torch-rb 0.16.0 → 0.17.0

data/codegen/native_functions.yaml

@@ -549,8 +549,8 @@
   structured_delegate: add.out
   variants: function, method
   dispatch:
-    SparseCPU, SparseCUDA: add_sparse
-    SparseCsrCPU, SparseCsrCUDA: add_sparse_csr
+    SparseCPU, SparseCUDA, SparseMeta: add_sparse
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: add_sparse_csr
     MkldnnCPU: mkldnn_add
     ZeroTensor: add_zerotensor
     NestedTensorCPU, NestedTensorCUDA: NestedTensor_add_Tensor
@@ -561,8 +561,8 @@
   variants: method
   structured_delegate: add.out
   dispatch:
-    SparseCPU, SparseCUDA: add_sparse_
-    SparseCsrCPU, SparseCsrCUDA: add_sparse_csr_
+    SparseCPU, SparseCUDA, SparseMeta: add_sparse_
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: add_sparse_csr_
     MkldnnCPU: mkldnn_add_
     NestedTensorCPU, NestedTensorCUDA: NestedTensor_add__Tensor
   tags: pointwise
@@ -575,9 +575,9 @@
     Generic: add (AllAndComplex, BFloat16, Half, ComplexHalf)
     ScalarOnly: add (Bool)
   dispatch:
-    SparseCPU: add_out_sparse_cpu
+    SparseCPU, SparseMeta: add_out_sparse_cpu
     SparseCUDA: add_out_sparse_cuda
-    SparseCsrCPU: add_out_sparse_compressed_cpu
+    SparseCsrCPU, SparseCsrMeta: add_out_sparse_compressed_cpu
     SparseCsrCUDA: add_out_sparse_compressed_cuda
     MkldnnCPU: mkldnn_add_out
     MPS: add_out_mps
@@ -1750,6 +1750,7 @@
 - func: copy(Tensor self, Tensor src, bool non_blocking=False) -> Tensor
   variants: function
   dispatch:
+    Meta: copy_meta
     CompositeExplicitAutogradNonFunctional: copy
   tags: core
 
@@ -3127,6 +3128,7 @@
   structured: True
   dispatch:
     CPU, CUDA: isin_Tensor_Tensor_out
+    MPS: isin_Tensor_Tensor_out_mps
 
 - func: isin.Tensor_Tensor(Tensor elements, Tensor test_elements, *, bool assume_unique=False, bool invert=False) -> Tensor
   variants: function
@@ -3268,6 +3270,8 @@
   autogen: native_layer_norm_backward.out
   tags: core
 
+- func: rms_norm(Tensor input, int[] normalized_shape, Tensor? weight=None, float? eps=None) -> Tensor
+
 - func: nan_to_num(Tensor self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor
   variants: function, method
   dispatch:
@@ -3340,10 +3344,31 @@
   dispatch:
     CUDA: _cslt_sparse_mm_search
 
+- func: _sparse_semi_structured_tile(Tensor input, str algorithm="", bool use_cutlass=True) -> (Tensor, Tensor, Tensor, Tensor, Tensor)
+  dispatch:
+    CUDA: _sparse_semi_structured_tile
+
+- func: _sparse_semi_structured_apply(Tensor input, Tensor thread_masks) -> (Tensor, Tensor)
+  dispatch:
+    CUDA: _sparse_semi_structured_apply
+
+- func: _sparse_semi_structured_apply_dense(Tensor input, Tensor thread_masks) -> Tensor
+  dispatch:
+    CUDA: _sparse_semi_structured_apply_dense
+
+# DEPRECATED: Use torch.__sparse_semi_structured_mm/torch._sparse_semi_structured_addmm instead
 - func: _sparse_semi_structured_linear(Tensor input, Tensor weight, Tensor meta, *, Tensor? bias=None, str? activation=None, ScalarType? out_dtype=None) -> Tensor
   dispatch:
     CUDA: _sparse_semi_structured_linear
 
+- func: _sparse_semi_structured_mm(Tensor mat1, Tensor mat1_meta, Tensor mat2, *, ScalarType? out_dtype=None) -> Tensor
+  dispatch:
+    CUDA: _sparse_semi_structured_mm
+
+- func: _sparse_semi_structured_addmm(Tensor input, Tensor mat1, Tensor mat1_meta, Tensor mat2, *, Scalar alpha=1, Scalar beta=1, ScalarType? out_dtype=None) -> Tensor
+  dispatch:
+    CUDA: _sparse_semi_structured_addmm
+
 - func: _mixed_dtypes_linear(Tensor input, Tensor weight, Tensor scale, *, Tensor? bias=None, str? activation=None) -> Tensor
   dispatch:
     CUDA: _mixed_dtypes_linear
@@ -4084,10 +4109,12 @@
 
 - func: _int_mm(Tensor self, Tensor mat2) -> Tensor
   dispatch:
+    CPU: _int_mm_cpu
     CUDA: _int_mm_cuda
 
 - func: _int_mm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
+    CPU: _int_mm_out_cpu
     CUDA: _int_mm_out_cuda
 
 - func: _convert_weight_to_int4pack(Tensor self, int innerKTiles) -> Tensor
@@ -4098,11 +4125,13 @@
 - func: _weight_int4pack_mm(Tensor self, Tensor mat2, int qGroupSize, Tensor qScaleAndZeros) -> Tensor
   dispatch:
     CPU: _weight_int4pack_mm_cpu
+    MPS: _weight_int4pack_mm_mps
     CUDA: _weight_int4pack_mm_cuda
 
 - func: _weight_int8pack_mm(Tensor self, Tensor mat2, Tensor scales) -> Tensor
   dispatch:
     CPU: _weight_int8pack_mm_cpu
+    MPS: _weight_int8pack_mm_mps
 
 - func: _sparse_mm(Tensor sparse, Tensor dense) -> Tensor
   python_module: sparse
@@ -5397,7 +5426,7 @@
   autogen: slice_backward.out
 
 # NB: This op exists to back the implementation of reverse view_funcs for various views (chunk,
-# slice.Tensor, split_with_sizes, et. al.). Currently, these are only used during fake-ification
+# slice.Tensor, split_with_sizes, et al.). Currently, these are only used during fake-ification
 # of PT2 graph input subclass instances that are views. This means:
 # * This op shouldn't really show up in eager mode (so e.g. XLA shouldn't have to implement it)
 # * This op shouldn't show up in a PT2 graph (so a PT2 backend shouldn't have to implement it)
@@ -5620,10 +5649,12 @@
 - func: _chunk_cat(Tensor[] tensors, int dim, int num_chunks) -> Tensor
   dispatch:
     CompositeExplicitAutograd: _chunk_cat
+    CUDA: _chunk_cat_cuda
 
 - func: _chunk_cat.out(Tensor[] tensors, int dim, int num_chunks, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CompositeExplicitAutograd: _chunk_cat_out
+    CUDA: _chunk_cat_out_cuda
 
 - func: stack(Tensor[] tensors, int dim=0) -> Tensor
   dispatch:
@@ -5689,8 +5720,8 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: sum
-    SparseCPU, SparseCUDA: sum_coo
-    SparseCsrCPU, SparseCsrCUDA: sum_csr
+    SparseCPU, SparseCUDA, SparseMeta: sum_coo
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sum_csr
   autogen: sum.out
 
 - func: sum.dim_IntList(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
@@ -6200,6 +6231,12 @@
   category_override: dummy
   dispatch: {}
 
+- func: _nested_compute_contiguous_strides_offsets(Tensor nested_size) -> (Tensor, Tensor)
+  variants: function
+  device_check: NoCheck
+  dispatch:
+    CPU, CUDA: _nested_compute_contiguous_strides_offsets
+
 - func: _trilinear(Tensor i1, Tensor i2, Tensor i3, int[] expand1, int[] expand2, int[] expand3, int[] sumdim, int unroll_dim=1) -> Tensor
   dispatch:
     # calls unsqueeze
@@ -6465,7 +6502,7 @@
     CPU: _efficientzerotensor
     CUDA: _efficientzerotensor_cuda
     MPS: _efficientzerotensor_mps
-    Meta: _efficientzerotensor_meta
+    Meta: _efficientzerotensor_meta_symint
   autogen: _efficientzerotensor.out
 
 - func: zeros(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
@@ -6542,6 +6579,32 @@
     SparseCPU, SparseCUDA: norm_sparse
   autogen: native_norm.ScalarOpt_dim_dtype_out
 
+- func: _batch_norm_with_update(Tensor input, Tensor? weight, Tensor? bias, Tensor(a!) running_mean, Tensor(b!) running_var, float momentum, float eps) -> (Tensor, Tensor, Tensor, Tensor)
+  dispatch:
+    CPU: _batch_norm_with_update_cpu
+    CUDA: _batch_norm_with_update_cuda
+    MPS: _batch_norm_with_update_mps
+    MkldnnCPU: _batch_norm_with_update_mkldnn
+  autogen: _batch_norm_with_update_functional
+
+- func: _batch_norm_with_update.out(Tensor input, Tensor? weight, Tensor? bias, Tensor(a!) running_mean, Tensor(b!) running_var, float momentum, float eps, *, Tensor(d!) out, Tensor(e!) save_mean, Tensor(f!) save_invstd, Tensor(g!) reserve) -> (Tensor(d!), Tensor(e!), Tensor(f!), Tensor(g!))
+  dispatch:
+    CPU: _batch_norm_with_update_cpu_out
+    CUDA: _batch_norm_with_update_cuda_out
+    MPS: _batch_norm_with_update_mps_out
+
+- func: _batch_norm_no_update(Tensor input, Tensor? weight, Tensor? bias, Tensor? running_mean, Tensor? running_var, float momentum, float eps) -> (Tensor, Tensor, Tensor, Tensor)
+  dispatch:
+    CompositeExplicitAutograd: _batch_norm_no_update
+  autogen: _batch_norm_no_update.out
+
+- func: batch_norm_backward(Tensor grad_out, Tensor input, Tensor weight, Tensor? running_mean, Tensor? running_var, Tensor? save_mean, Tensor? save_var, bool update, float eps, bool[3] output_mask, Tensor reserve) -> (Tensor, Tensor, Tensor)
+  dispatch:
+    CPU: _new_batch_norm_backward_cpu
+    CUDA: _new_batch_norm_backward_cuda
+    MPS: _new_batch_norm_backward_mps
+    MkldnnCPU: _new_batch_norm_backward_mkldnn
+
 # TODO: reduce signatures down to one when optional args is available
 - func: _sparse_sum(Tensor self) -> Tensor
 
@@ -7042,6 +7105,10 @@
 # FIXME: would be nicer if TensorOptions was optional based; not adding default arguments for options given
 # the default would never make sense.
 
+- func: _sparse_compressed_tensor_with_dims(int nnz, int dense_dim, int[] size, int[] blocksize, ScalarType index_dtype, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
+  dispatch:
+    CompositeExplicitAutograd: sparse_compressed_tensor_with_dims
+
 - func: sparse_compressed_tensor.comp_plain_value_size(Tensor compressed_indices, Tensor plain_indices, Tensor values, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
   dispatch:
     CompositeExplicitAutograd: sparse_compressed_tensor
@@ -7146,9 +7213,9 @@
 - func: sparse_dim(Tensor self) -> int
   variants: method
   dispatch:
-    CPU, CUDA: sparse_dim_strided
     SparseCPU, SparseCUDA, SparseMeta: sparse_dim_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sparse_dim_sparse_csr
+    CompositeExplicitAutograd: sparse_dim_default
   device_check: NoCheck
   device_guard: False
 
@@ -7163,9 +7230,9 @@
 - func: dense_dim(Tensor self) -> int
   variants: method
   dispatch:
-    CPU, CUDA: dense_dim_strided
     SparseCPU, SparseCUDA, SparseMeta: dense_dim_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: dense_dim_sparse_csr
+    CompositeExplicitAutograd: dense_dim_default
   device_check: NoCheck
   device_guard: False
 
@@ -7296,7 +7363,7 @@
   device_check: NoCheck  # Allows copy into different device
   variants: function
   dispatch:
-    SparseCPU, SparseCUDA: copy_sparse_
+    SparseCPU, SparseCUDA, SparseMeta: copy_sparse_
   autogen: copy_sparse_to_sparse, copy_sparse_to_sparse.out
 
 # By adding the AutogradNestedTensor this makes this function CompositeImplicit-like for nested tensors
@@ -7399,7 +7466,7 @@
     MkldnnCPU: mkldnn_reorder_conv2d_weight
   autogen: mkldnn_reorder_conv2d_weight.out
 
-- func: mkldnn_reorder_conv3d_weight(Tensor self, SymInt[3] padding=0, SymInt[3] stride=1, SymInt[3] dilation=1, SymInt groups=1) -> Tensor
+- func: mkldnn_reorder_conv3d_weight(Tensor self, SymInt[3] padding=0, SymInt[3] stride=1, SymInt[3] dilation=1, SymInt groups=1, SymInt[]? input_size=None) -> Tensor
   variants: function
   python_module: nn
   dispatch:
@@ -7647,7 +7714,7 @@
 
 - func: result_type.Scalar_Scalar(Scalar scalar1, Scalar scalar2) -> ScalarType
 
-- func: can_cast(ScalarType from, ScalarType to) -> bool
+- func: can_cast(ScalarType from_, ScalarType to) -> bool
   variants: function
 
 - func: promote_types(ScalarType type1, ScalarType type2) -> ScalarType
@@ -10222,6 +10289,7 @@
   variants: method, function
   dispatch:
     CompositeExplicitAutograd: alias
+    NestedTensorCPU, NestedTensorCUDA: alias_nested
   tags: core
 
 - func: _amp_foreach_non_finite_check_and_unscale_(Tensor(a!)[] self, Tensor(b!) found_inf, Tensor inv_scale) -> ()
@@ -10255,14 +10323,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_add_scalar_kernel_slow
     CUDA: foreach_tensor_add_scalar_kernel_cuda
 
 - func: _foreach_add_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_add_scalar_kernel_slow_
     CUDA: foreach_tensor_add_scalar_kernel_cuda_
   autogen: _foreach_add.Scalar_out
 
@@ -10270,14 +10338,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_add_list_kernel_slow
     CUDA: foreach_tensor_add_list_kernel_cuda
 
 - func: _foreach_add_.List(Tensor(a!)[] self, Tensor[] other, *, Scalar alpha=1) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_add_list_kernel_slow_
     CUDA: foreach_tensor_add_list_kernel_cuda_
   autogen: _foreach_add.List_out
 
@@ -10285,14 +10353,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_add_scalarlist_kernel_slow
     CUDA: foreach_tensor_add_scalarlist_kernel_cuda
 
 - func: _foreach_add_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_add_scalarlist_kernel_slow_
     CUDA: foreach_tensor_add_scalarlist_kernel_cuda_
   autogen: _foreach_add.ScalarList_out
 
@@ -10300,14 +10368,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_tensor_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_add_tensor_kernel_slow
     CUDA: foreach_tensor_add_tensor_kernel_cuda
 
 - func: _foreach_add_.Tensor(Tensor(a!)[] self, Tensor other, *, Scalar alpha=1) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_add_tensor_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_add_tensor_kernel_slow_
     CUDA: foreach_tensor_add_tensor_kernel_cuda_
   autogen: _foreach_add.Tensor_out
 
@@ -10315,14 +10383,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sub_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_sub_scalar_kernel_slow
     CUDA: foreach_tensor_sub_scalar_kernel_cuda
 
 - func: _foreach_sub_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sub_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_sub_scalar_kernel_slow_
     CUDA: foreach_tensor_sub_scalar_kernel_cuda_
   autogen: _foreach_sub.Scalar_out
 
@@ -10330,14 +10398,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sub_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_sub_list_kernel_slow
     CUDA: foreach_tensor_sub_list_kernel_cuda
 
 - func: _foreach_sub_.List(Tensor(a!)[] self, Tensor[] other, *, Scalar alpha=1) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sub_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_sub_list_kernel_slow_
     CUDA: foreach_tensor_sub_list_kernel_cuda_
   autogen: _foreach_sub.List_out
 
@@ -10345,14 +10413,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sub_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_sub_scalarlist_kernel_slow
     CUDA: foreach_tensor_sub_scalarlist_kernel_cuda
 
 - func: _foreach_sub_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sub_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_sub_scalarlist_kernel_slow_
     CUDA: foreach_tensor_sub_scalarlist_kernel_cuda_
   autogen: _foreach_sub.ScalarList_out
 
@@ -10360,14 +10428,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_mul_scalar_kernel_slow
     CUDA: foreach_tensor_mul_scalar_kernel_cuda
 
 - func: _foreach_mul_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_mul_scalar_kernel_slow_
     CUDA: foreach_tensor_mul_scalar_kernel_cuda_
   autogen: _foreach_mul.Scalar_out
 
@@ -10375,14 +10443,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_mul_list_kernel_slow
     CUDA: foreach_tensor_mul_list_kernel_cuda
 
 - func: _foreach_mul_.List(Tensor(a!)[] self, Tensor[] other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_mul_list_kernel_slow_
     CUDA: foreach_tensor_mul_list_kernel_cuda_
   autogen: _foreach_mul.List_out
 
@@ -10390,14 +10458,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_mul_scalarlist_kernel_slow
     CUDA: foreach_tensor_mul_scalarlist_kernel_cuda
 
 - func: _foreach_mul_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_mul_scalarlist_kernel_slow_
     CUDA: foreach_tensor_mul_scalarlist_kernel_cuda_
   autogen: _foreach_mul.ScalarList_out
 
@@ -10405,14 +10473,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_tensor_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_mul_tensor_kernel_slow
     CUDA: foreach_tensor_mul_tensor_kernel_cuda
 
 - func: _foreach_mul_.Tensor(Tensor(a!)[] self, Tensor other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_mul_tensor_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_mul_tensor_kernel_slow_
     CUDA: foreach_tensor_mul_tensor_kernel_cuda_
   autogen: _foreach_mul.Tensor_out
 
@@ -10420,14 +10488,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_div_scalar_kernel_slow
     CUDA: foreach_tensor_div_scalar_kernel_cuda
 
 - func: _foreach_div_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_div_scalar_kernel_slow_
     CUDA: foreach_tensor_div_scalar_kernel_cuda_
   autogen: _foreach_div.Scalar_out
 
@@ -10435,14 +10503,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_div_list_kernel_slow
     CUDA: foreach_tensor_div_list_kernel_cuda
 
 - func: _foreach_div_.List(Tensor(a!)[] self, Tensor[] other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_div_list_kernel_slow_
     CUDA: foreach_tensor_div_list_kernel_cuda_
   autogen: _foreach_div.List_out
 
@@ -10450,14 +10518,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_div_scalarlist_kernel_slow
     CUDA: foreach_tensor_div_scalarlist_kernel_cuda
 
 - func: _foreach_div_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_div_scalarlist_kernel_slow_
     CUDA: foreach_tensor_div_scalarlist_kernel_cuda_
   autogen: _foreach_div.ScalarList_out
 
@@ -10465,14 +10533,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_tensor_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_div_tensor_kernel_slow
     CUDA: foreach_tensor_div_tensor_kernel_cuda
 
 - func: _foreach_div_.Tensor(Tensor(a!)[] self, Tensor other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_div_tensor_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_div_tensor_kernel_slow_
     CUDA: foreach_tensor_div_tensor_kernel_cuda_
   autogen: _foreach_div.Tensor_out
 
@@ -10480,14 +10548,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalar_kernel_slow
     CUDA: foreach_tensor_clamp_max_scalar_kernel_cuda
 
 - func: _foreach_clamp_max_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalar_kernel_slow_
     CUDA: foreach_tensor_clamp_max_scalar_kernel_cuda_
   autogen: _foreach_clamp_max.Scalar_out
 
@@ -10495,14 +10563,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_list_kernel_slow
     CUDA: foreach_tensor_clamp_max_list_kernel_cuda
 
 - func: _foreach_clamp_max_.List(Tensor(a!)[] self, Tensor[] other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_list_kernel_slow_
     CUDA: foreach_tensor_clamp_max_list_kernel_cuda_
   autogen: _foreach_clamp_max.List_out
 
@@ -10510,14 +10578,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalarlist_kernel_slow
     CUDA: foreach_tensor_clamp_max_scalarlist_kernel_cuda
 
 - func: _foreach_clamp_max_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalarlist_kernel_slow_
     CUDA: foreach_tensor_clamp_max_scalarlist_kernel_cuda_
   autogen: _foreach_clamp_max.ScalarList_out
 
@@ -10525,14 +10593,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalar_kernel_slow
     CUDA: foreach_tensor_clamp_min_scalar_kernel_cuda
 
 - func: _foreach_clamp_min_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalar_kernel_slow_
     CUDA: foreach_tensor_clamp_min_scalar_kernel_cuda_
   autogen: _foreach_clamp_min.Scalar_out
 
@@ -10540,14 +10608,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_list_kernel_slow
     CUDA: foreach_tensor_clamp_min_list_kernel_cuda
 
 - func: _foreach_clamp_min_.List(Tensor(a!)[] self, Tensor[] other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_list_kernel_slow_
     CUDA: foreach_tensor_clamp_min_list_kernel_cuda_
   autogen: _foreach_clamp_min.List_out
 
@@ -10555,14 +10623,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalarlist_kernel_slow
     CUDA: foreach_tensor_clamp_min_scalarlist_kernel_cuda
 
 - func: _foreach_clamp_min_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalarlist_kernel_slow_
     CUDA: foreach_tensor_clamp_min_scalarlist_kernel_cuda_
   autogen: _foreach_clamp_min.ScalarList_out
 
@@ -10571,14 +10639,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalar_kernel_slow
     CUDA: foreach_tensor_clamp_min_scalar_kernel_cuda
 
 - func: _foreach_maximum_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalar_kernel_slow_
     CUDA: foreach_tensor_clamp_min_scalar_kernel_cuda_
   autogen: _foreach_maximum.Scalar_out
 
@@ -10587,14 +10655,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_list_kernel_slow
     CUDA: foreach_tensor_clamp_min_list_kernel_cuda
 
 - func: _foreach_maximum_.List(Tensor(a!)[] self, Tensor[] other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_list_kernel_slow_
     CUDA: foreach_tensor_clamp_min_list_kernel_cuda_
   autogen: _foreach_maximum.List_out
 
@@ -10603,14 +10671,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalarlist_kernel_slow
     CUDA: foreach_tensor_clamp_min_scalarlist_kernel_cuda
 
 - func: _foreach_maximum_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_min_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_min_scalarlist_kernel_slow_
     CUDA: foreach_tensor_clamp_min_scalarlist_kernel_cuda_
   autogen: _foreach_maximum.ScalarList_out
 
@@ -10618,14 +10686,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalar_kernel_slow
     CUDA: foreach_tensor_clamp_max_scalar_kernel_cuda
 
 - func: _foreach_minimum_.Scalar(Tensor(a!)[] self, Scalar scalar) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalar_kernel_slow_
     CUDA: foreach_tensor_clamp_max_scalar_kernel_cuda_
   autogen: _foreach_minimum.Scalar_out
 
@@ -10633,14 +10701,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_list_kernel_slow
     CUDA: foreach_tensor_clamp_max_list_kernel_cuda
 
 - func: _foreach_minimum_.List(Tensor(a!)[] self, Tensor[] other) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_list_kernel_slow_
     CUDA: foreach_tensor_clamp_max_list_kernel_cuda_
   autogen: _foreach_minimum.List_out
 
@@ -10648,14 +10716,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalarlist_kernel_slow
     CUDA: foreach_tensor_clamp_max_scalarlist_kernel_cuda
 
 - func: _foreach_minimum_.ScalarList(Tensor(a!)[] self, Scalar[] scalars) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_clamp_max_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_clamp_max_scalarlist_kernel_slow_
     CUDA: foreach_tensor_clamp_max_scalarlist_kernel_cuda_
   autogen: _foreach_minimum.ScalarList_out
 
@@ -10663,28 +10731,28 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcdiv_scalar_slow
+    CompositeExplicitAutograd: foreach_tensor_addcdiv_scalar_slow
     CUDA: foreach_tensor_addcdiv_scalar_cuda
 
 - func: _foreach_addcdiv.ScalarList(Tensor[] self, Tensor[] tensor1, Tensor[] tensor2, Scalar[] scalars) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcdiv_scalarlist_slow
+    CompositeExplicitAutograd: foreach_tensor_addcdiv_scalarlist_slow
     CUDA: foreach_tensor_addcdiv_scalarlist_cuda
 
 - func: _foreach_addcdiv.Tensor(Tensor[] self, Tensor[] tensor1, Tensor[] tensor2, Tensor scalars) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcdiv_tensor_slow
+    CompositeExplicitAutograd: foreach_tensor_addcdiv_tensor_slow
     CUDA: foreach_tensor_addcdiv_tensor_cuda
 
 - func: _foreach_addcdiv_.Scalar(Tensor(a!)[] self, Tensor[] tensor1, Tensor[] tensor2, Scalar value=1) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcdiv_scalar_slow_
+    CompositeExplicitAutograd: foreach_tensor_addcdiv_scalar_slow_
     CUDA: foreach_tensor_addcdiv_scalar_cuda_
   autogen: _foreach_addcdiv.Scalar_out
 
@@ -10692,7 +10760,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcdiv_scalarlist_slow_
+    CompositeExplicitAutograd: foreach_tensor_addcdiv_scalarlist_slow_
     CUDA: foreach_tensor_addcdiv_scalarlist_cuda_
   autogen: _foreach_addcdiv.ScalarList_out
 
@@ -10700,7 +10768,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcdiv_tensor_slow_
+    CompositeExplicitAutograd: foreach_tensor_addcdiv_tensor_slow_
     CUDA: foreach_tensor_addcdiv_tensor_cuda_
   autogen: _foreach_addcdiv.Tensor_out
 
@@ -10708,28 +10776,28 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcmul_scalar_slow
+    CompositeExplicitAutograd: foreach_tensor_addcmul_scalar_slow
     CUDA: foreach_tensor_addcmul_scalar_cuda
 
 - func: _foreach_addcmul.ScalarList(Tensor[] self, Tensor[] tensor1, Tensor[] tensor2, Scalar[] scalars) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcmul_scalarlist_slow
+    CompositeExplicitAutograd: foreach_tensor_addcmul_scalarlist_slow
     CUDA: foreach_tensor_addcmul_scalarlist_cuda
 
 - func: _foreach_addcmul.Tensor(Tensor[] self, Tensor[] tensor1, Tensor[] tensor2, Tensor scalars) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcmul_tensor_slow
+    CompositeExplicitAutograd: foreach_tensor_addcmul_tensor_slow
     CUDA: foreach_tensor_addcmul_tensor_cuda
 
 - func: _foreach_addcmul_.Scalar(Tensor(a!)[] self, Tensor[] tensor1, Tensor[] tensor2, Scalar value=1) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcmul_scalar_slow_
+    CompositeExplicitAutograd: foreach_tensor_addcmul_scalar_slow_
     CUDA: foreach_tensor_addcmul_scalar_cuda_
   autogen: _foreach_addcmul.Scalar_out
 
@@ -10737,7 +10805,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcmul_scalarlist_slow_
+    CompositeExplicitAutograd: foreach_tensor_addcmul_scalarlist_slow_
     CUDA: foreach_tensor_addcmul_scalarlist_cuda_
   autogen: _foreach_addcmul.ScalarList_out
 
@@ -10745,7 +10813,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_addcmul_tensor_slow_
+    CompositeExplicitAutograd: foreach_tensor_addcmul_tensor_slow_
     CUDA: foreach_tensor_addcmul_tensor_cuda_
   autogen: _foreach_addcmul.Tensor_out
 
@@ -10753,14 +10821,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_abs_slow
+    CompositeExplicitAutograd: foreach_tensor_abs_slow
     CUDA: foreach_tensor_abs_cuda
 
 - func: _foreach_abs_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_abs_slow_
+    CompositeExplicitAutograd: foreach_tensor_abs_slow_
     CUDA: foreach_tensor_abs_cuda_
   autogen: _foreach_abs.out
 
@@ -10768,14 +10836,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_acos_slow
+    CompositeExplicitAutograd: foreach_tensor_acos_slow
     CUDA: foreach_tensor_acos_cuda
 
 - func: _foreach_acos_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_acos_slow_
+    CompositeExplicitAutograd: foreach_tensor_acos_slow_
     CUDA: foreach_tensor_acos_cuda_
   autogen: _foreach_acos.out
 
@@ -10783,14 +10851,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_asin_slow
+    CompositeExplicitAutograd: foreach_tensor_asin_slow
     CUDA: foreach_tensor_asin_cuda
 
 - func: _foreach_asin_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_asin_slow_
+    CompositeExplicitAutograd: foreach_tensor_asin_slow_
     CUDA: foreach_tensor_asin_cuda_
   autogen: _foreach_asin.out
 
@@ -10798,14 +10866,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_atan_slow
+    CompositeExplicitAutograd: foreach_tensor_atan_slow
     CUDA: foreach_tensor_atan_cuda
 
 - func: _foreach_atan_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_atan_slow_
+    CompositeExplicitAutograd: foreach_tensor_atan_slow_
     CUDA: foreach_tensor_atan_cuda_
   autogen: _foreach_atan.out
 
@@ -10813,14 +10881,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_ceil_slow
+    CompositeExplicitAutograd: foreach_tensor_ceil_slow
     CUDA: foreach_tensor_ceil_cuda
 
 - func: _foreach_ceil_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_ceil_slow_
+    CompositeExplicitAutograd: foreach_tensor_ceil_slow_
     CUDA: foreach_tensor_ceil_cuda_
   autogen: _foreach_ceil.out
 
@@ -10828,14 +10896,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_cos_slow
+    CompositeExplicitAutograd: foreach_tensor_cos_slow
     CUDA: foreach_tensor_cos_cuda
 
 - func: _foreach_cos_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_cos_slow_
+    CompositeExplicitAutograd: foreach_tensor_cos_slow_
     CUDA: foreach_tensor_cos_cuda_
   autogen: _foreach_cos.out
 
@@ -10843,14 +10911,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_cosh_slow
+    CompositeExplicitAutograd: foreach_tensor_cosh_slow
     CUDA: foreach_tensor_cosh_cuda
 
 - func: _foreach_cosh_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_cosh_slow_
+    CompositeExplicitAutograd: foreach_tensor_cosh_slow_
     CUDA: foreach_tensor_cosh_cuda_
   autogen: _foreach_cosh.out
 
@@ -10858,14 +10926,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_erf_slow
+    CompositeExplicitAutograd: foreach_tensor_erf_slow
     CUDA: foreach_tensor_erf_cuda
 
 - func: _foreach_erf_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_erf_slow_
+    CompositeExplicitAutograd: foreach_tensor_erf_slow_
     CUDA: foreach_tensor_erf_cuda_
   autogen: _foreach_erf.out
 
@@ -10873,14 +10941,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_erfc_slow
+    CompositeExplicitAutograd: foreach_tensor_erfc_slow
     CUDA: foreach_tensor_erfc_cuda
 
 - func: _foreach_erfc_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_erfc_slow_
+    CompositeExplicitAutograd: foreach_tensor_erfc_slow_
     CUDA: foreach_tensor_erfc_cuda_
   autogen: _foreach_erfc.out
 
@@ -10888,14 +10956,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_exp_slow
+    CompositeExplicitAutograd: foreach_tensor_exp_slow
     CUDA: foreach_tensor_exp_cuda
 
 - func: _foreach_exp_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_exp_slow_
+    CompositeExplicitAutograd: foreach_tensor_exp_slow_
     CUDA: foreach_tensor_exp_cuda_
   autogen: _foreach_exp.out
 
@@ -10903,14 +10971,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_expm1_slow
+    CompositeExplicitAutograd: foreach_tensor_expm1_slow
     CUDA: foreach_tensor_expm1_cuda
 
 - func: _foreach_expm1_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_expm1_slow_
+    CompositeExplicitAutograd: foreach_tensor_expm1_slow_
     CUDA: foreach_tensor_expm1_cuda_
   autogen: _foreach_expm1.out
 
@@ -10918,14 +10986,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_floor_slow
+    CompositeExplicitAutograd: foreach_tensor_floor_slow
     CUDA: foreach_tensor_floor_cuda
 
 - func: _foreach_floor_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_floor_slow_
+    CompositeExplicitAutograd: foreach_tensor_floor_slow_
     CUDA: foreach_tensor_floor_cuda_
   autogen: _foreach_floor.out
 
@@ -10933,14 +11001,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_frac_slow
+    CompositeExplicitAutograd: foreach_tensor_frac_slow
     CUDA: foreach_tensor_frac_cuda
 
 - func: _foreach_frac_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_frac_slow_
+    CompositeExplicitAutograd: foreach_tensor_frac_slow_
     CUDA: foreach_tensor_frac_cuda_
   autogen: _foreach_frac.out
 
@@ -10948,7 +11016,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_ternary_lerp_slow
+    CompositeExplicitAutograd: foreach_tensor_ternary_lerp_slow
     CUDA: foreach_tensor_lerp_ternary_cuda
   autogen: _foreach_lerp.List_out
 
@@ -10956,7 +11024,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_ternary_lerp_slow_
+    CompositeExplicitAutograd: foreach_tensor_ternary_lerp_slow_
     CUDA: foreach_tensor_lerp_ternary_cuda_
   autogen: _foreach_lerp.List_out
 
@@ -10964,7 +11032,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_lerp_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_lerp_list_kernel_slow
     CUDA: foreach_tensor_lerp_list_cuda
   autogen: _foreach_lerp.Scalar_out
 
@@ -10972,7 +11040,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_lerp_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_lerp_list_kernel_slow_
     CUDA: foreach_tensor_lerp_list_cuda_
   autogen: _foreach_lerp.Scalar_out
 
@@ -10980,14 +11048,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_lgamma_slow
+    CompositeExplicitAutograd: foreach_tensor_lgamma_slow
     CUDA: foreach_tensor_lgamma_cuda
 
 - func: _foreach_lgamma_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_lgamma_slow_
+    CompositeExplicitAutograd: foreach_tensor_lgamma_slow_
     CUDA: foreach_tensor_lgamma_cuda_
   autogen: _foreach_lgamma.out
 
@@ -10995,14 +11063,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log_slow
+    CompositeExplicitAutograd: foreach_tensor_log_slow
     CUDA: foreach_tensor_log_cuda
 
 - func: _foreach_log_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log_slow_
+    CompositeExplicitAutograd: foreach_tensor_log_slow_
     CUDA: foreach_tensor_log_cuda_
   autogen: _foreach_log.out
 
@@ -11010,14 +11078,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log10_slow
+    CompositeExplicitAutograd: foreach_tensor_log10_slow
     CUDA: foreach_tensor_log10_cuda
 
 - func: _foreach_log10_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log10_slow_
+    CompositeExplicitAutograd: foreach_tensor_log10_slow_
     CUDA: foreach_tensor_log10_cuda_
   autogen: _foreach_log10.out
 
@@ -11025,14 +11093,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log1p_slow
+    CompositeExplicitAutograd: foreach_tensor_log1p_slow
     CUDA: foreach_tensor_log1p_cuda
 
 - func: _foreach_log1p_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log1p_slow_
+    CompositeExplicitAutograd: foreach_tensor_log1p_slow_
     CUDA: foreach_tensor_log1p_cuda_
   autogen: _foreach_log1p.out
 
@@ -11040,37 +11108,45 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log2_slow
+    CompositeExplicitAutograd: foreach_tensor_log2_slow
     CUDA: foreach_tensor_log2_cuda
 
 - func: _foreach_log2_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_log2_slow_
+    CompositeExplicitAutograd: foreach_tensor_log2_slow_
     CUDA: foreach_tensor_log2_cuda_
   autogen: _foreach_log2.out
 
+- func: _foreach_max(Tensor[] self) -> Tensor[]
+  device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: foreach_tensor_max_slow
+    CUDA: foreach_tensor_max_cuda
+  autogen: _foreach_max.out
+
 - func: _foreach_neg(Tensor[] self) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_neg_slow
+    CompositeExplicitAutograd: foreach_tensor_neg_slow
     CUDA: foreach_tensor_neg_cuda
 
 - func: _foreach_neg_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_neg_slow_
+    CompositeExplicitAutograd: foreach_tensor_neg_slow_
     CUDA: foreach_tensor_neg_cuda_
   autogen: _foreach_neg.out
 
-- func: _foreach_norm.Scalar(Tensor[] self, Scalar ord=2) -> Tensor[]
+- func: _foreach_norm.Scalar(Tensor[] self, Scalar ord=2, ScalarType? dtype=None) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_norm_slow
+    CompositeExplicitAutograd: foreach_tensor_norm_slow
     CUDA: foreach_tensor_norm_cuda
   autogen: _foreach_norm.Scalar_out
 
@@ -11078,35 +11154,35 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_pow_list_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_pow_list_kernel_slow
     CUDA: foreach_tensor_pow_list_kernel_cuda
 
 - func: _foreach_pow.Scalar(Tensor[] self, Scalar exponent) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_pow_scalar_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_pow_scalar_kernel_slow
     CUDA: foreach_tensor_pow_scalar_kernel_cuda
 
 - func: _foreach_pow.ScalarList(Tensor[] self, Scalar[] exponent) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_pow_scalarlist_kernel_slow
+    CompositeExplicitAutograd: foreach_tensor_pow_scalarlist_kernel_slow
     CUDA: foreach_tensor_pow_scalarlist_kernel_cuda
 
 - func: _foreach_pow.ScalarAndTensor(Scalar self, Tensor[] exponent) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_scalar_pow_list_kernel_slow
+    CompositeExplicitAutograd: foreach_scalar_pow_list_kernel_slow
     CUDA: foreach_scalar_pow_list_kernel_cuda
 
 - func: _foreach_pow_.List(Tensor(a!)[] self, Tensor[] exponent) -> ()
   device_check: NoCheck
   variants: function
   dispatch:
-    CPU: foreach_tensor_pow_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_pow_list_kernel_slow_
     CUDA: foreach_tensor_pow_list_kernel_cuda_
   autogen: _foreach_pow.List_out
 
@@ -11114,7 +11190,7 @@
   device_check: NoCheck
   variants: function
   dispatch:
-    CPU: foreach_tensor_pow_scalar_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_pow_scalar_kernel_slow_
     CUDA: foreach_tensor_pow_scalar_kernel_cuda_
   autogen: _foreach_pow.Scalar_out
 
@@ -11122,7 +11198,7 @@
   device_check: NoCheck
   variants: function
   dispatch:
-    CPU: foreach_tensor_pow_scalarlist_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_pow_scalarlist_kernel_slow_
     CUDA: foreach_tensor_pow_scalarlist_kernel_cuda_
   autogen: _foreach_pow.ScalarList_out
 
@@ -11130,14 +11206,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_reciprocal_slow
+    CompositeExplicitAutograd: foreach_tensor_reciprocal_slow
     CUDA: foreach_tensor_reciprocal_cuda
 
 - func: _foreach_reciprocal_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_reciprocal_slow_
+    CompositeExplicitAutograd: foreach_tensor_reciprocal_slow_
     CUDA: foreach_tensor_reciprocal_cuda_
   autogen: _foreach_reciprocal.out
 
@@ -11145,14 +11221,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_round_slow
+    CompositeExplicitAutograd: foreach_tensor_round_slow
     CUDA: foreach_tensor_round_cuda
 
 - func: _foreach_round_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_round_slow_
+    CompositeExplicitAutograd: foreach_tensor_round_slow_
     CUDA: foreach_tensor_round_cuda_
   autogen: _foreach_round.out
 
@@ -11160,14 +11236,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sigmoid_slow
+    CompositeExplicitAutograd: foreach_tensor_sigmoid_slow
     CUDA: foreach_tensor_sigmoid_cuda
 
 - func: _foreach_sigmoid_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sigmoid_slow_
+    CompositeExplicitAutograd: foreach_tensor_sigmoid_slow_
     CUDA: foreach_tensor_sigmoid_cuda_
   autogen: _foreach_sigmoid.out
 
@@ -11175,14 +11251,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sign_slow
+    CompositeExplicitAutograd: foreach_tensor_sign_slow
     CUDA: foreach_tensor_sign_cuda
 
 - func: _foreach_sign_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sign_slow_
+    CompositeExplicitAutograd: foreach_tensor_sign_slow_
     CUDA: foreach_tensor_sign_cuda_
   autogen: _foreach_sign.out
 
@@ -11190,14 +11266,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sin_slow
+    CompositeExplicitAutograd: foreach_tensor_sin_slow
     CUDA: foreach_tensor_sin_cuda
 
 - func: _foreach_sin_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sin_slow_
+    CompositeExplicitAutograd: foreach_tensor_sin_slow_
     CUDA: foreach_tensor_sin_cuda_
   autogen: _foreach_sin.out
 
@@ -11205,14 +11281,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sinh_slow
+    CompositeExplicitAutograd: foreach_tensor_sinh_slow
     CUDA: foreach_tensor_sinh_cuda
 
 - func: _foreach_sinh_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sinh_slow_
+    CompositeExplicitAutograd: foreach_tensor_sinh_slow_
     CUDA: foreach_tensor_sinh_cuda_
   autogen: _foreach_sinh.out
 
@@ -11220,14 +11296,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sqrt_slow
+    CompositeExplicitAutograd: foreach_tensor_sqrt_slow
     CUDA: foreach_tensor_sqrt_cuda
 
 - func: _foreach_sqrt_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_sqrt_slow_
+    CompositeExplicitAutograd: foreach_tensor_sqrt_slow_
     CUDA: foreach_tensor_sqrt_cuda_
   autogen: _foreach_sqrt.out
 
@@ -11235,14 +11311,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_tan_slow
+    CompositeExplicitAutograd: foreach_tensor_tan_slow
     CUDA: foreach_tensor_tan_cuda
 
 - func: _foreach_tan_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_tan_slow_
+    CompositeExplicitAutograd: foreach_tensor_tan_slow_
     CUDA: foreach_tensor_tan_cuda_
   autogen: _foreach_tan.out
 
@@ -11250,14 +11326,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_tanh_slow
+    CompositeExplicitAutograd: foreach_tensor_tanh_slow
     CUDA: foreach_tensor_tanh_cuda
 
 - func: _foreach_tanh_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_tanh_slow_
+    CompositeExplicitAutograd: foreach_tensor_tanh_slow_
     CUDA: foreach_tensor_tanh_cuda_
   autogen: _foreach_tanh.out
 
@@ -11265,14 +11341,14 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_trunc_slow
+    CompositeExplicitAutograd: foreach_tensor_trunc_slow
     CUDA: foreach_tensor_trunc_cuda
 
 - func: _foreach_trunc_(Tensor(a!)[] self) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_trunc_slow_
+    CompositeExplicitAutograd: foreach_tensor_trunc_slow_
     CUDA: foreach_tensor_trunc_cuda_
   autogen: _foreach_trunc.out
 
@@ -11280,7 +11356,7 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_zero_slow_
+    CompositeExplicitAutograd: foreach_tensor_zero_slow_
     CUDA: foreach_tensor_zero_cuda_
   autogen: _foreach_zero, _foreach_zero.out
 
@@ -11288,9 +11364,15 @@
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
   dispatch:
-    CPU: foreach_tensor_copy_list_kernel_slow_
+    CompositeExplicitAutograd: foreach_tensor_copy_list_kernel_slow_
     CUDA: foreach_tensor_copy_list_kernel_cuda_
-  autogen: _foreach_copy, _foreach_copy.out
+  autogen: _foreach_copy.out
+
+- func: _foreach_copy(Tensor[] self, Tensor[] src, bool non_blocking=False) -> Tensor[] self_out
+  device_check: NoCheck
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: _foreach_copy
 
 - func: bucketize.Tensor(Tensor self, Tensor boundaries, *, bool out_int32=False, bool right=False) -> Tensor
   dispatch:
@@ -14562,6 +14644,16 @@
     NestedTensorCUDA: NestedTensor_to_padded_tensor_cuda
   autogen: to_padded_tensor.out
 
+- func: _jagged_to_padded_dense_forward(Tensor values, Tensor[] offsets, SymInt[] max_lengths, float padding_value=0.0) -> Tensor
+  variants: function
+  dispatch:
+    CUDA: _fbgemm_jagged_to_padded_dense_forward
+
+- func: _padded_dense_to_jagged_forward(Tensor dense, Tensor[] offsets, SymInt? total_L=None) -> Tensor
+  variants: function
+  dispatch:
+    CUDA: _fbgemm_dense_to_jagged_forward_symint
+
 - func: _nested_tensor_softmax_with_shape(Tensor self, Tensor query) -> Tensor
   dispatch:
     NestedTensorCPU: NestedTensor_softmax_dropout
@@ -14636,31 +14728,36 @@
     CUDA: _scaled_dot_product_efficient_attention_backward_cuda
   tags: nondeterministic_seeded
 
-- func: _scaled_dot_product_cudnn_attention(Tensor query, Tensor key, Tensor value, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset)
+- func: _scaled_dot_product_cudnn_attention(Tensor query, Tensor key, Tensor value, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   dispatch:
     CUDA: _scaled_dot_product_cudnn_attention_cuda
   tags: nondeterministic_seeded
 
-- func: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
+- func: _scaled_dot_product_cudnn_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None) -> (Tensor, Tensor, Tensor)
+  dispatch:
+    CUDA: _scaled_dot_product_cudnn_attention_backward_cuda
+  tags: nondeterministic_seeded
+
+- func: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None, Tensor? seqused_k=None, Tensor? alibi_slopes=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   variants: function
   dispatch:
     CUDA: _flash_attention_forward
   tags: nondeterministic_seeded
 
-- func: _flash_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None) -> (Tensor, Tensor, Tensor)
+- func: _flash_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor out, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, Tensor philox_seed, Tensor philox_offset, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None) -> (Tensor, Tensor, Tensor)
   device_check: NoCheck
   variants: function
   dispatch:
     CUDA: _flash_attention_backward
 
 # Returns output, logsumexp if compute_logsumexp
-- func: _efficient_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? bias, Tensor? cu_seqlens_q, Tensor? cu_seqlens_k, int? max_seqlen_q, int? max_seqlen_k, float dropout_p, int custom_mask_type, bool compute_log_sumexp=False, *, float? scale=None, Tensor? causal_diagonal=None, Tensor? seqlen_k=None) -> (Tensor output, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset, SymInt max_seqlen_batch_q, SymInt max_seqlen_batch_k)
+- func: _efficient_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? bias, Tensor? cu_seqlens_q, Tensor? cu_seqlens_k, SymInt? max_seqlen_q, SymInt? max_seqlen_k, float dropout_p, int custom_mask_type, bool compute_log_sumexp=False, *, float? scale=None, Tensor? seqlen_k=None, int? window_size=None) -> (Tensor output, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset, SymInt max_seqlen_batch_q, SymInt max_seqlen_batch_k)
   variants: function
   dispatch:
     CUDA: _efficient_attention_forward
   tags: nondeterministic_seeded
 
-- func: _efficient_attention_backward(Tensor grad_out_, Tensor query, Tensor key, Tensor value, Tensor? bias, Tensor out, Tensor? cu_seqlens_q, Tensor? cu_seqlens_k, SymInt max_seqlen_q, SymInt max_seqlen_k, Tensor logsumexp, float dropout_p, Tensor philox_seed, Tensor philox_offset, int custom_mask_type, bool bias_requires_grad, *, float? scale=None, int? num_splits_key=None) -> (Tensor, Tensor, Tensor, Tensor)
+- func: _efficient_attention_backward(Tensor grad_out_, Tensor query, Tensor key, Tensor value, Tensor? bias, Tensor out, Tensor? cu_seqlens_q, Tensor? cu_seqlens_k, SymInt max_seqlen_q, SymInt max_seqlen_k, Tensor logsumexp, float dropout_p, Tensor philox_seed, Tensor philox_offset, int custom_mask_type, bool bias_requires_grad, *, float? scale=None, int? num_splits_key=None, int? window_size=None, bool shared_storage_dqdkdv=False) -> (Tensor, Tensor, Tensor, Tensor)
   device_check: NoCheck
   variants: function
   dispatch:
@@ -15460,11 +15557,11 @@
     CPU: foobar
   autogen: _foobar.out
 
-# Fused Optimizer CUDA kernels.
 - func: _fused_adam_(Tensor(a!)[] self, Tensor(b!)[] grads, Tensor(c!)[] exp_avgs, Tensor(d!)[] exp_avg_sqs, Tensor(e!)[] max_exp_avg_sqs, Tensor[] state_steps, *, float lr, float beta1, float beta2, float weight_decay, float eps, bool amsgrad, bool maximize, Tensor? grad_scale=None, Tensor? found_inf=None) -> ()
   # Unlike "foreach" functions, lists of tensors should be guaranteed to be on the same device (for now).
   variants: function
   dispatch:
+    CPU: _fused_adam_kernel_cpu_
     CUDA: _fused_adam_kernel_cuda_
   autogen: _fused_adam, _fused_adam.out
 
@@ -15474,6 +15571,7 @@
   device_check: NoCheck
   variants: function
   dispatch:
+    CPU: _fused_adam_kernel_cpu_
     CUDA: _fused_adam_kernel_cuda_
   autogen: _fused_adam.tensor_lr, _fused_adam.tensor_lr_out
 
@@ -15481,6 +15579,7 @@
   # Unlike "foreach" functions, lists of tensors should be guaranteed to be on the same device (for now).
   variants: function
   dispatch:
+    CPU: _fused_adamw_kernel_cpu_
     CUDA: _fused_adamw_kernel_cuda_
   autogen: _fused_adamw, _fused_adamw.out
 
@@ -15490,6 +15589,7 @@
   device_check: NoCheck
   variants: function
   dispatch:
+    CPU: _fused_adamw_kernel_cpu_
     CUDA: _fused_adamw_kernel_cuda_
   autogen: _fused_adamw.tensor_lr, _fused_adamw.tensor_lr_out
 
@@ -15497,6 +15597,7 @@
   # Unlike "foreach" functions, lists of tensors should be guaranteed to be on the same device (for now).
   variants: function
   dispatch:
+    CPU: _fused_sgd_kernel_cpu_
     CUDA: _fused_sgd_kernel_cuda_
   autogen: _fused_sgd, _fused_sgd.out
 
@@ -15506,9 +15607,16 @@
   device_check: NoCheck
   variants: function
   dispatch:
+    CPU: _fused_sgd_kernel_cpu_
     CUDA: _fused_sgd_kernel_cuda_
   autogen: _fused_sgd.tensor_lr, _fused_sgd.tensor_lr_out
 
+- func: _fused_adagrad_(Tensor(a!)[] self, Tensor(b!)[] grads, Tensor(c!)[] state_sums, Tensor(d!)[] state_steps, *, float lr, float lr_decay, float weight_decay, float eps, bool maximize, Tensor? grad_scale=None, Tensor? found_inf=None) -> ()
+  variants: function
+  dispatch:
+    CPU: _fused_adagrad_kernel_cpu_
+  autogen: _fused_adagrad, _fused_adagrad.out
+
 # This op is ONLY used by pytorch/XLA in functionalization, and should never show up in vanilla eager mode or in any pytorch tracing contexts.
 - func: _propagate_xla_data(Tensor input, Tensor output) -> ()
   variants: function