 - func: sum.dim_IntList(Tensor self, int[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
+  structured_delegate: sum.IntList_out
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CPU, CUDA: sum
 
 - func: sum.dim_DimnameList(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
 
 - func: sum.IntList_out(Tensor self, int[1] dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: sum_out
@@ -3986,12 +4197,12 @@
     CPU, CUDA: prod
 
 - func: prod.dim_int(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
+  structured_delegate: prod.int_out
   device_check: NoCheck   # TensorIterator
   variants: function, method
-  dispatch:
-    CPU, CUDA: prod
 
 - func: prod.int_out(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: prod_out
@@ -4136,8 +4347,7 @@
 - func: flip(Tensor self, int[] dims) -> Tensor
   variants: function, method
   dispatch:
-    CPU, QuantizedCPU: flip_cpu
-    CUDA: flip_cuda
+    CPU, QuantizedCPU, CUDA, QuantizedCUDA: flip
 
 - func: fliplr(Tensor self) -> Tensor
   variants: function, method
@@ -4158,6 +4368,10 @@
   dispatch:
     CompositeExplicitAutograd: rot90
 
+- func: trapezoid.x(Tensor y, Tensor x, *, int dim=-1) -> Tensor
+
+- func: trapezoid.dx(Tensor y, *, Scalar dx=1, int dim=-1) -> Tensor
+
 - func: trapz.x(Tensor y, Tensor x, *, int dim=-1) -> Tensor
 
 - func: trapz.dx(Tensor y, *, float dx=1, int dim=-1) -> Tensor
@@ -4476,32 +4690,36 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, SparseCPU, SparseCUDA: norm
+    CompositeExplicitAutograd: norm
 
 - func: norm.Scalar(Tensor self, Scalar p=2) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, SparseCPU, SparseCUDA: norm
+    CompositeExplicitAutograd: norm
 
 - func: norm.ScalarOpt_dim_dtype(Tensor self, Scalar? p, int[1] dim, bool keepdim, *, ScalarType dtype) -> Tensor
+  structured_delegate: norm.dtype_out
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, SparseCPU, SparseCUDA: norm
+    SparseCPU, SparseCUDA: sparse_dtype_norm
 
 - func: norm.ScalarOpt_dim(Tensor self, Scalar? p, int[1] dim, bool keepdim=False) -> Tensor
+  structured_delegate: norm.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, SparseCPU, SparseCUDA: norm
+    SparseCPU, SparseCUDA: sparse_norm
 
 - func: norm.dtype_out(Tensor self, Scalar? p, int[1] dim, bool keepdim, *, ScalarType dtype, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: norm_out
+    CPU, CUDA: norm_dtype_out
 
 - func: norm.out(Tensor self, Scalar? p, int[1] dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: norm_out
@@ -4573,7 +4791,7 @@
   variants: function
   dispatch:
     SparseCPU, SparseCUDA: resize_as_sparse_
-    SparseCsrCPU: resize_as_sparse_csr_
+    SparseCsrCPU, SparseCsrCUDA: resize_as_sparse_csr_
 
 - func: zero_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -4679,6 +4897,7 @@
     SparseCPU: addmm_out_sparse_dense_cpu
     SparseCUDA: addmm_out_sparse_dense_cuda
     SparseCsrCPU: addmm_out_sparse_csr_dense_cpu
+    SparseCsrCUDA: addmm_out_sparse_csr_dense_cuda
 
 - func: addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   structured_delegate: addmm.out
@@ -4686,7 +4905,7 @@
   dispatch:
     SparseCPU: addmm_sparse_dense_cpu
     SparseCUDA: addmm_sparse_dense_cuda
-    SparseCsrCPU: addmm_sparse_csr_dense_cpu
+    SparseCsrCPU, SparseCsrCUDA: addmm_sparse_csr_dense
 
 - func: addmm_(Tensor(a!) self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
   structured_delegate: addmm.out
@@ -4808,9 +5027,11 @@
 # FIXME: would be nicer if TensorOptions was optional based; not adding default arguments for options given
 # the default would never make sense.
 
-- func: _sparse_csr_tensor.crow_col_value_size(Tensor crow_indices, Tensor col_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
+- func: sparse_csr_tensor.crow_col_value_size(Tensor crow_indices, Tensor col_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
 
-- func: _sparse_csr_tensor.crow_col_value(Tensor crow_indices, Tensor col_indices, Tensor values, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
+- func: sparse_csr_tensor.crow_col_value(Tensor crow_indices, Tensor col_indices, Tensor values, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
+
+- func: _sparse_csr_tensor_unsafe(Tensor crow_indices, Tensor col_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 
 - func: sparse_coo_tensor.size(int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
 
@@ -4822,6 +5043,8 @@
 
 - func: _validate_sparse_coo_tensor_args(Tensor indices, Tensor values, int[] size) -> ()
 
+- func: _validate_sparse_csr_tensor_args(Tensor crow_indices, Tensor col_indices, Tensor values, int[] size) -> ()
+
 - func: _sparse_coo_tensor_with_dims(int sparse_dim, int dense_dim, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
   dispatch:
     SparseCPU, SparseCUDA: new_with_dims_sparse
@@ -4848,10 +5071,13 @@
     SparseCPU: sparse_mask_cpu
     SparseCUDA: sparse_mask_cuda
 
+- func: _to_cpu(Tensor[] tensors) -> Tensor[]
+  variants: function
+
 - func: to_dense(Tensor self, ScalarType? dtype=None) -> Tensor
   variants: method
   dispatch:
-    SparseCPU, SparseCUDA, SparseCsrCPU: sparse_to_dense
+    SparseCPU, SparseCUDA, SparseCsrCPU, SparseCsrCUDA: sparse_to_dense
     MkldnnCPU: mkldnn_to_dense
 
 - func: to_dense_backward(Tensor grad, Tensor input) -> Tensor
@@ -4890,7 +5116,7 @@
   variants: method
   dispatch:
     SparseCPU, SparseCUDA: _nnz_sparse
-    SparseCsrCPU: _nnz_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA: _nnz_sparse_csr
   device_check: NoCheck
   device_guard: False
 
@@ -4949,21 +5175,21 @@
   variants: method
   dispatch:
     SparseCPU, SparseCUDA: values_sparse
-    SparseCsrCPU: values_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA: values_sparse_csr
   device_check: NoCheck
   device_guard: False
 
 - func: crow_indices(Tensor(a) self) -> Tensor(a)
   variants: method
   dispatch:
-    SparseCsrCPU: crow_indices_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA: crow_indices_sparse_csr
   device_check: NoCheck
   device_guard: False
 
 - func: col_indices(Tensor(a) self) -> Tensor(a)
   variants: method
   dispatch:
-    SparseCsrCPU: col_indices_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA: col_indices_sparse_csr
   device_check: NoCheck
   device_guard: False
 
@@ -5025,6 +5251,11 @@
   dispatch:
     CPU, CUDA: quantize_per_tensor
 
+- func: quantize_per_tensor.tensor_qparams(Tensor self, Tensor scale, Tensor zero_point, ScalarType dtype) -> Tensor
+  variants: function
+  dispatch:
+    CPU, CUDA: quantize_per_tensor_tensor_qparams
+
 - func: quantize_per_tensor.tensors(Tensor[] tensors, Tensor scales, Tensor zero_points, ScalarType dtype) -> Tensor[]
   variants: function
   dispatch:
@@ -5033,13 +5264,13 @@
 - func: quantize_per_channel(Tensor self, Tensor scales, Tensor zero_points, int axis, ScalarType dtype) -> Tensor
   variants: function
   dispatch:
-    CPU: quantize_per_channel_cpu
+    CPU, CUDA: quantize_per_channel
 
 - func: dequantize.self(Tensor self) -> Tensor
   variants: function, method
   dispatch:
     CPU: dequantize_cpu
-    QuantizedCPU, QuantizedCUDA: dequantize_quantized_cpu
+    QuantizedCPU, QuantizedCUDA: dequantize_quantized
 
 - func: dequantize.tensors(Tensor[] tensors) -> Tensor[]
   variants: function
@@ -5086,6 +5317,7 @@
 - func: _make_per_channel_quantized_tensor(Tensor self, Tensor scale, Tensor zero_point, int axis) -> Tensor
   dispatch:
     CPU: make_per_channel_quantized_tensor_cpu
+    CUDA: make_per_channel_quantized_tensor_cuda
 
 - func: qscheme(Tensor self) -> QScheme
   variants: method
@@ -5096,11 +5328,20 @@
   device_check: NoCheck   # TensorIterator
   variants: function
 
+- func: fake_quantize_per_tensor_affine.tensor_qparams(Tensor self, Tensor scale, Tensor zero_point, int quant_min, int quant_max) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  variants: function
+
 - func: fake_quantize_per_tensor_affine_cachemask(Tensor self, float scale, int zero_point, int quant_min, int quant_max) -> (Tensor output, Tensor mask)
   variants: function
   dispatch:
     CPU, CUDA: fake_quantize_per_tensor_affine_cachemask
 
+- func: _fake_quantize_per_tensor_affine_cachemask_tensor_qparams(Tensor self, Tensor scale, Tensor zero_point, Tensor fake_quant_enabled, int quant_min, int quant_max) -> (Tensor output, Tensor mask)
+  variants: function
+  dispatch:
+    CPU, CUDA: _fake_quantize_per_tensor_affine_cachemask_tensor_qparams
+
 - func: fake_quantize_per_tensor_affine_cachemask_backward(Tensor grad, Tensor mask) -> Tensor
   variants: function
 
@@ -5132,6 +5373,15 @@
 - func: _fake_quantize_learnable_per_channel_affine_backward(Tensor grad, Tensor self, Tensor scale, Tensor zero_point, int axis, int quant_min, int quant_max, float grad_factor=1.0) -> (Tensor, Tensor, Tensor)
   variants: function
 
+- func: fused_moving_avg_obs_fake_quant(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor(a!) running_min, Tensor(b!) running_max, Tensor(c!) scale, Tensor(d!) zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False) -> Tensor
+  variants: function
+
+- func: _fused_moving_avg_obs_fq_helper(Tensor self, Tensor observer_on, Tensor fake_quant_on, Tensor(a!) running_min, Tensor(b!) running_max, Tensor(c!) scale, Tensor(d!) zero_point, float averaging_const, int quant_min, int quant_max, int ch_axis, bool per_row_fake_quant=False, bool symmetric_quant=False) -> (Tensor output, Tensor mask)
+  dispatch:
+    CPU: fused_moving_avg_obs_fake_quant_cpu
+    CUDA: fused_moving_avg_obs_fake_quant_cuda
+
+
 - func: _choose_qparams_per_tensor(Tensor self, bool reduce_range=False) -> (float, int)
   variants: function
 
@@ -5141,31 +5391,42 @@
 - func: choose_qparams_optimized(Tensor input, int numel, int n_bins, float ratio, int bit_width) -> (Tensor, Tensor)
   variants: function
 
+- func: _to_copy(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, MemoryFormat? memory_format=None) -> Tensor
+  device_check: NoCheck
+  device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: _to_copy
+
 # to(Device) must not exist because all constructors of Device also works for
 # TensorOptions. Otherwise, an ambiguity error is thrown.
 # See NOTE [ TensorOptions Constructors ].
-- func: to.dtype_layout(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor
+- func: to.dtype_layout(Tensor(a) self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)
   variants: method
   device_check: NoCheck
   device_guard: False
 
-- func: to.device(Tensor self, Device device, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor
+- func: to.device(Tensor(a) self, Device device, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)
   variants: method
   device_check: NoCheck
   device_guard: False
 
-- func: to.dtype(Tensor self, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor
+- func: to.dtype(Tensor(a) self, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)
   variants: method
   device_check: NoCheck
   device_guard: False
 
-- func: to.other(Tensor self, Tensor other, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor
+- func: to.other(Tensor(a) self, Tensor other, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)
   variants: method
   device_check: NoCheck
   device_guard: False
 
 - func: meshgrid(Tensor[] tensors) -> Tensor[]
 
+# TODO: Two weeks after this lands, combine these two overloads,
+#       making "indexing" optional. These are temporarily distinct for
+#       forward-compatibility reasons.
+- func: meshgrid.indexing(Tensor[] tensors, *, str indexing) -> Tensor[]
+
 - func: cartesian_prod(Tensor[] tensors) -> Tensor
   variants: function
 
@@ -5433,56 +5694,94 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
 
-- func: scatter_.src(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!)
-  variants: method
-  dispatch:
-    CPU, CUDA: scatter_
-
 - func: scatter.src(Tensor self, int dim, Tensor index, Tensor src) -> Tensor
+  structured_delegate: scatter.src_out
   variants: function, method
 
-- func: scatter_.value(Tensor(a!) self, int dim, Tensor index, Scalar value) -> Tensor(a!)
+- func: scatter_.src(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!)
+  structured_delegate: scatter.src_out
   variants: method
+
+- func: scatter.src_out(Tensor self, int dim, Tensor index, Tensor src, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
   dispatch:
-    CPU, CUDA: scatter_fill_
+    CPU, CUDA: scatter_src_out
 
 - func: scatter.value(Tensor self, int dim, Tensor index, Scalar value) -> Tensor
+  structured_delegate: scatter.value_out
   variants: function, method
 
-- func: scatter.dimname_src(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor
-  variants: function, method
+- func: scatter_.value(Tensor(a!) self, int dim, Tensor index, Scalar value) -> Tensor(a!)
+  structured_delegate: scatter.value_out
+  variants: method
 
-- func: scatter.dimname_value(Tensor self, Dimname dim, Tensor index, Scalar value) -> Tensor
+- func: scatter.value_out(Tensor self, int dim, Tensor index, Scalar value, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
+  dispatch:
+    CPU, CUDA: scatter_value_out
+
+- func: scatter.reduce(Tensor self, int dim, Tensor index, Tensor src, *, str reduce) -> Tensor
+  structured_delegate: scatter.reduce_out
   variants: function, method
 
 - func: scatter_.reduce(Tensor(a!) self, int dim, Tensor index, Tensor src, *, str reduce) -> Tensor(a!)
+  structured_delegate: scatter.reduce_out
   variants: method
+
+- func: scatter.reduce_out(Tensor self, int dim, Tensor index, Tensor src, *, str reduce, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
   dispatch:
-    CPU, CUDA: scatter_reduce_
+    CPU, CUDA: scatter_reduce_out
+
+- func: scatter.value_reduce(Tensor self, int dim, Tensor index, Scalar value, *, str reduce) -> Tensor
+  structured_delegate: scatter.value_reduce_out
+  variants: function, method
 
 - func: scatter_.value_reduce(Tensor(a!) self, int dim, Tensor index, Scalar value, *, str reduce) -> Tensor(a!)
+  structured_delegate: scatter.value_reduce_out
   variants: method
-  dispatch:
-    CPU, CUDA: scatter_scalar_reduce_
 
-- func: scatter_add_(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!)
-  variants: method
+- func: scatter.value_reduce_out(Tensor self, int dim, Tensor index, Scalar value, *, str reduce, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
   dispatch:
-    CPU, CUDA: scatter_add_
+    CPU, CUDA: scatter_value_reduce_out
+
+- func: scatter.dimname_src(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor
+  variants: function, method
+
+- func: scatter.dimname_value(Tensor self, Dimname dim, Tensor index, Scalar value) -> Tensor
+  variants: function, method
 
 - func: scatter_add(Tensor self, int dim, Tensor index, Tensor src) -> Tensor
+  structured_delegate: scatter_add.out
   variants: function, method
 
+- func: scatter_add_(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!)
+  structured_delegate: scatter_add.out
+  variants: method
+
+- func: scatter_add.out(Tensor self, int dim, Tensor index, Tensor src, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
+  dispatch:
+    CPU, CUDA: scatter_add
+
 - func: scatter_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor
   variants: function, method
 
 - func: eq_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  structured_delegate: eq.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
     CompositeExplicitAutograd: eq_
 
 - func: eq_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  structured_delegate: eq.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
@@ -5490,6 +5789,8 @@
 
 - func: bitwise_and.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   variants: function
   dispatch:
     CPU, CUDA: bitwise_and_out
@@ -5498,15 +5799,18 @@
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
-    CPU, CUDA: bitwise_and_out
+    CompositeExplicitAutograd: bitwise_and_out
 
 - func: bitwise_and.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
+  dispatch:
+    CompositeExplicitAutograd: bitwise_and
 
 - func: bitwise_and.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
+  structured_delegate: bitwise_and.Tensor_out
 
 - func: bitwise_and_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5515,6 +5819,7 @@
 - func: bitwise_and_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  structured_delegate: bitwise_and.Tensor_out
 
 - func: __and__.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -5534,6 +5839,8 @@
 
 - func: bitwise_or.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   variants: function
   dispatch:
     CPU, CUDA: bitwise_or_out
@@ -5542,7 +5849,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
-    CPU, CUDA: bitwise_or_out
+    CompositeExplicitAutograd: bitwise_or_out
 
 - func: bitwise_or.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -5551,6 +5858,7 @@
 - func: bitwise_or.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
+  structured_delegate: bitwise_or.Tensor_out
 
 - func: bitwise_or_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5559,6 +5867,7 @@
 - func: bitwise_or_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  structured_delegate: bitwise_or.Tensor_out
 
 - func: __or__.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -5578,6 +5887,8 @@
 
 - func: bitwise_xor.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   variants: function
   dispatch:
     CPU, CUDA: bitwise_xor_out
@@ -5586,7 +5897,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function
   dispatch:
-    CPU, CUDA: bitwise_xor_out
+    CompositeExplicitAutograd: bitwise_xor_out
 
 - func: bitwise_xor.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -5595,6 +5906,7 @@
 - func: bitwise_xor.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
+  structured_delegate: bitwise_xor.Tensor_out
 
 - func: bitwise_xor_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5603,6 +5915,7 @@
 - func: bitwise_xor_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
+  structured_delegate: bitwise_xor.Tensor_out
 
 - func: __xor__.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -5644,6 +5957,47 @@
   dispatch:
     CPU, CUDA: __ilshift__
 
+- func: bitwise_left_shift.Tensor(Tensor self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  variants: function, method
+  structured_delegate: bitwise_left_shift.Tensor_out
+
+- func: bitwise_left_shift_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  variants: method
+  structured_delegate: bitwise_left_shift.Tensor_out
+
+- func: bitwise_left_shift.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
+  dispatch:
+    CPU, CUDA: bitwise_left_shift_out
+
+- func: bitwise_left_shift.Tensor_Scalar(Tensor self, Scalar other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  variants: method, function
+  dispatch:
+    CPU, CUDA: bitwise_left_shift
+
+- func: bitwise_left_shift_.Tensor_Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  variants: method
+  dispatch:
+    CPU, CUDA: bitwise_left_shift_
+
+- func: bitwise_left_shift.Tensor_Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  variants: function
+  dispatch:
+    CPU, CUDA: bitwise_left_shift_out
+
+- func: bitwise_left_shift.Scalar_Tensor(Scalar self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  variants: function
+  dispatch:
+    CPU, CUDA: bitwise_left_shift
+
 - func: __rshift__.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
@@ -5668,67 +6022,77 @@
   dispatch:
     CPU, CUDA: __irshift__
 
-- func: tril_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
-  variants: method
-  dispatch:
-    CPU: tril_cpu_
-    CUDA: tril_cuda_
+- func: bitwise_right_shift.Tensor(Tensor self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  variants: function, method
+  structured_delegate: bitwise_right_shift.Tensor_out
 
-- func: triu_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
+- func: bitwise_right_shift_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
   variants: method
+  structured_delegate: bitwise_right_shift.Tensor_out
+
+- func: bitwise_right_shift.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
-    CPU: triu_cpu_
-    CUDA: triu_cuda_
+    CPU, CUDA: bitwise_right_shift_out
 
-- func: digamma_(Tensor(a!) self) -> Tensor(a!)
+- func: bitwise_right_shift.Tensor_Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
-  structured_delegate: digamma.out
-  variants: method
+  variants: method, function
+  dispatch:
+    CPU, CUDA: bitwise_right_shift
 
-- func: renorm_(Tensor(a!) self, Scalar p, int dim, Scalar maxnorm) -> Tensor(a!)
+- func: bitwise_right_shift_.Tensor_Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
-    CPU: legacy::cpu::_th_renorm_
-    CUDA: legacy::cuda::_th_renorm_
+    CPU, CUDA: bitwise_right_shift_
 
-- func: lerp_.Scalar(Tensor(a!) self, Tensor end, Scalar weight) -> Tensor(a!)
+- func: bitwise_right_shift.Tensor_Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
-  variants: method
+  variants: function
   dispatch:
-    CPU: lerp_cpu_scalar_
-    CUDA: lerp_cuda_scalar_
+    CPU, CUDA: bitwise_right_shift_out
 
-- func: lerp_.Tensor(Tensor(a!) self, Tensor end, Tensor weight) -> Tensor(a!)
+- func: bitwise_right_shift.Scalar_Tensor(Scalar self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
+  variants: function
+  dispatch:
+    CPU, CUDA: bitwise_right_shift
+
+- func: tril_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
   variants: method
   dispatch:
-    CPU: lerp_cpu_tensor_
-    CUDA: lerp_cuda_tensor_
+    CPU: tril_cpu_
+    CUDA: tril_cuda_
 
-- func: fmod_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
-  device_check: NoCheck   # TensorIterator
+- func: triu_(Tensor(a!) self, int diagonal=0) -> Tensor(a!)
   variants: method
   dispatch:
-    CPU, CUDA: fmod_
+    CPU: triu_cpu_
+    CUDA: triu_cuda_
 
-- func: fmod_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+- func: digamma_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured_delegate: digamma.out
   variants: method
-  dispatch:
-    CPU, CUDA: fmod_
 
-- func: remainder_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+- func: lerp_.Scalar(Tensor(a!) self, Tensor end, Scalar weight) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
-    CPU, CUDA: remainder_
+    CPU: lerp_cpu_scalar_
+    CUDA: lerp_cuda_scalar_
 
-- func: remainder_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+- func: lerp_.Tensor(Tensor(a!) self, Tensor end, Tensor weight) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
-    CPU, CUDA: remainder_
+    CPU: lerp_cpu_tensor_
+    CUDA: lerp_cuda_tensor_
 
 - func: addbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
   variants: method
@@ -5744,12 +6108,6 @@
   dispatch:
     CPU, CUDA: addbmm
 
-- func: addcdiv_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!)
-  device_check: NoCheck   # TensorIterator
-  variants: method
-  dispatch:
-    CompositeExplicitAutograd: addcdiv_
-
 - func: random_.from(Tensor(a!) self, int from, int? to, *, Generator? generator=None) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   variants: method
@@ -5870,38 +6228,44 @@
   device_guard: False
 
 - func: ne.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: ne_out
+    CPU, CUDA: ne_Scalar_out
     QuantizedCPU: ne_out_quantized_cpu
 
 - func: ne.Scalar(Tensor self, Scalar other) -> Tensor
+  structured_delegate: ne.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: ne
     QuantizedCPU: ne_quantized_cpu
 
 - func: ne.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: ne_out
+    CPU, CUDA: ne_Tensor_out
     QuantizedCPU: ne_out_quantized_cpu
 
 - func: ne.Tensor(Tensor self, Tensor other) -> Tensor
+  structured_delegate: ne.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: ne
     QuantizedCPU: ne_quantized_cpu
 
 - func: ne_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  structured_delegate: ne.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
     CompositeExplicitAutograd: ne_
 
 - func: ne_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  structured_delegate: ne.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
@@ -5925,64 +6289,74 @@
   variants: method
 
 - func: eq.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: eq_out
+    CPU, CUDA: eq_Scalar_out
     QuantizedCPU: eq_out_quantized_cpu
 
 - func: eq.Scalar(Tensor self, Scalar other) -> Tensor
+  structured_delegate: eq.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: eq
     QuantizedCPU: eq_quantized_cpu
 
 - func: eq.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: eq_out
+    CPU, CUDA: eq_Tensor_out
     QuantizedCPU: eq_out_quantized_cpu
 
 - func: eq.Tensor(Tensor self, Tensor other) -> Tensor
+  structured_delegate: eq.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: eq
     QuantizedCPU: eq_quantized_cpu
 
 - func: ge.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: ge_out
+    CPU, CUDA: ge_Scalar_out
     QuantizedCPU: ge_out_quantized_cpu
 
 - func: ge.Scalar(Tensor self, Scalar other) -> Tensor
+  structured_delegate: ge.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: ge
     QuantizedCPU: ge_quantized_cpu
 
 - func: ge.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: ge_out
+    CPU, CUDA: ge_Tensor_out
     QuantizedCPU: ge_out_quantized_cpu
 
 - func: ge.Tensor(Tensor self, Tensor other) -> Tensor
+  structured_delegate: ge.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: ge
     QuantizedCPU: ge_quantized_cpu
 
 - func: ge_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  structured_delegate: ge.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
     CompositeExplicitAutograd: ge_
 
 - func: ge_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  structured_delegate: ge.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
@@ -6006,38 +6380,44 @@
   variants: method
 
 - func: le.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: le_out
+    CPU, CUDA: le_Scalar_out
     QuantizedCPU: le_out_quantized_cpu
 
 - func: le.Scalar(Tensor self, Scalar other) -> Tensor
+  structured_delegate: le.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: le
     QuantizedCPU: le_quantized_cpu
 
 - func: le.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: le_out
+    CPU, CUDA: le_Tensor_out
     QuantizedCPU: le_out_quantized_cpu
 
 - func: le.Tensor(Tensor self, Tensor other) -> Tensor
+  structured_delegate: le.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: le
     QuantizedCPU: le_quantized_cpu
 
 - func: le_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  structured_delegate: le.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
     CompositeExplicitAutograd: le_
 
 - func: le_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  structured_delegate: le.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
@@ -6061,38 +6441,44 @@
   variants: method
 
 - func: gt.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: gt_out
+    CPU, CUDA: gt_Scalar_out
     QuantizedCPU: gt_out_quantized_cpu
 
 - func: gt.Scalar(Tensor self, Scalar other) -> Tensor
+  structured_delegate: gt.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: gt
     QuantizedCPU: gt_quantized_cpu
 
 - func: gt.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: gt_out
+    CPU, CUDA: gt_Tensor_out
     QuantizedCPU: gt_out_quantized_cpu
 
 - func: gt.Tensor(Tensor self, Tensor other) -> Tensor
+  structured_delegate: gt.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: gt
     QuantizedCPU: gt_quantized_cpu
 
 - func: gt_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  structured_delegate: gt.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
     CompositeExplicitAutograd: gt_
 
 - func: gt_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  structured_delegate: gt.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
@@ -6116,38 +6502,44 @@
   variants: method
 
 - func: lt.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: lt_out
+    CPU, CUDA: lt_Scalar_out
     QuantizedCPU: lt_out_quantized_cpu
 
 - func: lt.Scalar(Tensor self, Scalar other) -> Tensor
+  structured_delegate: lt.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: lt
     QuantizedCPU: lt_quantized_cpu
 
 - func: lt.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: lt_out
+    CPU, CUDA: lt_Tensor_out
     QuantizedCPU: lt_out_quantized_cpu
 
 - func: lt.Tensor(Tensor self, Tensor other) -> Tensor
+  structured_delegate: lt.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: lt
     QuantizedCPU: lt_quantized_cpu
 
 - func: lt_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  structured_delegate: lt.Scalar_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
     CompositeExplicitAutograd: lt_
 
 - func: lt_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  structured_delegate: lt.Tensor_out
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
@@ -6186,14 +6578,14 @@
 
 - func: index_select.out(Tensor self, int dim, Tensor index, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: index_select_out_cpu_
-    CUDA: index_select_out_cuda
+    CPU, QuantizedCPU: index_select_out_cpu_
+    CUDA, QuantizedCUDA: index_select_out_cuda
 
 - func: index_select(Tensor self, int dim, Tensor index) -> Tensor
   variants: method, function
   dispatch:
-    CPU: index_select_cpu_
-    CUDA: index_select_cuda
+    CPU, QuantizedCPU: index_select_cpu_
+    CUDA, QuantizedCUDA: index_select_cuda
     SparseCPU: index_select_sparse
     SparseCUDA: index_select_sparse
 
@@ -6225,27 +6617,26 @@
 
 - func: nonzero.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: legacy::cpu::_th_nonzero_out
+    CPU: nonzero_out_cpu
     CUDA: nonzero_out_cuda
 
 - func: nonzero(Tensor self) -> Tensor
   variants: method, function
   dispatch:
-    CPU: legacy::cpu::_th_nonzero
+    CPU: nonzero_cpu
     CUDA: nonzero_cuda
 
 - func: nonzero_numpy(Tensor self) -> Tensor[]
   variants: method, function
 
 - func: gather.out(Tensor self, int dim, Tensor index, *, bool sparse_grad=False, Tensor(a!) out) -> Tensor(a!)
+  structured: True
   dispatch:
-    CPU: gather_out_cpu_cuda
-    CUDA: gather_out_cpu_cuda
+    CPU, CUDA: gather_out
 
 - func: gather(Tensor self, int dim, Tensor index, *, bool sparse_grad=False) -> Tensor
   variants: method, function
-  dispatch:
-    CPU, CUDA: gather
+  structured_delegate: gather.out
 
 - func: gather_backward(Tensor grad, Tensor self, int dim, Tensor index, bool sparse_grad) -> Tensor
   variants: function
@@ -6260,46 +6651,52 @@
 - func: _gather_sparse_backward(Tensor self, int dim, Tensor index, Tensor grad) -> Tensor
 
 - func: addcmul.out(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: addcmul_out
 
 - func: addcmul(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor
+  structured_delegate: addcmul.out
   device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CompositeExplicitAutograd: addcmul
 
 - func: addcmul_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!)
+  structured_delegate: addcmul.out
   device_check: NoCheck   # TensorIterator
   variants: method
-  dispatch:
-    CompositeExplicitAutograd: addcmul_
 
 - func: addcdiv.out(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: addcdiv_out
 
 - func: addcdiv(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor
+  structured_delegate: addcdiv.out
   device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CompositeExplicitAutograd: addcdiv
 
-- func: cross_entropy_loss(Tensor self, Tensor target, Tensor? weight=None, int reduction=Mean, int ignore_index=-100) -> Tensor
+- func: addcdiv_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!)
+  structured_delegate: addcdiv.out
+  device_check: NoCheck   # TensorIterator
+  variants: method
+
+- func: cross_entropy_loss(Tensor self, Tensor target, Tensor? weight=None, int reduction=Mean, int ignore_index=-100, float label_smoothing=0.0) -> Tensor
   python_module: nn
 
 - func: lstsq.X(Tensor self, Tensor A, *, Tensor(a!) X, Tensor(b!) qr) -> (Tensor(a!) solution, Tensor(b!) QR)
   dispatch:
-    CPU: legacy::cpu::_th_gels_out
-    CUDA: legacy::cuda::_th_gels_out
+    CPU: legacy_lstsq_out
+    CUDA: legacy_lstsq_out_cuda
 
 - func: lstsq(Tensor self, Tensor A) -> (Tensor solution, Tensor QR)
   variants: method, function
   dispatch:
-    CPU: legacy::cpu::_th_gels
-    CUDA: legacy::cuda::_th_gels
+    CPU: legacy_lstsq
+    CUDA: legacy_lstsq_cuda
 
 - func: triangular_solve.X(Tensor self, Tensor A, bool upper=True, bool transpose=False, bool unitriangular=False, *, Tensor(a!) X, Tensor(b!) M) -> (Tensor(a!) solution, Tensor(b!) cloned_coefficient)
   dispatch:
@@ -6444,19 +6841,19 @@
   dispatch:
     CPU, CUDA: ormqr
 
-- func: _lu_with_info(Tensor self, bool pivot=True, bool check_errors=True) -> (Tensor, Tensor, Tensor)
+- func: _lu_with_info(Tensor self, bool pivot=True, bool check_errors=True) -> (Tensor LU, Tensor pivots, Tensor info)
   variants: function
   dispatch:
     CPU, CUDA: _lu_with_info
 
 - func: lu_solve.out(Tensor self, Tensor LU_data, Tensor LU_pivots, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CompositeExplicitAutograd: lu_solve_out
+    CPU, CUDA: lu_solve_out
 
 - func: lu_solve(Tensor self, Tensor LU_data, Tensor LU_pivots) -> Tensor
   variants: method, function
   dispatch:
-    CompositeExplicitAutograd: lu_solve
+    CPU, CUDA: lu_solve
 
 - func: lu_unpack(Tensor LU_data, Tensor LU_pivots, bool unpack_data=True, bool unpack_pivots=True) -> (Tensor P, Tensor L, Tensor U)
   variants: function
@@ -6579,8 +6976,11 @@
 
 - func: signbit(Tensor self) -> Tensor
   variants: function, method
+  structured_delegate: signbit.out
 
 - func: signbit.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU: signbit_out
     CUDA: signbit_out
@@ -6636,36 +7036,67 @@
 
 - func: histc.out(Tensor self, int bins=100, Scalar min=0, Scalar max=0, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU: legacy::cpu::_th_histc_out
+    CPU: histogram_histc_cpu_out
     CUDA: _histc_out_cuda
 
 - func: histc(Tensor self, int bins=100, Scalar min=0, Scalar max=0) -> Tensor
   variants: method, function
   dispatch:
-    CPU: legacy::cpu::_th_histc
+    CPU: histogram_histc_cpu
     CUDA: _histc_cuda
 
+- func: histogram.bins_tensor_out(Tensor self, Tensor bins, *, Tensor? weight=None, bool density=False, Tensor(a!) hist, Tensor(b!) bin_edges) -> (Tensor(a!) hist, Tensor(b!) bin_edges)
+  dispatch:
+    CPU: histogram_out_cpu
+
+- func: histogram.bins_tensor(Tensor self, Tensor bins, *, Tensor? weight=None, bool density=False) -> (Tensor hist, Tensor bin_edges)
+  variants: method, function
+  dispatch:
+    CPU: histogram_cpu
+
+- func: histogram.bin_ct_out(Tensor self, int bins=100, *, float[]? range=None, Tensor? weight=None, bool density=False, Tensor(a!) hist, Tensor(b!) bin_edges) -> (Tensor(a!) hist, Tensor(b!) bin_edges)
+  dispatch:
+    CPU: histogram_out_cpu
+
+- func: histogram.bin_ct(Tensor self, int bins=100, *, float[]? range=None, Tensor? weight=None, bool density=False) -> (Tensor hist, Tensor bin_edges)
+  variants: method, function
+  dispatch:
+    CPU: histogram_cpu
+
 - func: fmod.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: fmod_out
+    CompositeExplicitAutograd: fmod_out
 
 - func: fmod.Scalar(Tensor self, Scalar other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: fmod
+    CompositeExplicitAutograd: fmod
+
+- func: fmod_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  variants: method
+  dispatch:
+    CompositeExplicitAutograd: fmod_
 
 - func: fmod.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: fmod_out
 
 - func: fmod.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: fmod.Tensor_out
   variants: method, function
-  dispatch:
-    CPU, CUDA: fmod
+
+
+- func: fmod_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  variants: method
+  structured_delegate: fmod.Tensor_out
 
 - func: hypot.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -6728,24 +7159,39 @@
     CompositeExplicitAutograd: nextafter_
 
 - func: remainder.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
-  device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: remainder_out
+    CompositeExplicitAutograd: remainder_out
 
 - func: remainder.Scalar(Tensor self, Scalar other) -> Tensor
-  device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: remainder
+    CompositeExplicitAutograd: remainder
+
+- func: remainder_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  variants: method
+  dispatch:
+    CompositeExplicitAutograd: remainder_
 
 - func: remainder.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: remainder_out
 
 - func: remainder.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: remainder.Tensor_out
   variants: method, function
+
+- func: remainder_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  structured_delegate: remainder.Tensor_out
+  variants: method
+
+- func: remainder.Scalar_Tensor(Scalar self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  variants: function
   dispatch:
     CPU, CUDA: remainder
 
@@ -6757,11 +7203,14 @@
     QuantizedCPU: min_quantized_cpu
 
 - func: fmin(Tensor self, Tensor other) -> Tensor
+  structured_delegate: fmin.out
+  device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CPU, CUDA: fmin
 
 - func: fmin.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: fmin_out
 
@@ -6773,11 +7222,14 @@
     QuantizedCPU: max_quantized_cpu
 
 - func: fmax(Tensor self, Tensor other) -> Tensor
+  structured_delegate: fmax.out
+  device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CPU, CUDA: fmax
 
 - func: fmax.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: fmax_out
 
@@ -6928,29 +7380,43 @@
 
 - func: all(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: all.all_out
   variants: method, function
+
+- func: all.all_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck
+  structured: True
   dispatch:
-    CPU, CUDA: all
+    CPU, CUDA: all_all_out
 
 - func: any(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
+  structured_delegate: any.all_out
   variants: method, function
   dispatch:
-    CPU, CUDA: any
     SparseCPU, SparseCUDA: any_sparse
 
+- func: any.all_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck
+  structured: True
+  dispatch:
+    CPU, CUDA: any_all_out
+
 - func: renorm.out(Tensor self, Scalar p, int dim, Scalar maxnorm, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  structured: True
   dispatch:
-    CPU: legacy::cpu::_th_renorm_out
-    CUDA: legacy::cuda::_th_renorm_out
+    CPU, CUDA: renorm_out
 
 - func: renorm(Tensor self, Scalar p, int dim, Scalar maxnorm) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
-  dispatch:
-    CPU: legacy::cpu::_th_renorm
-    CUDA: legacy::cuda::_th_renorm
+  structured_delegate: renorm.out
+
+- func: renorm_(Tensor(a!) self, Scalar p, int dim, Scalar maxnorm) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  variants: method
+  structured_delegate: renorm.out
 
 - func: unfold(Tensor(a) self, int dimension, int size, int step) -> Tensor(a)
   variants: method
@@ -7084,26 +7550,6 @@
     CPU: _index_copy_impl_
     CUDA: _index_copy_impl_
 
-- func: _cumsum(Tensor self, int dim) -> Tensor
-  dispatch:
-    CPU: _cumsum_cpu
-    CUDA: _cumsum_cuda
-
-- func: _cumsum.out(Tensor self, int dim, *, Tensor(a!) out) -> Tensor(a!)
-  dispatch:
-    CPU: _cumsum_out_cpu
-    CUDA: _cumsum_out_cuda
-
-- func: _cumprod(Tensor self, int dim) -> Tensor
-  dispatch:
-    CPU: _cumprod_cpu
-    CUDA: _cumprod_cuda
-
-- func: _cumprod.out(Tensor self, int dim, *, Tensor(a!) out) -> Tensor(a!)
-  dispatch:
-    CPU: _cumprod_out_cpu
-    CUDA: _cumprod_out_cuda
-
 - func: _amp_foreach_non_finite_check_and_unscale_(Tensor(a!)[] self, Tensor(b!) found_inf, Tensor inv_scale) -> ()
   variants: function
   dispatch:
@@ -7793,6 +8239,15 @@
     CPU: searchsorted_cpu
     CUDA: searchsorted_cuda
 
+- func: _convert_indices_from_coo_to_csr(Tensor self, int size, *, bool out_int32=False) -> Tensor
+  structured_delegate: _convert_indices_from_coo_to_csr.out
+
+- func: _convert_indices_from_coo_to_csr.out(Tensor self, int size, *, bool out_int32=False, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  dispatch:
+    CPU: _convert_indices_from_coo_to_csr_structured_cpu
+    CUDA: _convert_indices_from_coo_to_csr_structured_cuda
+
 ## NN wrappers
 
 - func: mse_loss.out(Tensor self, Tensor target, int reduction=Mean, *, Tensor(a!) out) -> Tensor(a!)
@@ -7841,25 +8296,25 @@
   python_module: nn
   dispatch:
     CPU: multi_margin_loss_cpu_out
-    CUDA: legacy::cuda::_thnn_multi_margin_loss_forward_out
+    CUDA: multi_margin_loss_cuda_out
 
 - func: multi_margin_loss(Tensor self, Tensor target, Scalar p=1, Scalar margin=1, Tensor? weight=None, int reduction=Mean) -> Tensor
   python_module: nn
   dispatch:
     CPU: multi_margin_loss_cpu
-    CUDA: legacy::cuda::_thnn_multi_margin_loss_forward
+    CUDA: multi_margin_loss_cuda
 
 - func: multi_margin_loss_backward.grad_input(Tensor grad_output, Tensor self, Tensor target, Scalar p, Scalar margin, Tensor? weight=None, int reduction=Mean, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   dispatch:
     CPU: multi_margin_loss_cpu_backward_out
-    CUDA: legacy::cuda::_thnn_multi_margin_loss_backward_out
+    CUDA: multi_margin_loss_cuda_backward_out
 
 - func: multi_margin_loss_backward(Tensor grad_output, Tensor self, Tensor target, Scalar p, Scalar margin, Tensor? weight=None, int reduction=Mean) -> Tensor
   python_module: nn
   dispatch:
     CPU: multi_margin_loss_cpu_backward
-    CUDA: legacy::cuda::_thnn_multi_margin_loss_backward
+    CUDA: multi_margin_loss_cuda_backward
 
 - func: multilabel_margin_loss.out(Tensor self, Tensor target, int reduction=Mean, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
@@ -7871,25 +8326,25 @@
   python_module: nn
   dispatch:
     CPU: multilabel_margin_loss_forward_out_cpu
-    CUDA: legacy::cuda::_thnn_multilabel_margin_loss_forward_out
+    CUDA: multilabel_margin_loss_forward_out_cuda
 
 - func: multilabel_margin_loss_forward(Tensor self, Tensor target, int reduction) -> (Tensor output, Tensor is_target)
   python_module: nn
   dispatch:
     CPU: multilabel_margin_loss_forward_cpu
-    CUDA: legacy::cuda::_thnn_multilabel_margin_loss_forward
+    CUDA: multilabel_margin_loss_forward_cuda
 
 - func: multilabel_margin_loss_backward.grad_input(Tensor grad_output, Tensor self, Tensor target, int reduction, Tensor is_target, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   dispatch:
     CPU: multilabel_margin_loss_backward_cpu_out
-    CUDA: legacy::cuda::_thnn_multilabel_margin_loss_backward_out
+    CUDA: multilabel_margin_loss_backward_cuda_out
 
 - func: multilabel_margin_loss_backward(Tensor grad_output, Tensor self, Tensor target, int reduction, Tensor is_target) -> Tensor
   python_module: nn
   dispatch:
     CPU: multilabel_margin_loss_backward_cpu
-    CUDA: legacy::cuda::_thnn_multilabel_margin_loss_backward
+    CUDA: multilabel_margin_loss_backward_cuda
 
 - func: nll_loss.out(Tensor self, Tensor target, Tensor? weight=None, int reduction=Mean, int ignore_index=-100, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
@@ -7902,27 +8357,25 @@
 
 - func: nll_loss_forward.output(Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index, *, Tensor(a!) output, Tensor(b!) total_weight) -> (Tensor(a!), Tensor(b!))
   python_module: nn
+  structured: True
   dispatch:
     CPU: nll_loss_forward_out_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss_forward_out
+    CUDA: nll_loss_forward_out_cuda
 
 - func: nll_loss_forward(Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index) -> (Tensor output, Tensor total_weight)
   python_module: nn
-  dispatch:
-    CPU: nll_loss_forward_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss_forward
+  structured_delegate: nll_loss_forward.output
 
 - func: nll_loss_backward.grad_input(Tensor grad_output, Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index, Tensor total_weight, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
     CPU: nll_loss_backward_out_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss_backward_out
+    CUDA: nll_loss_backward_out_cuda
 
 - func: nll_loss_backward(Tensor grad_output, Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index, Tensor total_weight) -> Tensor
   python_module: nn
-  dispatch:
-    CPU: nll_loss_backward_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss_backward
+  structured_delegate: nll_loss_backward.grad_input
 
 - func: nll_loss2d.out(Tensor self, Tensor target, Tensor? weight=None, int reduction=Mean, int ignore_index=-100, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
@@ -7934,25 +8387,25 @@
   python_module: nn
   dispatch:
     CPU: nll_loss2d_forward_out_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss2d_forward_out
+    CUDA: nll_loss2d_forward_out_cuda
 
 - func: nll_loss2d_forward(Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index) -> (Tensor output, Tensor total_weight)
   python_module: nn
   dispatch:
     CPU: nll_loss2d_forward_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss2d_forward
+    CUDA: nll_loss2d_forward_cuda
 
 - func: nll_loss2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index, Tensor total_weight, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   dispatch:
     CPU: nll_loss2d_backward_out_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss2d_backward_out
+    CUDA: nll_loss2d_backward_out_cuda
 
 - func: nll_loss2d_backward(Tensor grad_output, Tensor self, Tensor target, Tensor? weight, int reduction, int ignore_index, Tensor total_weight) -> Tensor
   python_module: nn
   dispatch:
     CPU: nll_loss2d_backward_cpu
-    CUDA: legacy::cuda::_thnn_nll_loss2d_backward
+    CUDA: nll_loss2d_backward_cuda
 
 - func: smooth_l1_loss.out(Tensor self, Tensor target, int reduction=Mean, float beta=1.0, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -8031,10 +8484,16 @@
   device_check: NoCheck   # TensorIterator
   python_module: nn
 
-- func: elu_backward(Tensor grad_output, Scalar alpha, Scalar scale, Scalar input_scale, bool is_result, Tensor self_or_result) -> Tensor
+- func: elu_backward.grad_input(Tensor grad_output, Scalar alpha, Scalar scale, Scalar input_scale, bool is_result, Tensor self_or_result, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA: elu_backward
+    CPU, CUDA: elu_backward_out
+
+- func: elu_backward(Tensor grad_output, Scalar alpha, Scalar scale, Scalar input_scale, bool is_result, Tensor self_or_result) -> Tensor
+  structured_delegate: elu_backward.grad_input
+  python_module: nn
 
 - func: elu_(Tensor(a!) self, Scalar alpha=1, Scalar scale=1, Scalar input_scale=1) -> Tensor(a!)
   structured_delegate: elu.out
@@ -8044,28 +8503,28 @@
     CompositeExplicitAutograd: elu_
 
 - func: glu.out(Tensor self, int dim=-1, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU: glu_out
-    CUDA: legacy::cuda::_thnn_glu_forward_out
+    CPU, CUDA: glu_out
 
 - func: glu(Tensor self, int dim=-1) -> Tensor
+  structured_delegate: glu.out
+  device_check: NoCheck   # TensorIterator
   python_module: nn
-  dispatch:
-    CPU: glu
-    CUDA: legacy::cuda::_thnn_glu_forward
 
 - func: glu_backward.grad_input(Tensor grad_output, Tensor self, int dim, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   dispatch:
-    CPU: glu_backward_out
-    CUDA: legacy::cuda::_thnn_glu_backward_out
+    CPU: glu_backward_cpu_out
+    CUDA: glu_backward_cuda_out
 
 - func: glu_backward(Tensor grad_output, Tensor self, int dim) -> Tensor
   python_module: nn
   dispatch:
-    CPU: glu_backward
-    CUDA: legacy::cuda::_thnn_glu_backward
+    CPU: glu_backward_cpu
+    CUDA: glu_backward_cuda
 
 - func: hardsigmoid.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -8087,10 +8546,16 @@
   device_check: NoCheck   # TensorIterator
   python_module: nn
 
-- func: hardsigmoid_backward(Tensor grad_output, Tensor self) -> Tensor
+- func: hardsigmoid_backward.grad_input(Tensor grad_output, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA: hardsigmoid_backward
+    CPU, CUDA: hardsigmoid_backward_out
+
+- func: hardsigmoid_backward(Tensor grad_output, Tensor self) -> Tensor
+  structured_delegate: hardsigmoid_backward.grad_input
+  python_module: nn
 
 - func: hardtanh.out(Tensor self, Scalar min_val=-1, Scalar max_val=1, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -8162,10 +8627,16 @@
   dispatch:
     QuantizedCPU: leaky_relu_quantized_cpu
 
-- func: leaky_relu_backward(Tensor grad_output, Tensor self, Scalar negative_slope, bool self_is_result) -> Tensor
+- func: leaky_relu_backward.grad_input(Tensor grad_output, Tensor self, Scalar negative_slope, bool self_is_result, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA: leaky_relu_backward
+    CPU, CUDA: leaky_relu_backward_out
+
+- func: leaky_relu_backward(Tensor grad_output, Tensor self, Scalar negative_slope, bool self_is_result) -> Tensor
+  structured_delegate: leaky_relu_backward.grad_input
+  python_module: nn
 
 - func: leaky_relu_(Tensor(a!) self, Scalar negative_slope=0.01) -> Tensor(a!)
   structured_delegate: leaky_relu.out
@@ -8187,38 +8658,38 @@
   python_module: nn
   dispatch:
     CPU: log_sigmoid_forward_out_cpu
-    CUDA: legacy::cuda::_thnn_log_sigmoid_forward_out
+    CUDA: log_sigmoid_forward_out_cuda
 
 - func: log_sigmoid_forward(Tensor self) -> (Tensor output, Tensor buffer)
   device_check: NoCheck   # TensorIterator
   python_module: nn
   dispatch:
     CPU: log_sigmoid_forward_cpu
-    CUDA: legacy::cuda::_thnn_log_sigmoid_forward
+    CUDA: log_sigmoid_forward_cuda
 
 - func: log_sigmoid_backward.grad_input(Tensor grad_output, Tensor self, Tensor buffer, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
   dispatch:
-    CPU: log_sigmoid_backward_out_cpu
-    CUDA: legacy::cuda::_thnn_log_sigmoid_backward_out
+    CPU: log_sigmoid_backward_cpu_out
+    CUDA: log_sigmoid_backward_cuda_out
 
 - func: log_sigmoid_backward(Tensor grad_output, Tensor self, Tensor buffer) -> Tensor
   python_module: nn
   dispatch:
     CPU: log_sigmoid_backward_cpu
-    CUDA: legacy::cuda::_thnn_log_sigmoid_backward
+    CUDA: log_sigmoid_backward_cuda
 
 - func: rrelu_with_noise.out(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   dispatch:
     CPU: rrelu_with_noise_out_cpu
-    CUDA: legacy::cuda::_thnn_rrelu_with_noise_forward_out
+    CUDA: rrelu_with_noise_out_cuda
 
 - func: rrelu_with_noise(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
   python_module: nn
   dispatch:
     CPU: rrelu_with_noise_cpu
-    CUDA: legacy::cuda::_thnn_rrelu_with_noise_forward
+    CUDA: rrelu_with_noise_cuda
 
 - func: rrelu_with_noise_backward(Tensor grad_output, Tensor self, Tensor noise, Scalar lower, Scalar upper, bool training, bool self_is_result) -> Tensor
   python_module: nn
@@ -8229,7 +8700,7 @@
   python_module: nn
   dispatch:
     CPU: rrelu_with_noise_cpu_
-    CUDA: legacy::cuda::_thnn_rrelu_with_noise_forward_
+    CUDA: rrelu_with_noise_cuda_
 
 - func: softplus.out(Tensor self, Scalar beta=1, Scalar threshold=20, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -8245,14 +8716,15 @@
   python_module: nn
 
 - func: softplus_backward.grad_input(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold, Tensor output, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
     CPU, CUDA: softplus_backward_out
 
 - func: softplus_backward(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold, Tensor output) -> Tensor
+  structured_delegate: softplus_backward.grad_input
   python_module: nn
-  dispatch:
-    CPU, CUDA: softplus_backward
 
 - func: softshrink.out(Tensor self, Scalar lambd=0.5, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -8268,19 +8740,21 @@
   python_module: nn
 
 - func: softshrink_backward.grad_input(Tensor grad_output, Tensor self, Scalar lambd, *, Tensor(a!) grad_input) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
     CPU, CUDA: softshrink_backward_out
 
 - func: softshrink_backward(Tensor grad_output, Tensor self, Scalar lambd) -> Tensor
+  structured_delegate: softshrink_backward.grad_input
   python_module: nn
-  dispatch:
-    CPU, CUDA: softshrink_backward
 
 - func: adaptive_avg_pool2d.out(Tensor self, int[2] output_size, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   dispatch:
-    CPU, CUDA: adaptive_avg_pool2d_out_cpu
+    CPU: adaptive_avg_pool2d_out_cpu
+    CUDA: adaptive_avg_pool2d_out_cuda
     MkldnnCPU: mkldnn_adaptive_avg_pool2d_out
 
 - func: adaptive_avg_pool2d(Tensor self, int[2] output_size) -> Tensor
@@ -8384,6 +8858,11 @@
 
 - func: avg_pool2d.out(Tensor self, int[2] kernel_size, int[2] stride=[], int[2] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
+  structured: True
+  precomputed:
+  - kernel_size -> int kH, int kW
+  - stride -> int dH, int dW
+  - padding -> int padH, int padW
   dispatch:
     CPU: avg_pool2d_out_cpu
     CUDA: avg_pool2d_out_cuda
@@ -8391,14 +8870,14 @@
 
 - func: avg_pool2d(Tensor self, int[2] kernel_size, int[2] stride=[], int[2] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None) -> Tensor
   python_module: nn
+  structured_delegate: avg_pool2d.out
   dispatch:
-    CPU: avg_pool2d_cpu
-    CUDA: avg_pool2d_cuda
     MkldnnCPU: mkldnn_avg_pool2d
     QuantizedCPU: avg_pool2d_quantized_cpu
 
 - func: avg_pool2d_backward.grad_input(Tensor grad_output, Tensor self, int[2] kernel_size, int[2] stride, int[2] padding, bool ceil_mode, bool count_include_pad, int? divisor_override, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
     CPU: avg_pool2d_backward_out_cpu
     CUDA: avg_pool2d_backward_out_cuda
@@ -8406,13 +8885,13 @@
 
 - func: avg_pool2d_backward(Tensor grad_output, Tensor self, int[2] kernel_size, int[2] stride, int[2] padding, bool ceil_mode, bool count_include_pad, int? divisor_override) -> Tensor
   python_module: nn
+  structured_delegate: avg_pool2d_backward.grad_input
   dispatch:
-    CPU: avg_pool2d_backward_cpu
-    CUDA: avg_pool2d_backward_cuda
     MkldnnCPU: mkldnn_avg_pool2d_backward
 
 - func: avg_pool3d.out(Tensor self, int[3] kernel_size, int[3] stride=[], int[3] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
     CPU: avg_pool3d_out_cpu
     CUDA: avg_pool3d_out_cuda
@@ -8420,14 +8899,14 @@
 
 - func: avg_pool3d(Tensor self, int[3] kernel_size, int[3] stride=[], int[3] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None) -> Tensor
   python_module: nn
+  structured_delegate: avg_pool3d.out
   dispatch:
-    CPU: avg_pool3d_cpu
-    CUDA: avg_pool3d_cuda
     MkldnnCPU: mkldnn_avg_pool3d
     QuantizedCPU: avg_pool3d_quantized_cpu
 
 - func: avg_pool3d_backward.grad_input(Tensor grad_output, Tensor self, int[3] kernel_size, int[3] stride, int[3] padding, bool ceil_mode, bool count_include_pad, int? divisor_override, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
     CPU: avg_pool3d_backward_out_cpu
     CUDA: avg_pool3d_backward_out_cuda
@@ -8435,9 +8914,8 @@
 
 - func: avg_pool3d_backward(Tensor grad_output, Tensor self, int[3] kernel_size, int[3] stride, int[3] padding, bool ceil_mode, bool count_include_pad, int? divisor_override) -> Tensor
   python_module: nn
+  structured_delegate: avg_pool3d_backward.grad_input
   dispatch:
-    CPU: avg_pool3d_backward_cpu
-    CUDA: avg_pool3d_backward_cuda
     MkldnnCPU: mkldnn_avg_pool3d_backward
 
 # Return: (Tensor output, Tensor indices)
@@ -8604,15 +9082,14 @@
 
 - func: reflection_pad1d_backward.grad_input(Tensor grad_output, Tensor self, int[2] padding, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
     CPU: reflection_pad1d_backward_out_cpu
     CUDA: reflection_pad1d_backward_out_cuda
 
 - func: reflection_pad1d_backward(Tensor grad_output, Tensor self, int[2] padding) -> Tensor
   python_module: nn
-  dispatch:
-    CPU: reflection_pad1d_backward_cpu
-    CUDA: reflection_pad1d_backward_cuda
+  structured_delegate: reflection_pad1d_backward.grad_input
 
 - func: reflection_pad2d.out(Tensor self, int[4] padding, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
@@ -8638,6 +9115,28 @@
     CPU: reflection_pad2d_backward_cpu
     CUDA: reflection_pad2d_backward_cuda
 
+- func: reflection_pad3d.out(Tensor self, int[6] padding, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: reflection_pad3d_out_cpu
+    CUDA: reflection_pad3d_out_cuda
+
+- func: reflection_pad3d(Tensor self, int[6] padding) -> Tensor
+  python_module: nn
+  structured_delegate: reflection_pad3d.out
+
+- func: reflection_pad3d_backward.grad_input(Tensor grad_output, Tensor self, int[6] padding, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: reflection_pad3d_backward_out_cpu
+    CUDA: reflection_pad3d_backward_out_cuda
+
+- func: reflection_pad3d_backward(Tensor grad_output, Tensor self, int[6] padding) -> Tensor
+  python_module: nn
+  structured_delegate: reflection_pad3d_backward.grad_input
+
 - func: replication_pad1d.out(Tensor self, int[2] padding, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -8942,33 +9441,36 @@
 
 - func: sigmoid_backward.grad_input(Tensor grad_output, Tensor output, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: sigmoid_backward_out
 
 - func: sigmoid_backward(Tensor grad_output, Tensor output) -> Tensor
   python_module: nn
-  dispatch:
-    CPU, CUDA: sigmoid_backward
+  structured_delegate: sigmoid_backward.grad_input
 
 - func: logit_backward.grad_input(Tensor grad_output, Tensor self, float? eps=None, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: logit_backward_out
 
 - func: logit_backward(Tensor grad_output, Tensor self, float? eps=None) -> Tensor
   python_module: nn
-  dispatch:
-    CPU, CUDA: logit_backward
+  structured_delegate: logit_backward.grad_input
 
 - func: tanh_backward.grad_input(Tensor grad_output, Tensor output, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: tanh_backward_out
 
 - func: tanh_backward(Tensor grad_output, Tensor output) -> Tensor
   python_module: nn
-  dispatch:
-    CPU, CUDA: tanh_backward
+  structured_delegate: tanh_backward.grad_input
 
 # What's a thnn_conv_ versus a slow_conv_?
 #
@@ -8990,15 +9492,14 @@
 
 - func: slow_conv_transpose2d.out(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, int[2] output_padding=0, int[2] dilation=1, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
+  structured: True
   dispatch:
-    CPU: slow_conv_transpose2d_out_cpu
-    CUDA: slow_conv_transpose2d_out_cuda
+    CPU: slow_conv_transpose2d_structured_cpu
+    CUDA: slow_conv_transpose2d_structured_cuda
 
 - func: slow_conv_transpose2d(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, int[2] output_padding=0, int[2] dilation=1) -> Tensor
   python_module: nn
-  dispatch:
-    CPU: slow_conv_transpose2d_cpu
-    CUDA: slow_conv_transpose2d_cuda
+  structured_delegate: slow_conv_transpose2d.out
 
 - func: slow_conv_transpose2d_backward.grad_output(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] output_padding, int[2] dilation, Tensor columns, Tensor ones, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
   python_module: nn
@@ -9046,13 +9547,13 @@
   python_module: nn
   dispatch:
     CPU: slow_conv2d_forward_out_cpu
-    CUDA: legacy::cuda::_thnn_conv2d_forward_out
+    CUDA: slow_conv2d_forward_out_cuda
 
 - func: thnn_conv2d_forward(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding) -> (Tensor output, Tensor finput, Tensor fgrad_input)
   python_module: nn
   dispatch:
     CPU: slow_conv2d_forward_cpu
-    CUDA: legacy::cuda::_thnn_conv2d_forward
+    CUDA: slow_conv2d_forward_cuda
 
 - func: thnn_conv2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, Tensor finput, Tensor fgrad_input, *, Tensor(a!) grad_input, Tensor(b!) grad_weight, Tensor(c!) grad_bias) -> (Tensor(a!), Tensor(b!), Tensor(c!))
   python_module: nn
@@ -9066,31 +9567,26 @@
     CPU: slow_conv2d_backward_cpu
     CUDA: slow_conv2d_backward_cuda
 
-- func: thnn_conv_depthwise2d.out(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, int[2] dilation=1, *, Tensor(a!) out) -> Tensor(a!)
-  python_module: nn
-
-- func: thnn_conv_depthwise2d(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, int[2] dilation=1) -> Tensor
-  python_module: nn
-
-- func: thnn_conv_depthwise2d_forward.out(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, int[2] dilation, *, Tensor(a!) out) -> Tensor(a!)
+- func: _conv_depthwise2d.out(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, int[2] dilation, *, Tensor(a!) out) -> Tensor(a!)
+  use_const_ref_for_mutable_tensors: True
   python_module: nn
   dispatch:
-    CUDA: legacy::cuda::_thnn_conv_depthwise2d_forward_out
+    CUDA: conv_depthwise2d_cuda_out
 
-- func: thnn_conv_depthwise2d_forward(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, int[2] dilation) -> Tensor
+- func: _conv_depthwise2d(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, int[2] dilation) -> Tensor
   python_module: nn
   dispatch:
-    CUDA: legacy::cuda::_thnn_conv_depthwise2d_forward
+    CUDA: conv_depthwise2d_cuda
 
-- func: thnn_conv_depthwise2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, *, Tensor(a!) grad_input, Tensor(b!) grad_weight) -> (Tensor(a!), Tensor(b!))
+- func: _conv_depthwise2d_backward.grad_input(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, *, Tensor(a!) grad_input, Tensor(b!) grad_weight) -> (Tensor(a!), Tensor(b!))
   python_module: nn
   dispatch:
-    CUDA: thnn_conv_depthwise2d_backward_out
+    CUDA: conv_depthwise2d_backward_cuda_out
 
-- func: thnn_conv_depthwise2d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, bool[2] output_mask) -> (Tensor grad_input, Tensor grad_weight)
+- func: _conv_depthwise2d_backward.output_mask(Tensor grad_output, Tensor self, Tensor weight, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, bool[2] output_mask) -> (Tensor grad_input, Tensor grad_weight)
   python_module: nn
   dispatch:
-    CUDA: thnn_conv_depthwise2d_backward
+    CUDA: conv_depthwise2d_backward_cuda
 
 - func: conv_depthwise3d(Tensor self, Tensor weight, int[3] kernel_size, Tensor? bias, int[3] stride, int[3] padding, int[3] dilation) -> Tensor
   python_module: nn
@@ -9226,15 +9722,21 @@
 
 - func: isposinf(Tensor self) -> Tensor
   variants: function, method
+  structured_delegate: isposinf.out
 
 - func: isposinf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: isposinf_out
 
 - func: isneginf(Tensor self) -> Tensor
   variants: function, method
+  structured_delegate: isneginf.out
 
 - func: isneginf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA: isneginf_out
 
@@ -9269,6 +9771,19 @@
   dispatch:
     CPU, CUDA: special_entr_out
 
+- func: special_ndtri(Tensor self) -> Tensor
+  structured_delegate: special_ndtri.out
+  python_module: special
+  variants: function
+
+- func: special_ndtri.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  structured_inherits: TensorIteratorBase
+  python_module: special
+  variants: function
+  dispatch:
+    CPU, CUDA: special_ndtri_out
+
 - func: special_expm1(Tensor self) -> Tensor
   python_module: special
   variants: function
@@ -9285,6 +9800,22 @@
   python_module: special
   variants: function
 
+- func: special_psi(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_psi.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
+- func: special_digamma(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_digamma.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
 - func: special_gammaln(Tensor self) -> Tensor
   python_module: special
   variants: function
@@ -9308,6 +9839,18 @@
 - func: special_erfc.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   python_module: special
 
+- func: special_erfcx(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+  structured_delegate: special_erfcx.out
+
+- func: special_erfcx.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  structured: True
+  structured_inherits: TensorIteratorBase
+  dispatch:
+    CPU, CUDA: special_erfcx_out
+
 - func: special_erfinv(Tensor self) -> Tensor
   python_module: special
   variants: function
@@ -9315,6 +9858,14 @@
 - func: special_erfinv.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   python_module: special
 
+- func: special_ndtr(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_ndtr.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
 - func: special_xlog1py(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   python_module: special
@@ -9358,6 +9909,89 @@
   dispatch:
     CompositeExplicitAutograd: special_xlog1py_out
 
+- func: special_xlogy(Tensor self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+
+- func: special_xlogy.self_scalar(Scalar self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+
+- func: special_xlogy.other_scalar(Tensor self, Scalar other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+
+- func: special_xlogy.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+
+- func: special_xlogy.self_scalar_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+
+- func: special_xlogy.other_scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+
+- func: special_zeta(Tensor self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+  structured_delegate: special_zeta.out
+  dispatch:
+    CompositeExplicitAutograd: special_zeta
+
+- func: special_zeta.self_scalar(Scalar self, Tensor other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: special_zeta
+
+- func: special_zeta.other_scalar(Tensor self, Scalar other) -> Tensor
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: special_zeta
+
+- func: special_zeta.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  structured: True
+  structured_inherits: TensorIteratorBase
+  python_module: special
+  variants: function
+  dispatch:
+    CPU, CUDA: special_zeta_out
+
+- func: special_zeta.self_scalar_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: special_zeta_out
+
+- func: special_zeta.other_scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  python_module: special
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: special_zeta_out
+
+- func: special_i0(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_i0.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
 - func: special_i0e(Tensor self) -> Tensor
   python_module: special
   variants: function
@@ -9370,6 +10004,30 @@
   dispatch:
     CPU, CUDA: special_i0e_out
 
+- func: special_i1(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+  structured_delegate: special_i1.out
+
+- func: special_i1.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  structured: True
+  structured_inherits: TensorIteratorBase
+  dispatch:
+    CPU, CUDA: special_i1_out
+
+- func: special_i1e(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+  structured_delegate: special_i1e.out
+
+- func: special_i1e.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  structured: True
+  structured_inherits: TensorIteratorBase
+  dispatch:
+    CPU, CUDA: special_i1e_out
+
 - func: special_logit(Tensor self, float? eps=None) -> Tensor
   python_module: special
   variants: function
@@ -9377,6 +10035,20 @@
 - func: special_logit.out(Tensor self, float? eps=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: special
 
+- func: special_polygamma(int n, Tensor self) -> Tensor
+  python_module: special
+  variants: function, method
+
+- func: special_polygamma.out(int n, Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+
+- func: special_logsumexp(Tensor self, int[1] dim, bool keepdim=False) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_logsumexp.out(Tensor self, int[1] dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+
 - func: special_expit(Tensor self) -> Tensor
   python_module: special
   variants: function
@@ -9385,6 +10057,58 @@
   python_module: special
   variants: function
 
+- func: special_sinc(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_sinc.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
+- func: special_round(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_round.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
+- func: special_log1p(Tensor self) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_log1p.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
+- func: special_log_softmax(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_gammainc.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
+- func: special_gammainc(Tensor self, Tensor other) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_gammaincc.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
+- func: special_gammaincc(Tensor self, Tensor other) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_multigammaln(Tensor self, int p) -> Tensor
+  python_module: special
+  variants: function
+
+- func: special_multigammaln.out(Tensor self, int p, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: special
+  variants: function
+
 ## Functions related to the fast Fourier transform and the torch.fft namespace
 # Note [FFT namespace binding]
 # Functions in the fft python module should have their names start with
@@ -9542,41 +10266,47 @@
 # See linalg_det as an example.
 
 # "_ex" stands for experimental
-- func: linalg_cholesky_ex(Tensor self, *, bool check_errors=False) -> (Tensor L, Tensor info)
+- func: linalg_cholesky_ex(Tensor self, *, bool upper=False, bool check_errors=False) -> (Tensor L, Tensor info)
   python_module: linalg
   variants: function
   dispatch:
     CPU, CUDA: linalg_cholesky_ex
 
-- func: linalg_cholesky_ex.L(Tensor self, *, bool check_errors=False, Tensor(a!) L, Tensor(b!) info) -> (Tensor(a!) L, Tensor(b!) info)
+- func: linalg_cholesky_ex.L(Tensor self, *, bool upper=False, bool check_errors=False, Tensor(a!) L, Tensor(b!) info) -> (Tensor(a!) L, Tensor(b!) info)
   python_module: linalg
   variants: function
   dispatch:
     CPU, CUDA: linalg_cholesky_ex_out
 
-- func: linalg_cholesky(Tensor self) -> Tensor
+- func: linalg_cholesky(Tensor self, *, bool upper=False) -> Tensor
   python_module: linalg
   variants: function
 
-- func: linalg_cholesky.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+- func: linalg_cholesky.out(Tensor self, *, bool upper=False, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
   variants: function
 
 - func: linalg_det(Tensor self) -> Tensor
   python_module: linalg
   variants: function
-  dispatch:
-    CompositeExplicitAutograd: linalg_det
 
 - func: linalg_det.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
-  dispatch:
-    CompositeExplicitAutograd: linalg_det_out
 
 # torch.det, alias for torch.linalg.det
 - func: det(Tensor self) -> Tensor
   variants: function, method
 
+- func: _det_lu_based_helper(Tensor self) -> (Tensor det, Tensor lu, Tensor pivs)
+  variants: function
+  dispatch:
+    CPU, CUDA: _det_lu_based_helper
+
+- func: _det_lu_based_helper_backward_helper(Tensor det_grad, Tensor det, Tensor self, Tensor lu, Tensor pivs) -> Tensor
+  variants: function
+  dispatch:
+    CPU, CUDA: _det_lu_based_helper_backward_helper
+
 - func: linalg_lstsq(Tensor self, Tensor b, float? rcond=None, *, str? driver=None) -> (Tensor solution, Tensor residuals, Tensor rank, Tensor singular_values)
   python_module: linalg
   variants: function
@@ -9589,6 +10319,14 @@
   dispatch:
     CPU, CUDA: linalg_lstsq_out
 
+# torch.linalg.matmul, alias for torch.matmul
+- func: linalg_matmul(Tensor self, Tensor other) -> Tensor
+  python_module: linalg
+  variants: function
+
+- func: linalg_matmul.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: linalg
+
 - func: linalg_slogdet(Tensor self) -> (Tensor sign, Tensor logabsdet)
   python_module: linalg
   variants: function
@@ -9621,12 +10359,12 @@
   python_module: linalg
   variants: function
   dispatch:
-    CompositeExplicitAutograd: linalg_eigh
+    CPU, CUDA: linalg_eigh
 
 - func: linalg_eigh.eigvals(Tensor self, str UPLO="L", *, Tensor(a!) eigvals, Tensor(b!) eigvecs) -> (Tensor(a!) eigenvalues, Tensor(b!) eigenvectors)
   python_module: linalg
   dispatch:
-    CompositeExplicitAutograd: linalg_eigh_out
+    CPU, CUDA: linalg_eigh_out
 
 - func: linalg_eigvalsh(Tensor self, str UPLO="L") -> Tensor
   python_module: linalg
@@ -9634,6 +10372,8 @@
 
 - func: linalg_eigvalsh.out(Tensor self, str UPLO='L', *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
+  dispatch:
+    CPU, CUDA: linalg_eigvalsh_out
 
 - func: linalg_householder_product(Tensor input, Tensor tau) -> Tensor
   python_module: linalg
@@ -9677,20 +10417,16 @@
 
 - func: inner.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
 
-# torch.outer, alias for torch.ger
 - func: outer(Tensor self, Tensor vec2) -> Tensor
   variants: function, method
 
 - func: outer.out(Tensor self, Tensor vec2, *, Tensor(a!) out) -> Tensor(a!)
 
+# torch.ger, alias for torch.outer
 - func: ger(Tensor self, Tensor vec2) -> Tensor
   variants: function, method
-  dispatch:
-    CompositeExplicitAutograd: ger
 
 - func: ger.out(Tensor self, Tensor vec2, *, Tensor(a!) out) -> Tensor(a!)
-  dispatch:
-    CompositeExplicitAutograd: ger_out
 
 - func: linalg_norm(Tensor self, Scalar? ord=None, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
   python_module: linalg
@@ -9778,22 +10514,16 @@
   python_module: linalg
   variants: function
 
-- func: _linalg_solve_out_helper_(Tensor(a!) self, Tensor(b!) other, Tensor(c!) infos) -> Tensor(a!)
-  variants: function
-  dispatch:
-    CPU: _linalg_solve_out_helper_cpu
-    CUDA: _linalg_solve_out_helper_cuda
-
 - func: linalg_solve(Tensor input, Tensor other) -> Tensor
   python_module: linalg
   variants: function
   dispatch:
-    CompositeExplicitAutograd: linalg_solve
+    CPU, CUDA: linalg_solve
 
 - func: linalg_solve.out(Tensor input, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
   dispatch:
-    CompositeExplicitAutograd: linalg_solve_out
+    CPU, CUDA: linalg_solve_out
 
 - func: linalg_tensorinv(Tensor self, int ind=2) -> Tensor
   python_module: linalg
@@ -9897,10 +10627,10 @@
   dispatch:
     CPU, CUDA: segment_reduce_kernel
 
-- func: segment_reduce_backward(Tensor grad, Tensor output, Tensor data, *, Tensor? lengths=None) -> Tensor
+- func: _segment_reduce_backward(Tensor grad, Tensor output, Tensor data, str reduce, *, Tensor? lengths=None, int axis=0) -> Tensor
   variants: function
   dispatch:
-    CPU, CUDA: segment_reduce_backward_kernel
+    CPU, CUDA: _segment_reduce_backward_kernel
 
 - func: pad_sequence(Tensor[] sequences, bool batch_first=False, float padding_value=0.0) -> Tensor
   python_module: nn