torch-rb 0.18.0 → 0.19.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 0f044a5934fa8a296fe407fd79d3b9ac7e5e582b110df6e024ce01fe88d4dbd6
-  data.tar.gz: ac662c6a27bba6c173f9631b16e8832d24cfb72af9b4d845780b9c717404f923
+  metadata.gz: 03164cc479d8f8a32f0669d597e8fe5310d91955e6954cfdc0fffdc8983c5768
+  data.tar.gz: 87fc733016b6f4489b38a419a3879cacbdb1e190cfaa5c02397aceb57c012d16
 SHA512:
-  metadata.gz: 8da9c0bb9d466a81483a31b63d8c7670d14025bc364d70461e9a26264ac0567ae10af7ef464d43d6cfc4e00bca5701e846093bc6651ded5c84ef22e6652a59a2
-  data.tar.gz: 4420bf303e8ef0ed9b96c607eead13b33aa9f323d0422015868b93d28a1ccd1dd9999053792c04d8810e9dcc3ff8f4594c99587b033653031882f3d1c25310ff
+  metadata.gz: 6ba0480138a10ba43dff625dc1bcf99e2287f238dc4607ea6813e82914f1e335133f55408fa59579343b161c54850316f744c594cf6687b7f2de64a0d71746d1
+  data.tar.gz: 859015641dd14bf919a7982c6673acb296f858518552b8c924fc7e59b9c1b2a9491aa598c01b019b392c8c2bba7b9f65ff0923f838e6cbde7ddaede9c4b69191

data/CHANGELOG.md

@@ -1,3 +1,10 @@
+## 0.19.0 (2025-01-29)
+
+- Updated LibTorch to 2.6.0
+- Improved `inspect` for `Device`
+- Fixed equality for `Device`
+- Fixed `index` method for `Device` when no index
+
 ## 0.18.0 (2024-10-22)
 
 - Updated LibTorch to 2.5.0

data/README.md

@@ -17,7 +17,7 @@ Check out:
 First, [download LibTorch](https://pytorch.org/get-started/locally/). For Mac arm64, use:
 
 ```sh
-curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.5.0.zip > libtorch.zip
+curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.6.0.zip > libtorch.zip
 unzip -q libtorch.zip
 ```
 
@@ -413,6 +413,7 @@ Here’s the list of compatible versions.
 
 Torch.rb | LibTorch
 --- | ---
+0.19.x | 2.6.x
 0.18.x | 2.5.x
 0.17.x | 2.4.x
 0.16.x | 2.3.x

data/codegen/generate_functions.rb

@@ -53,7 +53,9 @@ def skip_functions(functions)
     f.base_name == "sym_size" ||
     f.base_name == "sym_numel" ||
     f.base_name == "sym_storage_offset" ||
-    f.base_name == "sym_stride"
+    f.base_name == "sym_stride" ||
+    # TODO fix LibTorch 2.6 changes
+    f.base_name == "rrelu_with_noise"
   end
 end
 

data/codegen/native_functions.yaml

@@ -187,7 +187,10 @@
   dispatch:
     CPU: _functional_assert_async_msg_cpu
 
-- func: _assert_tensor_metadata(Tensor a, SymInt[]? size=None, SymInt[]? stride=None, ScalarType? dtype=None) -> ()
+- func: _assert_tensor_metadata(Tensor a, SymInt[]? size=None, SymInt[]? stride=None, ScalarType? dtype=None, *, Device? device=None, Layout? layout=None) -> ()
+  dispatch:
+    CompositeExplicitAutograd: _assert_tensor_metadata
+    Meta: _assert_tensor_metadata_meta_symint
 
 - func: _print(str s) -> ()
   dispatch:
@@ -309,25 +312,25 @@
 - func: _shape_as_tensor(Tensor self) -> Tensor
 
 - func: dropout(Tensor input, float p, bool train) -> Tensor
-  tags: nondeterministic_seeded
+  tags: [nondeterministic_seeded, maybe_aliasing_or_mutating]
 
 - func: dropout_(Tensor(a!) self, float p, bool train) -> Tensor(a!)
   tags: nondeterministic_seeded
 
 - func: feature_dropout(Tensor input, float p, bool train) -> Tensor
-  tags: nondeterministic_seeded
+  tags: [nondeterministic_seeded, maybe_aliasing_or_mutating]
 
 - func: feature_dropout_(Tensor(a!) self, float p, bool train) -> Tensor(a!)
   tags: nondeterministic_seeded
 
 - func: alpha_dropout(Tensor input, float p, bool train) -> Tensor
-  tags: nondeterministic_seeded
+  tags: [nondeterministic_seeded, maybe_aliasing_or_mutating]
 
 - func: alpha_dropout_(Tensor(a!) self, float p, bool train) -> Tensor(a!)
   tags: nondeterministic_seeded
 
 - func: feature_alpha_dropout(Tensor input, float p, bool train) -> Tensor
-  tags: nondeterministic_seeded
+  tags: [nondeterministic_seeded, maybe_aliasing_or_mutating]
 
 - func: feature_alpha_dropout_(Tensor(a!) self, float p, bool train) -> Tensor(a!)
   tags: nondeterministic_seeded
@@ -477,7 +480,7 @@
 
 - func: conj_physical(Tensor self) -> Tensor
   variants: function, method
-  tags: pointwise
+  tags: [pointwise, maybe_aliasing_or_mutating]
 
 - func: conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -641,6 +644,7 @@
     CPU: addmv_out_cpu
     CUDA: addmv_out_cuda
     MPS: addmv_out_mps
+    XPU: addmv_out_xpu
     SparseCsrCPU: addmv_out_sparse_compressed
     SparseCsrCUDA: addmv_out_sparse_compressed_cuda
 
@@ -1031,17 +1035,20 @@
 
 - func: atleast_1d(Tensor self) -> Tensor
   variants: function
+  tags: maybe_aliasing_or_mutating
 
 - func: atleast_1d.Sequence(Tensor[] tensors) -> Tensor[]
 
 - func: atleast_2d(Tensor self) -> Tensor
   variants: function
+  tags: maybe_aliasing_or_mutating
 
 - func: atleast_2d.Sequence(Tensor[] tensors) -> Tensor[]
   variants: function
 
 - func: atleast_3d(Tensor self) -> Tensor
   variants: function
+  tags: maybe_aliasing_or_mutating
 
 - func: atleast_3d.Sequence(Tensor[] tensors) -> Tensor[]
   variants: function
@@ -1061,6 +1068,7 @@
     CPU: baddbmm_out_cpu
     CUDA: baddbmm_out_cuda
     MPS: baddbmm_out_mps
+    XPU: baddbmm_out_xpu
     SparseCsrCUDA: baddbmm_out_sparse_csr_cuda
 
 - func: bartlett_window(int window_length, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
@@ -1074,6 +1082,7 @@
   autogen: bartlett_window.periodic_out
 
 - func: batch_norm(Tensor input, Tensor? weight, Tensor? bias, Tensor? running_mean, Tensor? running_var, bool training, float momentum, float eps, bool cudnn_enabled) -> Tensor
+  tags: maybe_aliasing_or_mutating
 
 - func: quantized_batch_norm(Tensor input, Tensor? weight, Tensor? bias, Tensor mean, Tensor var, float eps, float output_scale, int output_zero_point) -> Tensor
   dispatch:
@@ -1081,6 +1090,7 @@
   autogen: quantized_batch_norm.out
 
 - func: _batch_norm_impl_index(Tensor input, Tensor? weight, Tensor? bias, Tensor? running_mean, Tensor? running_var, bool training, float momentum, float eps, bool cudnn_enabled) -> (Tensor, Tensor, Tensor, Tensor, int)
+  tags: maybe_aliasing_or_mutating
 
 - func: _batch_norm_impl_index_backward(int impl_index, Tensor input, Tensor grad_output, Tensor? weight, Tensor? running_mean, Tensor? running_var, Tensor? save_mean, Tensor? save_var_transform, bool train, float eps, bool[3] output_mask, Tensor reservedSpace) -> (Tensor, Tensor, Tensor)
 
@@ -1358,6 +1368,7 @@
     CPU: bmm_out_cpu
     CUDA: bmm_out_cuda
     MPS: bmm_out_mps
+    XPU: bmm_out_xpu
     SparseCPU: bmm_out_sparse_cpu
     SparseCUDA: bmm_out_sparse_cuda
     SparseCsrCUDA: bmm_out_sparse_csr_cuda
@@ -1462,6 +1473,7 @@
   variants: function, method
   device_check: NoCheck
   device_guard: False
+  tags: maybe_aliasing_or_mutating
 
 - func: chunk(Tensor(a -> *) self, int chunks, int dim=0) -> Tensor(a)[]
   variants: function, method
@@ -1788,7 +1800,7 @@
   variants: function, method
   structured_delegate: cos.out
   dispatch:
-    NestedTensorCPU, NestedTensorCUDA: cos_nested
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_cos
   tags: [core, pointwise]
 
 - func: cos_(Tensor(a!) self) -> Tensor(a!)
@@ -2821,6 +2833,7 @@
     # non-differentiable so NonFunctional doesn't apply
     CompositeExplicitAutograd: full_like
   autogen: full_like.out
+  tags: core
 
 - func: from_file(str filename, bool? shared=None, int? size=0, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
@@ -3179,6 +3192,7 @@
   device_guard: False
   dispatch:
     CPU, CUDA, MPS: isnan
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_isnan
     SparseCPU, SparseCUDA: isnan_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: isnan_sparse_csr
   autogen: isnan.out
@@ -3289,7 +3303,9 @@
   autogen: native_layer_norm_backward.out
   tags: core
 
-- func: rms_norm(Tensor input, int[] normalized_shape, Tensor? weight=None, float? eps=None) -> Tensor
+- func: rms_norm(Tensor input, SymInt[] normalized_shape, Tensor? weight=None, float? eps=None) -> Tensor
+  dispatch:
+    CompositeImplicitAutograd: rms_norm_symint
 
 - func: nan_to_num(Tensor self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor
   variants: function, method
@@ -3355,9 +3371,10 @@
   dispatch:
     CUDA: _cslt_compress
 
-- func: _cslt_sparse_mm(Tensor compressed_A, Tensor dense_B, Tensor? bias=None, Tensor? alpha=None, ScalarType? out_dtype=None, bool transpose_result=False, int alg_id=0) -> Tensor
+- func: _cslt_sparse_mm(Tensor compressed_A, Tensor dense_B, Tensor? bias=None, Tensor? alpha=None, ScalarType? out_dtype=None, bool transpose_result=False, int alg_id=0, int split_k=1, bool split_k_one_kernel=True) -> Tensor
   dispatch:
     CUDA: _cslt_sparse_mm
+  tags: needs_fixed_stride_order
 
 - func: _cslt_sparse_mm_search(Tensor compressed_A, Tensor dense_B, Tensor? bias=None, Tensor? alpha=None, ScalarType? out_dtype=None, bool transpose_result=False) -> int
   dispatch:
@@ -4126,6 +4143,7 @@
     CPU: mm_out_cpu
     CUDA: mm_out_cuda
     MPS: mm_out_mps
+    XPU: mm_out_xpu
     SparseCPU, SparseCUDA: _sparse_mm_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: _sparse_csr_mm_out
 
@@ -4141,16 +4159,24 @@
 
 - func: _convert_weight_to_int4pack(Tensor self, int innerKTiles) -> Tensor
   dispatch:
-    CPU: _convert_weight_to_int4pack_cpu
     CUDA: _convert_weight_to_int4pack_cuda
     MPS: _convert_weight_to_int4pack_mps
 
 - 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
 
+# Split int4 pack weight between cpu and other devices due to
+# https://github.com/pytorch/ao/issues/1117#issuecomment-2451252756.
+- func: _convert_weight_to_int4pack_for_cpu(Tensor self, int innerKTiles) -> Tensor
+  dispatch:
+    CPU: _convert_weight_to_int4pack_cpu
+
+- func: _weight_int4pack_mm_for_cpu(Tensor self, Tensor mat2, int qGroupSize, Tensor qScaleAndZeros) -> Tensor
+  dispatch:
+    CPU: _weight_int4pack_mm_cpu
+
 - func: _weight_int8pack_mm(Tensor self, Tensor mat2, Tensor scales) -> Tensor
   dispatch:
     CPU: _weight_int8pack_mm_cpu
@@ -4585,6 +4611,7 @@
     CompositeExplicitAutograd: rad2deg
     SparseCPU, SparseCUDA: rad2deg_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: rad2deg_sparse_csr
+  tags: pointwise
 
 - func: rad2deg_(Tensor(a!) self) -> Tensor(a!)
   variants: function, method
@@ -4592,12 +4619,14 @@
     CompositeExplicitAutograd: rad2deg_
     SparseCPU, SparseCUDA: rad2deg_sparse_
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: rad2deg_sparse_csr_
+  tags: pointwise
 
 - func: rad2deg.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CompositeExplicitAutograd: rad2deg_out
     SparseCPU, SparseCUDA: rad2deg_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: rad2deg_sparse_csr_out
+  tags: pointwise
 
 - func: deg2rad(Tensor self) -> Tensor
   variants: function, method
@@ -4990,7 +5019,7 @@
 
 - func: rrelu(Tensor self, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
   device_check: NoCheck   # TensorIterator
-  tags: nondeterministic_seeded
+  tags: [pointwise, nondeterministic_seeded]
 
 - func: rrelu_(Tensor(a!) self, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
   tags: nondeterministic_seeded
@@ -5027,6 +5056,7 @@
 
 - func: relu6(Tensor self) -> Tensor
   python_module: nn
+  tags: pointwise
 
 - func: relu6_(Tensor(a!) self) -> Tensor(a!)
   python_module: nn
@@ -5111,6 +5141,7 @@
   structured_delegate: hardshrink.out
   device_check: NoCheck   # TensorIterator
   variants: function, method
+  tags: pointwise
 
 - func: hardshrink_backward.grad_input(Tensor grad_out, Tensor self, Scalar lambd, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
@@ -5175,6 +5206,7 @@
 
 - func: selu(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
+  tags: pointwise
 
 - func: selu_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5183,6 +5215,7 @@
   device_check: NoCheck   # TensorIterator
   dispatch:
     CompositeExplicitAutograd: celu
+  tags: pointwise
 
 - func: celu_(Tensor(a!) self, Scalar alpha=1.0) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -5233,6 +5266,7 @@
 - func: mish(Tensor self) -> Tensor
   structured_delegate: mish.out
   python_module: nn
+  tags: pointwise
 
 - func: mish_(Tensor(a!) self) -> Tensor(a!)
   structured_delegate: mish.out
@@ -5305,7 +5339,7 @@
   dispatch:
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sin_sparse_csr
     SparseCPU, SparseCUDA: sin_sparse
-    NestedTensorCPU, NestedTensorCUDA: sin_nested
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_sin
   tags: [core, pointwise]
 
 - func: sin_(Tensor(a!) self) -> Tensor(a!)
@@ -5803,6 +5837,7 @@
   structured_delegate: sqrt.out
   variants: function, method
   dispatch:
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_sqrt
     SparseCPU, SparseCUDA: sqrt_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sqrt_sparse_csr
   tags: [core, pointwise]
@@ -6032,6 +6067,7 @@
   structured_delegate: threshold.out
   dispatch:
     QuantizedCPU: threshold_quantized_cpu
+  tags: pointwise
 
 - func: threshold_(Tensor(a!) self, Scalar threshold, Scalar value) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -6486,6 +6522,7 @@
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA, MPS: where_self_out
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_where_out
 
 - func: where.ScalarSelf(Tensor condition, Scalar self, Tensor other) -> Tensor
   variants: function
@@ -6988,6 +7025,7 @@
     CPU: addmm_out_cpu
     CUDA: addmm_out_cuda
     MPS: addmm_out_mps
+    XPU: addmm_out_xpu
     SparseCPU: addmm_out_sparse_dense_cpu
     SparseCUDA: addmm_out_sparse_dense_cuda
     SparseCsrCPU: addmm_out_sparse_compressed_cpu
@@ -7016,6 +7054,7 @@
   dispatch:
     CPU: addmm_activation_out_cpu
     CUDA: addmm_activation_out_cuda
+    XPU: addmm_activation_out_xpu
 
 - func: _addmm_activation(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, bool use_gelu=False) -> Tensor
   structured_delegate: _addmm_activation.out
@@ -7732,6 +7771,7 @@
 
 - func: cartesian_prod(Tensor[] tensors) -> Tensor
   variants: function
+  tags: maybe_aliasing_or_mutating
 
 - func: combinations(Tensor self, int r=2, bool with_replacement=False) -> Tensor
   variants: function
@@ -8013,6 +8053,7 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: masked_scatter
+  tags: core
 
 - func: masked_scatter_backward(Tensor grad_output, Tensor mask, SymInt[] sizes) -> Tensor
   dispatch:
@@ -8247,7 +8288,7 @@
   structured: True
   variants: function
   dispatch:
-    CPU, CUDA: scatter_reduce_two
+    CPU, CUDA, MPS: scatter_reduce_two
 
 - func: eq_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
   structured_delegate: eq.Scalar_out
@@ -8649,18 +8690,18 @@
 - func: addbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)
   variants: method
   dispatch:
-    CPU, CUDA: addbmm_
+    CPU, CUDA, XPU: addbmm_
     MPS: addbmm_mps_
 
 - func: addbmm.out(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU, CUDA: addbmm_out
+    CPU, CUDA, XPU: addbmm_out
     MPS: addbmm_out_mps
 
 - func: addbmm(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
   variants: method, function
   dispatch:
-    CPU, CUDA: addbmm
+    CPU, CUDA, XPU: addbmm
     MPS: addbmm_mps
 
 - func: random_.from(Tensor(a!) self, int from, int? to, *, Generator? generator=None) -> Tensor(a!)
@@ -8774,12 +8815,14 @@
   dispatch:
     CPU: tril_indices_cpu
     CUDA: tril_indices_cuda
+    MPS: tril_indices_mps
   autogen: tril_indices.out
 
 - func: triu_indices(int row, int col, int offset=0, *, ScalarType? dtype=long, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
     CPU: triu_indices_cpu
     CUDA: triu_indices_cuda
+    MPS: triu_indices_mps
   autogen: triu_indices.out
 
 - func: trace(Tensor self) -> Tensor
@@ -9234,11 +9277,13 @@
 - func: nonzero_static.out(Tensor self, *, int size, int fill_value=-1, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU: nonzero_static_out_cpu
+    CUDA: nonzero_static_out_cuda
 
 - func: nonzero_static(Tensor self, *, int size, int fill_value=-1) -> Tensor
   variants: method, function
   dispatch:
     CPU: nonzero_static_cpu
+    CUDA: nonzero_static_cuda
 
 - func: nonzero_numpy(Tensor self) -> Tensor[]
   variants: method, function
@@ -9577,7 +9622,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: i0_out
+    CPU, CUDA, MPS: i0_out
   tags: pointwise
 
 - func: sign(Tensor self) -> Tensor
@@ -10153,7 +10198,7 @@
 - func: unfold_backward(Tensor grad_in, SymInt[] input_sizes, int dim, int size, int step) -> Tensor
   variants: function
   dispatch:
-    CPU, CUDA: unfold_backward
+    CPU, CUDA, MPS: unfold_backward
   autogen: unfold_backward.out
 
 - func: equal(Tensor self, Tensor other) -> bool
@@ -11083,6 +11128,22 @@
     CUDA: foreach_tensor_lerp_list_cuda_
   autogen: _foreach_lerp.Scalar_out
 
+- func: _foreach_lerp.ScalarList(Tensor[] self, Tensor[] tensors1, Scalar[] weight) -> Tensor[]
+  device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: foreach_tensor_lerp_scalarlist_kernel_slow
+    CUDA: foreach_tensor_lerp_scalarlist_cuda
+  autogen: _foreach_lerp.ScalarList_out
+
+- func: _foreach_lerp_.ScalarList(Tensor(a!)[] self, Tensor[] tensors1, Scalar[] weight) -> ()
+  device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: foreach_tensor_lerp_scalarlist_kernel_slow_
+    CUDA: foreach_tensor_lerp_scalarlist_cuda_
+  autogen: _foreach_lerp.ScalarList_out
+
 - func: _foreach_lgamma(Tensor[] self) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
@@ -11271,6 +11332,21 @@
     CUDA: foreach_tensor_round_cuda_
   autogen: _foreach_round.out
 
+- func: _foreach_rsqrt(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_rsqrt_slow
+    CUDA: foreach_tensor_rsqrt_cuda
+
+- func: _foreach_rsqrt_(Tensor(a!)[] self) -> ()
+  device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: foreach_tensor_rsqrt_slow_
+    CUDA: foreach_tensor_rsqrt_cuda_
+  autogen: _foreach_rsqrt.out
+
 - func: _foreach_sigmoid(Tensor[] self) -> Tensor[]
   device_check: NoCheck   # foreach kernels fall back to slow path when tensor are on different devices
   variants: function
@@ -11714,6 +11790,7 @@
   structured_delegate: elu.out
   device_check: NoCheck   # TensorIterator
   python_module: nn
+  tags: pointwise
 
 - func: elu_backward.grad_input(Tensor grad_output, Scalar alpha, Scalar scale, Scalar input_scale, bool is_result, Tensor self_or_result, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
@@ -11787,6 +11864,7 @@
   python_module: nn
   dispatch:
     QuantizedCPU: hardsigmoid_quantized_cpu
+  tags: pointwise
 
 - func: hardsigmoid_(Tensor(a!) self) -> Tensor(a!)
   structured_delegate: hardsigmoid.out
@@ -11818,7 +11896,7 @@
   dispatch:
     CPU, CUDA, MPS: hardtanh
     QuantizedCPU: hardtanh_quantized_cpu
-  tags: core
+  tags: [pointwise, core]
 
 - func: hardtanh_backward.grad_input(Tensor grad_output, Tensor self, Scalar min_val, Scalar max_val, *, Tensor(a!) grad_input) -> Tensor(a!)
   python_module: nn
@@ -11942,19 +12020,20 @@
     CUDA: log_sigmoid_backward_cuda
     MPS: log_sigmoid_backward_mps
 
-- func: rrelu_with_noise.out(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None, *, Tensor(a!) out) -> Tensor(a!)
+- func: rrelu_with_noise.out(Tensor self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   tags: nondeterministic_seeded
   dispatch:
     CPU: rrelu_with_noise_out_cpu
     CUDA: rrelu_with_noise_out_cuda
 
-- func: rrelu_with_noise(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
+- func: rrelu_with_noise(Tensor self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
   python_module: nn
   dispatch:
     CPU: rrelu_with_noise_cpu
     CUDA: rrelu_with_noise_cuda
   tags: nondeterministic_seeded
+  autogen: rrelu_with_noise_functional
 
 - func: rrelu_with_noise_backward(Tensor grad_output, Tensor self, Tensor noise, Scalar lower, Scalar upper, bool training, bool self_is_result) -> Tensor
   python_module: nn
@@ -11962,7 +12041,7 @@
     CompositeExplicitAutograd: rrelu_with_noise_backward
   autogen: rrelu_with_noise_backward.out
 
-- func: rrelu_with_noise_(Tensor(a!) self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
+- func: rrelu_with_noise_(Tensor(a!) self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
   python_module: nn
   tags: nondeterministic_seeded
   dispatch:
@@ -11982,6 +12061,7 @@
   structured_delegate: softplus.out
   device_check: NoCheck   # TensorIterator
   python_module: nn
+  tags: pointwise
 
 - func: softplus_backward.grad_input(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
@@ -12008,6 +12088,7 @@
   structured_delegate: softshrink.out
   device_check: NoCheck   # TensorIterator
   python_module: nn
+  tags: pointwise
 
 - func: softshrink_backward.grad_input(Tensor grad_output, Tensor self, Scalar lambd, *, Tensor(a!) grad_input) -> Tensor(a!)
   structured: True
@@ -12652,6 +12733,7 @@
   dispatch:
     CPU: upsample_bicubic2d_out_cpu
     CUDA: upsample_bicubic2d_out_cuda
+    MPS: upsample_bicubic2d_out_mps
 
 - func: upsample_bicubic2d(Tensor self, SymInt[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
   python_module: nn
@@ -12663,6 +12745,7 @@
   dispatch:
     CPU: upsample_bicubic2d_backward_out_cpu
     CUDA: upsample_bicubic2d_backward_out_cuda
+    MPS: upsample_bicubic2d_backward_out_mps
 
 - func: upsample_bicubic2d_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
@@ -13043,17 +13126,20 @@
   dispatch:
     CPU: im2col_out_cpu
     CUDA: im2col_out_cuda
+    MPS: im2col_out_mps
 
 - func: im2col(Tensor self, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride) -> Tensor
   python_module: nn
   dispatch:
     CPU: im2col_cpu
     CUDA: im2col_cuda
+    MPS: im2col_mps
 
 - func: isfinite(Tensor self) -> Tensor
   variants: function, method
   device_check: NoCheck
   device_guard: False
+  tags: pointwise
 
 - func: isinf(Tensor self) -> Tensor
   variants: function, method
@@ -13061,6 +13147,7 @@
   device_guard: False
   dispatch:
     CompositeExplicitAutograd: isinf
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_isinf
     SparseCPU, SparseCUDA: isinf_sparse
     SparseMeta: isinf_sparse_meta
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: isinf_sparse_csr
@@ -13076,6 +13163,7 @@
   variants: function, method
   structured_delegate: isposinf.out
   dispatch:
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_isposinf
     SparseCPU, SparseCUDA: isposinf_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: isposinf_sparse_csr
   tags: pointwise
@@ -13084,7 +13172,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: isposinf_out
+    CPU, CUDA, MPS: isposinf_out
     SparseCPU, SparseCUDA: isposinf_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: isposinf_sparse_csr_out
   tags: pointwise
@@ -13093,6 +13181,7 @@
   variants: function, method
   structured_delegate: isneginf.out
   dispatch:
+    NestedTensorCPU, NestedTensorCUDA: NestedTensor_isneginf
     SparseCPU, SparseCUDA: isneginf_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: isneginf_sparse_csr
   tags: pointwise
@@ -13101,7 +13190,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: isneginf_out
+    CPU, CUDA, MPS: isneginf_out
     SparseCPU, SparseCUDA: isneginf_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: isneginf_sparse_csr_out
   tags: pointwise
@@ -13114,7 +13203,7 @@
   variants: function
 
 # See NOTE [_add_batch_dim and _remove_batch_dim]
-- func: _remove_batch_dim(Tensor self, int level, int batch_size, int out_dim) -> Tensor
+- func: _remove_batch_dim(Tensor self, int level, SymInt batch_size, int out_dim) -> Tensor
   variants: function
 
 ## Functions related to the `torch.special` namespace
@@ -13414,7 +13503,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: special_i1_out
+    CPU, CUDA, MPS: special_i1_out
   tags: pointwise
 
 - func: special_i1e(Tensor self) -> Tensor
@@ -14706,6 +14795,11 @@
     CUDA: _fbgemm_dense_to_jagged_forward_symint
     CPU: _padded_dense_to_jagged_forward_cpu
 
+- func: _nested_from_padded_tensor(Tensor padded, Tensor offsets, Tensor dummy, int ragged_idx=1, Tensor? min_seqlen=None, Tensor? max_seqlen=None, SymInt? sum_S=None) -> Tensor
+  variants: function
+  device_check: NoCheck
+  dispatch: {}
+
 - func: _nested_tensor_softmax_with_shape(Tensor self, Tensor query) -> Tensor
   dispatch:
     NestedTensorCPU: NestedTensor_softmax_dropout

data/ext/torch/device.cpp

@@ -8,7 +8,7 @@ void init_device(Rice::Module& m) {
   Rice::define_class_under<torch::Device>(m, "Device")
     .define_constructor(Rice::Constructor<torch::Device, const std::string&>())
     .define_method(
-      "index",
+      "_index",
       [](torch::Device& self) {
         return self.index();
       })
@@ -23,5 +23,10 @@ void init_device(Rice::Module& m) {
         std::stringstream s;
         s << self.type();
         return s.str();
+      })
+    .define_method(
+      "_str",
+      [](torch::Device& self) {
+        return self.str();
       });
 }

data/ext/torch/ext.cpp

@@ -31,6 +31,7 @@ void Init_ext()
 
   // keep this order
   init_torch(m);
+  init_device(m);
   init_tensor(m, rb_cTensor, rb_cTensorOptions);
   init_nn(m);
   init_fft(m);
@@ -39,7 +40,6 @@ void Init_ext()
 
   init_backends(m);
   init_cuda(m);
-  init_device(m);
   init_generator(m, rb_cGenerator);
   init_ivalue(m, rb_cIValue);
   init_random(m);

data/ext/torch/tensor.cpp

@@ -212,11 +212,9 @@ void init_tensor(Rice::Module& m, Rice::Class& c, Rice::Class& rb_cTensorOptions
         return s.str();
       })
     .define_method(
-      "device",
+      "_device",
       [](Tensor& self) {
-        std::stringstream s;
-        s << self.device();
-        return s.str();
+        return self.device();
       })
     .define_method(
       "_data_str",

data/ext/torch/torch.cpp

@@ -9,19 +9,14 @@
 #include "utils.h"
 
 template<typename T>
-torch::Tensor make_tensor(Rice::Array a, std::vector<int64_t> size, const torch::TensorOptions &options) {
+torch::Tensor make_tensor(Rice::Array a, const std::vector<int64_t> &size, const torch::TensorOptions &options) {
   std::vector<T> vec;
+  vec.reserve(a.size());
   for (long i = 0; i < a.size(); i++) {
     vec.push_back(Rice::detail::From_Ruby<T>().convert(a[i].value()));
   }
 
-  // hack for requires_grad error
-  auto requires_grad = options.requires_grad();
-  torch::Tensor t = torch::tensor(vec, options.requires_grad(c10::nullopt));
-  if (requires_grad) {
-    t.set_requires_grad(true);
-  }
-
+  torch::Tensor t = torch::tensor(vec, options);
   return t.reshape(size);
 }
 
@@ -46,12 +41,12 @@ void init_torch(Rice::Module& m) {
     // config
     .define_singleton_function(
       "show_config",
-      [] {
+      []() {
         return torch::show_config();
       })
     .define_singleton_function(
       "parallel_info",
-      [] {
+      []() {
         return torch::get_parallel_info();
       })
     // begin operations
@@ -74,13 +69,13 @@ void init_torch(Rice::Module& m) {
       })
     .define_singleton_function(
       "_from_blob",
-      [](Rice::String s, std::vector<int64_t> size, const torch::TensorOptions &options) {
+      [](Rice::String s, const std::vector<int64_t> &size, const torch::TensorOptions &options) {
         void *data = const_cast<char *>(s.c_str());
         return torch::from_blob(data, size, options);
       })
     .define_singleton_function(
       "_tensor",
-      [](Rice::Array a, std::vector<int64_t> size, const torch::TensorOptions &options) {
+      [](Rice::Array a, const std::vector<int64_t> &size, const torch::TensorOptions &options) {
         auto dtype = options.dtype();
         if (dtype == torch::kByte) {
           return make_tensor<uint8_t>(a, size, options);

data/ext/torch/utils.h

@@ -6,7 +6,7 @@
 #include <rice/stl.hpp>
 
 static_assert(
-  TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 5,
+  TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 6,
   "Incompatible LibTorch version"
 );
 

data/lib/torch/device.rb

@@ -0,0 +1,25 @@
+module Torch
+  class Device
+    def index
+      index? ? _index : nil
+    end
+
+    def inspect
+      extra = ", index: #{index.inspect}" if index?
+      "device(type: #{type.inspect}#{extra})"
+    end
+    alias_method :to_s, :inspect
+
+    def ==(other)
+      eql?(other)
+    end
+
+    def eql?(other)
+      other.is_a?(Device) && other.type == type && other.index == index
+    end
+
+    def hash
+      [type, index].hash
+    end
+  end
+end

data/lib/torch/tensor.rb

@@ -209,5 +209,10 @@ module Torch
         raise TypeError, "#{self.class} can't be coerced into #{other.class}"
       end
     end
+
+    # TODO return Device instead of String in 0.19.0
+    def device
+      _device._str
+    end
   end
 end

data/lib/torch/version.rb

@@ -1,3 +1,3 @@
 module Torch
-  VERSION = "0.18.0"
+  VERSION = "0.19.0"
 end

data/lib/torch.rb

@@ -8,6 +8,7 @@ require "set"
 require "tmpdir"
 
 # modules
+require_relative "torch/device"
 require_relative "torch/inspector"
 require_relative "torch/tensor"
 require_relative "torch/version"
@@ -382,7 +383,11 @@ module Torch
     alias_method :set_grad_enabled, :grad_enabled
 
     def device(str)
-      Device.new(str)
+      if str.is_a?(Device)
+        str
+      else
+        Device.new(str)
+      end
     end
 
     def save(obj, f)

metadata

@@ -1,14 +1,13 @@
 --- !ruby/object:Gem::Specification
 name: torch-rb
 version: !ruby/object:Gem::Version
-  version: 0.18.0
+  version: 0.19.0
 platform: ruby
 authors:
 - Andrew Kane
-autorequire:
 bindir: bin
 cert_chain: []
-date: 2024-10-22 00:00:00.000000000 Z
+date: 2025-01-30 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: rice
@@ -24,7 +23,6 @@ dependencies:
     - - ">="
       - !ruby/object:Gem::Version
         version: 4.3.3
-description:
 email: andrew@ankane.org
 executables: []
 extensions:
@@ -65,6 +63,7 @@ files:
 - ext/torch/wrap_outputs.h
 - lib/torch-rb.rb
 - lib/torch.rb
+- lib/torch/device.rb
 - lib/torch/hub.rb
 - lib/torch/inspector.rb
 - lib/torch/nn/adaptive_avg_pool1d.rb
@@ -224,7 +223,6 @@ homepage: https://github.com/ankane/torch.rb
 licenses:
 - BSD-3-Clause
 metadata: {}
-post_install_message:
 rdoc_options: []
 require_paths:
 - lib
@@ -239,8 +237,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.5.16
-signing_key:
+rubygems_version: 3.6.2
 specification_version: 4
 summary: Deep learning for Ruby, powered by LibTorch
 test_files: []