torch-rb 0.23.1 → 0.24.0

data/codegen/native_functions.yaml

@@ -418,11 +418,13 @@
   variants: function
   dispatch:
     CPU, CUDA, MPS, Meta: view_as_real
+    SparseCPU, SparseCUDA, SparseMPS: view_as_real_sparse
 
 - func: view_as_complex(Tensor(a) self) -> Tensor(a)
   variants: function
   dispatch:
     CPU, CUDA, MPS, Meta: view_as_complex
+    SparseCPU, SparseCUDA, SparseMPS: view_as_complex_sparse
 
 - func: sgn(Tensor self) -> Tensor
   variants: function, method
@@ -478,6 +480,7 @@
     CompositeExplicitAutograd: _conj_physical
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMPS, SparseCsrMeta: conj_physical_sparse_csr
   autogen: _conj_physical.out
+  tags: pointwise
 
 - func: conj_physical(Tensor self) -> Tensor
   variants: function, method
@@ -1089,11 +1092,13 @@
   variants: function
   dispatch:
     CUDA: _baddbmm_dtype_cuda
+    XPU: _baddbmm_dtype_xpu
 
 - func: baddbmm.dtype_out(Tensor self, Tensor batch1, Tensor batch2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
     CUDA: _baddbmm_out_dtype_cuda
+    XPU: _baddbmm_out_dtype_xpu
 
 - func: bartlett_window(int window_length, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
@@ -1318,7 +1323,7 @@
 - func: logical_xor.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: logical_xor_out
+    CPU, CUDA, MTIA: logical_xor_out
     MPS: logical_xor_out_mps
   tags: pointwise
 
@@ -1403,11 +1408,13 @@
   variants: function
   dispatch:
     CUDA: _bmm_dtype_cuda
+    XPU: _bmm_dtype_xpu
 
 - func: bmm.dtype_out(Tensor self, Tensor mat2, ScalarType out_dtype, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
     CUDA: _bmm_out_dtype_cuda
+    XPU: _bmm_out_dtype_xpu
 
 - func: broadcast_tensors(Tensor[] tensors) -> Tensor[]
   device_check: NoCheck
@@ -2311,6 +2318,7 @@
   dispatch:
     CPU: vdot
     CUDA: vdot_cuda
+    MPS: vdot_mps
 
 - func: vdot.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -2728,8 +2736,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: fill_
-    MPS: fill_scalar_mps
+    CPU, CUDA, MPS: fill_
     QuantizedCPU, QuantizedCUDA: fill_quantized_
     Meta: fill_meta_
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: fill_sparse_csr_
@@ -2740,8 +2747,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: fill_
-    MPS: fill_tensor_mps_
+    CPU, CUDA, MPS: fill_
     QuantizedCPU, QuantizedCUDA: fill_quantized_
     Meta: fill_meta_
     NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: fill_nested_
@@ -2870,7 +2876,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: gcd_out
+    CPU, CUDA, MPS: gcd_out
   tags: pointwise
 
 - func: gcd(Tensor self, Tensor other) -> Tensor
@@ -2941,6 +2947,8 @@
   autogen: _grid_sampler_2d_cpu_fallback.out
 
 - func: _grid_sampler_2d_cpu_fallback_backward(Tensor grad_output, Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners) -> (Tensor, Tensor)
+  dispatch:
+    CompositeExplicitAutograd: _grid_sampler_2d_cpu_fallback_backward
 
 - func: grid_sampler_3d(Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners) -> Tensor
   dispatch:
@@ -3043,6 +3051,7 @@
     CPU: _fft_c2r_mkl
     CUDA: _fft_c2r_cufft
     MPS: _fft_c2r_mps
+  tags: core
 
 - func: _fft_c2r.out(Tensor self, int[] dim, int normalization, SymInt last_dim_size, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
@@ -3288,6 +3297,42 @@
   device_guard: False
   manual_cpp_binding: True
 
+- func: numel(Tensor self) -> int
+  variants: method
+  device_check: NoCheck
+  device_guard: False
+  manual_cpp_binding: True
+
+- func: dim(Tensor self) -> int
+  variants: method
+  device_check: NoCheck
+  device_guard: False
+  manual_cpp_binding: True
+
+- func: get_device(Tensor self) -> int
+  variants: method
+  device_check: NoCheck
+  device_guard: False
+  manual_cpp_binding: True
+
+- func: storage_offset(Tensor self) -> int
+  variants: method
+  device_check: NoCheck
+  device_guard: False
+  manual_cpp_binding: True
+
+- func: is_contiguous(Tensor self) -> bool
+  variants: method
+  device_check: NoCheck
+  device_guard: False
+  manual_cpp_binding: True
+
+- func: is_contiguous.memory_format(Tensor self, MemoryFormat memory_format) -> bool
+  variants: 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
 
 - func: kron(Tensor self, Tensor other) -> Tensor
@@ -3342,11 +3387,13 @@
   dispatch:
     CUDA: _fused_rms_norm_cuda
     MPS: _fused_rms_norm_mps
+    XPU: _fused_rms_norm_xpu
     CompositeImplicitAutograd: rms_norm_composite
 
 - func: _fused_rms_norm_backward(Tensor grad_out, Tensor input, int[] normalized_shape, Tensor rstd, Tensor? weight, bool[2] output_mask) -> (Tensor, Tensor)
   dispatch:
     CUDA: _fused_rms_norm_backward_cuda
+    XPU: _fused_rms_norm_backward_xpu
 
 - func: nan_to_num(Tensor self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor
   variants: function, method
@@ -3476,13 +3523,20 @@
 
 - func: ldexp.Tensor(Tensor self, Tensor other) -> Tensor
   variants: function, method
+  tags: pointwise
+  dispatch:
+    CompositeExplicitAutograd: ldexp
 
 - func: ldexp_(Tensor(a!) self, Tensor other) -> Tensor(a!)
   variants: function, method
   tags: pointwise
+  dispatch:
+    CompositeExplicitAutograd: ldexp_
 
 - func: ldexp.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   tags: pointwise
+  dispatch:
+    CompositeExplicitAutograd: ldexp_out
 
 - func: linspace(Scalar start, Scalar end, int steps, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
@@ -3676,8 +3730,7 @@
   structured_inherits: TensorIteratorBase
   variants: function
   dispatch:
-    CPU, CUDA: xlogy_out
-    MPS: xlogy_out_mps
+    CPU, CUDA, MPS: xlogy_out
   tags: pointwise
 
 - func: xlogy.OutScalar_Self(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
@@ -3871,7 +3924,7 @@
   device_check: NoCheck   # TensorIterator
   structured: True
   dispatch:
-    CPU, CUDA, MTIA: aminmax_out
+    CPU, CUDA: aminmax_out
     MPS: aminmax_out_mps
   tags: reduction
 
@@ -3926,7 +3979,7 @@
 - func: amax.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
-    CPU, CUDA, MTIA: amax_out
+    CPU, CUDA: amax_out
     MPS: amax_out_mps
   tags: reduction
 
@@ -4115,7 +4168,7 @@
 - func: amin.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
   dispatch:
-    CPU, CUDA, MTIA: amin_out
+    CPU, CUDA: amin_out
     MPS: amin_out_mps
   tags: reduction
 
@@ -4193,6 +4246,27 @@
     CUDA: miopen_rnn_backward
   autogen: miopen_rnn_backward.out
 
+- func: _use_miopen_ctc_loss(Tensor log_probs, Tensor targets, int[] input_lengths, int[] target_lengths, int blank) -> bool
+  device_check: NoCheck  # Tensor arguments allowed to be on different devices, see also miopen_ctc_loss
+  dispatch:
+    CUDA: _use_miopen_ctc_loss
+
+- func: _use_miopen_ctc_loss.Tensor(Tensor log_probs, Tensor targets, Tensor input_lengths, Tensor target_lengths, int blank) -> bool
+  device_check: NoCheck  # Tensor arguments allowed to be on different devices, see also miopen_ctc_loss
+  dispatch:
+    CUDA: _use_miopen_ctc_loss_tensor
+
+- func: miopen_ctc_loss(Tensor log_probs, Tensor targets, int[] input_lengths, int[] target_lengths, int blank, bool deterministic, bool zero_infinity) -> (Tensor, Tensor)
+  device_check: NoCheck  # log_probs is expected to be on CUDA while targets is expected to be on CPU
+  dispatch:
+    CUDA: miopen_ctc_loss
+  autogen: miopen_ctc_loss.out
+
+- func: miopen_ctc_loss.Tensor(Tensor log_probs, Tensor targets, Tensor input_lengths, Tensor target_lengths, int blank, bool deterministic, bool zero_infinity) -> (Tensor, Tensor)
+  device_check: NoCheck  # log_probs is expected to be on CUDA while targets is expected to be on CPU
+  dispatch:
+    CUDA: miopen_ctc_loss_tensor
+
 - func: mm(Tensor self, Tensor mat2) -> Tensor
   structured_delegate: mm.out
   variants: function, method
@@ -4215,10 +4289,12 @@
 - func: mm.dtype(Tensor self, Tensor mat2, ScalarType out_dtype) -> Tensor
   dispatch:
     CUDA: _mm_dtype_cuda
+    XPU: _mm_dtype_xpu
 
 - func: mm.dtype_out(Tensor self, Tensor mat2, ScalarType out_dtype, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CUDA: _mm_dtype_out_cuda
+    XPU: _mm_dtype_out_xpu
 
 - func: _int_mm(Tensor self, Tensor mat2) -> Tensor
   dispatch:
@@ -4381,7 +4457,7 @@
 
 - func: mvlgamma.out(Tensor self, int p, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU, CUDA: mvlgamma_out
+    CPU, CUDA, MPS: mvlgamma_out
   tags: pointwise
 
 - func: mvlgamma(Tensor self, int p) -> Tensor
@@ -4993,8 +5069,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA, MTIA: reciprocal_out
-    MPS: reciprocal_out_mps
+    CPU, CUDA, MPS, MTIA: reciprocal_out
   tags: pointwise
 
 - func: neg(Tensor self) -> Tensor
@@ -5344,6 +5419,7 @@
 
 - func: selu_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
+  tags: pointwise
 
 - func: celu(Tensor self, Scalar alpha=1.0) -> Tensor
   device_check: NoCheck   # TensorIterator
@@ -5356,6 +5432,7 @@
   dispatch:
     CompositeExplicitAutograd: celu_
   autogen: celu.out
+  tags: pointwise
 
 - func: silu(Tensor self) -> Tensor
   structured_delegate: silu.out
@@ -5376,8 +5453,7 @@
   structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA, MTIA: silu_out
-    MPS: silu_out_mps
+    CPU, CUDA, MPS, MTIA: silu_out
   tags: pointwise
 
 - func: silu_backward.grad_input(Tensor grad_output, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
@@ -5385,8 +5461,7 @@
   structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA: silu_backward_out
-    MPS: silu_backward_out_mps
+    CPU, CUDA, MPS: silu_backward_out
   tags: pointwise
 
 - func: silu_backward(Tensor grad_output, Tensor self) -> Tensor
@@ -6532,6 +6607,7 @@
   dispatch:
     CPU: _unique_cpu
     CUDA: _unique_cuda
+    MPS: _unique_mps
   autogen: _unique.out
 
 - func: unique_dim(Tensor self, int dim, bool sorted=True, bool return_inverse=False, bool return_counts=False) -> (Tensor, Tensor, Tensor)
@@ -6618,7 +6694,6 @@
   dispatch:
     CPU, CUDA: var
     MPS: var_mps
-    MTIA: var_mtia
   tags: [core, reduction]
 
 - func: var.out(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
@@ -6780,6 +6855,7 @@
   dispatch:
     CPU: _standard_gamma_grad_cpu
     CUDA: _standard_gamma_grad_cuda
+    MPS: _standard_gamma_grad_mps
   autogen: _standard_gamma_grad.out
 
 - func: _standard_gamma(Tensor self, Generator? generator=None) -> Tensor
@@ -6787,9 +6863,32 @@
   dispatch:
     CPU: _s_gamma_cpu
     CUDA: _s_gamma_cuda
+    MPS: _s_gamma_mps
   tags: nondeterministic_seeded
   autogen: _standard_gamma.out
 
+- func: _philox_key_split(Tensor key, int num_splits) -> Tensor
+  variants: function
+  dispatch:
+    CUDA: _philox_key_split_cuda
+
+- func: _philox_key_fold_in(Tensor key, int data) -> Tensor
+  variants: function
+  dispatch:
+    CUDA: _philox_key_fold_in_cuda
+
+- func: _philox_normal_(Tensor(a!) self, Tensor key, float mean=0, float std=1) -> Tensor(a!)
+  variants: function, method
+  dispatch:
+    CUDA: _philox_normal_cuda_
+  autogen: _philox_normal, _philox_normal.out
+
+- func: _philox_uniform_(Tensor(a!) self, Tensor key, float low=0, float high=1) -> Tensor(a!)
+  variants: function, method
+  dispatch:
+    CUDA: _philox_uniform_cuda_
+  autogen: _philox_uniform, _philox_uniform.out
+
 - func: _dirichlet_grad(Tensor x, Tensor alpha, Tensor total) -> Tensor
   dispatch:
     CPU: _dirichlet_grad_cpu
@@ -6978,16 +7077,14 @@
   structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: norm_dtype_out
-    MPS: norm_dtype_out_mps
+    CPU, CUDA, MPS: norm_dtype_out
   tags: reduction
 
 - 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
-    MPS: norm_out_mps
+    CPU, CUDA, MPS: norm_out
   tags: reduction
 
 # These four redispatch in their implementation, so OK to be CompositeImplicitAutograd
@@ -7080,8 +7177,7 @@
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA: zero_
-    MPS: zero_mps_
+    CPU, CUDA, MPS: zero_
     Meta: zero_meta_
     SparseCPU, SparseCUDA, SparseMPS, SparseMeta: zero_sparse_
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: zero_sparse_csr_
@@ -7242,10 +7338,12 @@
 - func: addmm.dtype(Tensor self, Tensor mat1, Tensor mat2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   dispatch:
     CUDA: _addmm_dtype_cuda
+    XPU: _addmm_dtype_xpu
 
 - func: addmm.dtype_out(Tensor self, Tensor mat1, Tensor mat2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CUDA: _addmm_dtype_out_cuda
+    XPU: _addmm_dtype_out_xpu
 
 - func: addmm_(Tensor(a!) self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
   structured_delegate: addmm.out
@@ -7287,14 +7385,18 @@
 - func: _scaled_mm_v2(Tensor self, Tensor mat2, Tensor[] scale_a, int[] recipe_a, int[] swizzle_a, Tensor[] scale_b, int[] recipe_b, int[] swizzle_b, Tensor? bias, ScalarType? out_dtype, int[] contraction_dim=[], bool use_fast_accum=False) -> Tensor
   variants: function
   dispatch:
+    CPU: _scaled_mm_cpu_v2
     CUDA: _scaled_mm_cuda_v2
     XPU: _scaled_mm_xpu_v2
+  tags: needs_exact_strides
 
 - func: _scaled_mm_v2.out(Tensor self, Tensor mat2, Tensor[] scale_a, int[] recipe_a, int[] swizzle_a, Tensor[] scale_b, int[] recipe_b, int[] swizzle_b, Tensor? bias, ScalarType? out_dtype, int[] contraction_dim=[], bool use_fast_accum=False, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
+    CPU: _scaled_mm_cpu_v2_out
     CUDA: _scaled_mm_cuda_v2_out
     XPU: _scaled_mm_xpu_v2_out
+  tags: needs_exact_strides
 
 
 - func: _scaled_grouped_mm(Tensor self, Tensor mat2, Tensor scale_a, Tensor scale_b, Tensor? offs=None, Tensor? bias=None, Tensor? scale_result=None, ScalarType? out_dtype=None, bool use_fast_accum=False) -> Tensor
@@ -8380,6 +8482,7 @@
   dispatch:
     CPU: index_reduce_cpu_out
     CUDA: index_reduce_cuda_out
+    MPS: index_reduce_mps_out
 
 - func: index_reduce_(Tensor(a!) self, int dim, Tensor index, Tensor source, str reduce, *, bool include_self=True) -> Tensor(a!)
   structured_delegate: index_reduce.out
@@ -8393,9 +8496,7 @@
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
-    CPU: index_fill_
-    CUDA: index_fill_
-    MPS: index_fill_mps_
+    CPU, CUDA, MPS: index_fill_
   autogen: index_fill.int_Scalar_out
 
 - func: index_fill.int_Scalar(Tensor self, int dim, Tensor index, Scalar value) -> Tensor
@@ -8408,8 +8509,7 @@
   device_check: NoCheck   # TensorIterator
   variants: method
   dispatch:
-    CPU, CUDA: index_fill_
-    MPS: index_fill_mps_
+    CPU, CUDA, MPS: index_fill_
   autogen: index_fill.int_Tensor_out
 
 - func: index_fill.int_Tensor(Tensor self, int dim, Tensor index, Tensor value) -> Tensor
@@ -8755,14 +8855,14 @@
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA, MPS: __lshift__
+    CPU, CUDA, MPS, MTIA: __lshift__
   tags: pointwise
 
 - func: __lshift__.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA, MPS: __lshift__
+    CPU, CUDA, MPS, MTIA: __lshift__
   tags: pointwise
 
 - func: __ilshift__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
@@ -8834,14 +8934,14 @@
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA, MPS: __rshift__
+    CPU, CUDA, MPS, MTIA: __rshift__
   tags: pointwise
 
 - func: __rshift__.Tensor(Tensor self, Tensor other) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA, MPS: __rshift__
+    CPU, CUDA, MPS, MTIA: __rshift__
   tags: pointwise
 
 - func: __irshift__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
@@ -8996,6 +9096,7 @@
   tags: nondeterministic_seeded
   dispatch:
     CPU, CUDA: cauchy_
+    MPS: cauchy_mps_
   autogen: cauchy, cauchy.out
 
 - func: log_normal_(Tensor(a!) self, float mean=1, float std=2, *, Generator? generator=None) -> Tensor(a!)
@@ -9004,6 +9105,7 @@
   variants: method
   dispatch:
     CPU, CUDA: log_normal_
+    MPS: log_normal_mps_
   autogen: log_normal, log_normal.out
 
 - func: exponential_(Tensor(a!) self, float lambd=1, *, Generator? generator=None) -> Tensor(a!)
@@ -9021,6 +9123,7 @@
   variants: method
   dispatch:
     CPU, CUDA: geometric_
+    MPS: geometric_mps_
 
   # wrappers for TH functions
   autogen: geometric, geometric.out
@@ -9525,12 +9628,14 @@
   dispatch:
     CPU: nonzero_static_out_cpu
     CUDA: nonzero_static_out_cuda
+    MPS: nonzero_static_out_mps
 
 - func: nonzero_static(Tensor self, *, SymInt size, int fill_value=-1) -> Tensor
   variants: method, function
   dispatch:
     CPU: nonzero_static_cpu
     CUDA: nonzero_static_cuda
+    MPS: nonzero_static_mps
 
 - func: nonzero_numpy(Tensor self) -> Tensor[]
   variants: method, function
@@ -9695,16 +9800,17 @@
   dispatch:
     CPU: _cholesky_solve_helper_cpu
     CUDA: _cholesky_solve_helper_cuda
+    MPS: _cholesky_solve_helper_mps
   autogen: _cholesky_solve_helper.out
 
 - func: cholesky_inverse(Tensor self, bool upper=False) -> Tensor
   variants: method, function
   dispatch:
-    CPU, CUDA: cholesky_inverse
+    CPU, CUDA, MPS: cholesky_inverse
 
 - func: cholesky_inverse.out(Tensor self, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU, CUDA: cholesky_inverse_out
+    CPU, CUDA, MPS: cholesky_inverse_out
 
 - func: qr.Q(Tensor self, bool some=True, *, Tensor(a!) Q, Tensor(b!) R) -> (Tensor(a!) Q, Tensor(b!) R)
 
@@ -9773,8 +9879,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: lgamma_out
-    MPS: lgamma_out_mps
+    CPU, CUDA, MPS: lgamma_out
   tags: pointwise
 
 - func: lgamma_(Tensor(a!) self) -> Tensor(a!)
@@ -9794,8 +9899,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: digamma_out
-    MPS: digamma_out_mps
+    CPU, CUDA, MPS: digamma_out
   tags: pointwise
 
 - func: digamma(Tensor self) -> Tensor
@@ -9809,8 +9913,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: polygamma_out
-    MPS: polygamma_out_mps
+    CPU, CUDA, MPS: polygamma_out
   tags: pointwise
 
 - func: polygamma(int n, Tensor self) -> Tensor
@@ -9931,8 +10034,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: atan2_out
-    MPS: atan2_out_mps
+    CPU, CUDA, MPS: atan2_out
   tags: [core, pointwise]
 
 - func: atan2_(Tensor(a!) self, Tensor other) -> Tensor(a!)
@@ -9970,8 +10072,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: lerp_Tensor
-    MPS: lerp_Tensor_mps
+    CPU, CUDA, MPS: lerp_Tensor
   tags: pointwise
 
 - func: lerp.Scalar(Tensor self, Tensor end, Scalar weight) -> Tensor
@@ -10256,8 +10357,7 @@
   structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA: maximum_out
-    MPS: maximum_out_mps
+    CPU, CUDA, MTIA, MPS: maximum_out
   tags: pointwise
 
 # binary max, alias of maximum
@@ -10289,8 +10389,7 @@
   structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA: minimum_out
-    MPS: minimum_out_mps
+    CPU, CUDA, MTIA, MPS: minimum_out
   tags: pointwise
 
 # binary min, alias for minimum
@@ -10496,9 +10595,8 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: pow_Tensor_Scalar_out
+    CPU, CUDA, MPS: pow_Tensor_Scalar_out
     SparseCPU, SparseCUDA, SparseMPS: pow_out_sparse_scalar
-    MPS: pow_tensor_scalar_out_mps
   tags: pointwise
 
 - func: pow.Tensor_Scalar(Tensor self, Scalar exponent) -> Tensor
@@ -11524,6 +11622,15 @@
     MTIA: foreach_tensor_norm_mtia
   autogen: _foreach_norm.Scalar_out
 
+# Like _foreach_norm but returns sum(|x|^ord) without the final root
+- func: _foreach_powsum.Scalar(Tensor[] self, Scalar ord=2, ScalarType? dtype=None) -> Tensor[]
+  device_check: NoCheck
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: foreach_tensor_powsum_slow
+    CUDA: foreach_tensor_powsum_cuda
+  autogen: _foreach_powsum.Scalar_out
+
 - func: _foreach_pow.List(Tensor[] self, Tensor[] exponent) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
@@ -11750,6 +11857,14 @@
     CUDA: foreach_tensor_zero_cuda_
   autogen: _foreach_zero, _foreach_zero.out
 
+- func: _foreach_clone(Tensor[] self, *, MemoryFormat? memory_format=None) -> Tensor[]
+  device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: foreach_tensor_clone_slow
+    CUDA: foreach_tensor_clone_cuda
+  autogen: _foreach_clone.out
+
 - func: _foreach_copy_(Tensor(a!)[] self, Tensor[] src, bool non_blocking=False) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
@@ -12083,6 +12198,7 @@
   structured_delegate: elu.out
   device_check: NoCheck   # TensorIterator
   python_module: nn
+  tags: pointwise
 
 - func: glu.out(Tensor self, int dim=-1, *, Tensor(a!) out) -> Tensor(a!)
   structured: True
@@ -12144,6 +12260,7 @@
   structured_delegate: hardsigmoid.out
   device_check: NoCheck   # TensorIterator
   python_module: nn
+  tags: pointwise
 
 - func: hardsigmoid_backward.grad_input(Tensor grad_output, Tensor self, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
@@ -12189,6 +12306,7 @@
   dispatch:
     CPU, CUDA, MPS: hardtanh_
     QuantizedCPU: hardtanh_quantized_cpu_
+  tags: pointwise
 
 - func: hardswish.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -12229,7 +12347,7 @@
   python_module: nn
   dispatch:
     QuantizedCPU: leaky_relu_quantized_cpu
-  tags: core
+  tags: [core, pointwise]
 
 - 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
@@ -12248,6 +12366,7 @@
   python_module: nn
   dispatch:
     QuantizedCPU: leaky_relu_quantized_cpu_
+  tags: pointwise
 
 - func: log_sigmoid.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -12904,6 +13023,10 @@
   python_module: nn
   autogen: _upsample_bicubic2d_aa.vec_out
 
+- func: _upsample_lanczos2d_aa.vec(Tensor input, SymInt[]? output_size, bool align_corners, float[]? scale_factors) -> Tensor
+  python_module: nn
+  autogen: _upsample_lanczos2d_aa.vec_out
+
 - func: upsample_nearest1d.vec(Tensor input, SymInt[]? output_size, float[]? scale_factors) -> Tensor
   python_module: nn
   autogen: upsample_nearest1d.vec_out
@@ -13050,6 +13173,26 @@
   python_module: nn
   structured_delegate: _upsample_bicubic2d_aa_backward.grad_input
 
+- func: _upsample_lanczos2d_aa.out(Tensor self, SymInt[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_lanczos2d_aa_out_cpu
+
+- func: _upsample_lanczos2d_aa(Tensor self, SymInt[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_lanczos2d_aa.out
+
+- func: _upsample_lanczos2d_aa_backward.grad_input(Tensor grad_output, SymInt[2] output_size, SymInt[4] input_size, bool align_corners, float? scales_h=None, float? scales_w=None, *, Tensor(a!) grad_input) -> Tensor(a!)
+  python_module: nn
+  structured: True
+  dispatch:
+    CPU: _upsample_lanczos2d_aa_backward_out_cpu
+
+- func: _upsample_lanczos2d_aa_backward(Tensor grad_output, SymInt[2] output_size, SymInt[4] input_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
+  python_module: nn
+  structured_delegate: _upsample_lanczos2d_aa_backward.grad_input
+
 - func: upsample_trilinear3d.out(Tensor self, SymInt[3] output_size, bool align_corners, float? scales_d=None, float? scales_h=None, float? scales_w=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   structured: True
@@ -13608,7 +13751,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: special_erfcx_out
+    CPU, CUDA, MPS: special_erfcx_out
   tags: pointwise
 
 - func: special_erfinv(Tensor self) -> Tensor
@@ -14471,8 +14614,18 @@
   python_module: linalg
   structured: True
   dispatch:
-    CPU, CUDA: linalg_vector_norm_out
-    MPS: linalg_vector_norm_out_mps
+    CPU, CUDA, MPS: linalg_vector_norm_out
+  tags: reduction
+
+# Computes sum(|x|^ord) - the "power sum" without the final root.
+# This is useful for distributed computing where partial power sums
+# can be reduced across shards before taking the final root.
+- func: linalg__powsum(Tensor self, Scalar ord=2, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
+  python_module: linalg
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: linalg__powsum_slow
+    CPU, CUDA: linalg__powsum
   tags: reduction
 
 - func: linalg_matrix_norm(Tensor self, Scalar ord, int[] dim=[-2,-1], bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
@@ -14622,6 +14775,7 @@
   structured: True
   dispatch:
     CPU, CUDA: linalg_qr_out
+    MPS: linalg_qr_out_mps
 
 - func: linalg_matrix_power(Tensor self, int n) -> Tensor
   python_module: linalg
@@ -15122,7 +15276,7 @@
   variants: function
   tags: nondeterministic_seeded
 
-- func: _scaled_dot_product_attention_math_for_mps(Tensor query, Tensor key, Tensor value, Tensor? attn_mask=None, float dropout_p=0.0, bool is_causal=False, Tensor? dropout_mask=None, *, float? scale=None) -> (Tensor, Tensor)
+- func: _scaled_dot_product_attention_math_for_mps(Tensor query, Tensor key, Tensor value, Tensor? attn_mask=None, float dropout_p=0.0, bool is_causal=False, Tensor? dropout_mask=None, *, float? scale=None, bool enable_gqa=False) -> (Tensor, Tensor)
   dispatch:
     MPS: _scaled_dot_product_attention_math_mps
   tags: nondeterministic_seeded
@@ -15134,6 +15288,11 @@
     NestedTensorCUDA: _scaled_dot_product_flash_attention_nestedtensor_cuda
   tags: nondeterministic_seeded
 
+- func: _scaled_dot_product_flash_attention.quantized(Tensor query, Tensor key, Tensor value, Tensor? q_descale, Tensor? k_descale, Tensor? v_descale, 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 rng_state, Tensor unused, Tensor debug_attn_mask)
+  dispatch:
+    CUDA: _scaled_dot_product_flash_attention_cuda_quantized
+  tags: nondeterministic_seeded
+
 - func: _scaled_dot_product_flash_attention_for_cpu(Tensor query, Tensor key, Tensor value, float dropout_p=0.0, bool is_causal=False, *, Tensor? attn_mask=None, float? scale=None) -> (Tensor output, Tensor logsumexp)
   dispatch:
     CPU: _scaled_dot_product_flash_attention_cpu
@@ -15189,12 +15348,24 @@
     NestedTensorCUDA: _scaled_dot_product_cudnn_attention_nestedtensor_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 rng_state, Tensor unused, Tensor debug_attn_mask)
+- 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? block_table=None, int? num_splits=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor rng_state, Tensor unused, Tensor debug_attn_mask)
   variants: function
   dispatch:
     CUDA: _flash_attention_forward
   tags: nondeterministic_seeded
 
+- func: _flash_attention_forward_no_dropout_inplace(Tensor(a!) out, 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? block_table=None, int? num_splits=None) -> Tensor softmax_logsumexp
+  variants: function
+  dispatch:
+    CUDA: _flash_attention_forward_no_dropout_inplace
+  tags: nondeterministic_seeded
+
+- func: _flash_attention_forward.quantized(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, Tensor? q_descale, Tensor? k_descale, Tensor? v_descale, *, 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 rng_state, Tensor unused, Tensor debug_attn_mask)
+  variants: function
+  dispatch:
+    CUDA: _flash_attention_forward_quantized
+  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 rng_state, Tensor unused, *, float? scale=None, SymInt? window_size_left=None, SymInt? window_size_right=None) -> (Tensor, Tensor, Tensor)
   device_check: NoCheck
   variants: function