torch-rb 0.26.0 → 0.27.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: '048eb5719197dec6d4a69bdd250fd5489c57c3798190b0ac86fe8a0851ce35fb'
-  data.tar.gz: b6e8bbd5a4b663f48ab4f2a1a883813beb2bfccc73759faa4611e404e54b7656
+  metadata.gz: 640e66dad79610e8cbe3559f46edbb6673e60376dd12c52408e2c966bd8418c7
+  data.tar.gz: 7b4753cbd521f6724feeac18a2d908dc76394b890918d86e62ce92ef605155c5
 SHA512:
-  metadata.gz: 637f2fc8203a4fa16247ba2241bdec891d47aebbcb6258f6b95cb2395caf25f1b1d4712315a609a766500813805a7434cf0910e10314f8360fe6eede74a1ad4e
-  data.tar.gz: 7db34407bbc4ac7b80a7cd7a60ed9500974d06e76f89a58c9aa67d94f5f4d0883c1d0d24544db88ea3d5426430cd33c75f798e3566e0c89b743bf3979d78ab0d
+  metadata.gz: 9bf90f990d386e240af32fb177571f4cd60a523ecf5e31b4c50d0a0634d812443309af7c3f2b5f17ce0e9e5b17724c4cb4e967b2a7e3a0a003ee0d2fde692218
+  data.tar.gz: 5962a968d8815a76999eee43a9ef6f8858e7cacb870484b5f762715c0fef9aa2dc06c5361b251559684e672f40ef8668639d37b8fe32eaa5194a93bf92d3bb83

data/CHANGELOG.md

@@ -1,3 +1,7 @@
+## 0.27.0 (2026-10-28)
+
+- Updated LibTorch to 2.15.0
+
 ## 0.26.0 (2026-09-05)
 
 - Updated LibTorch to 2.14.0

data/README.md

@@ -23,7 +23,7 @@ As well as:
 First, [download LibTorch](https://pytorch.org/get-started/locally/). For Mac arm64, use:
 
 ```sh
-curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.14.0.zip > libtorch.zip
+curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.15.0.zip > libtorch.zip
 unzip -q libtorch.zip
 ```
 
@@ -419,12 +419,12 @@ Here’s the list of compatible versions.
 
 Torch.rb | LibTorch
 --- | ---
+0.27.x | 2.15.x
 0.26.x | 2.14.x
 0.25.x | 2.13.x
 0.24.x | 2.12.x
 0.23.x | 2.10.x
 0.22.x | 2.9.x
-0.21.x | 2.8.x
 
 ## Performance
 

data/codegen/native_functions.yaml

@@ -475,8 +475,7 @@
 
 - func: conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU, CUDA, XPU: conj_physical_out
-    MPS: conj_physical_out_mps
+    CPU, CUDA, XPU, MPS: conj_physical_out
     SparseCPU, SparseCUDA, SparseMPS, SparseXPU: conj_physical_out_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMPS, SparseCsrMeta, SparseCsrXPU: conj_physical_sparse_csr_out
   tags: pointwise
@@ -766,10 +765,12 @@
 - func: arange(Scalar end, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
     CompositeExplicitAutograd: arange
+    Meta: arange_meta
 
 - func: arange.start(Scalar start, Scalar end, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
     CompositeExplicitAutograd: arange
+    Meta: arange_meta
 
 # This operator should be named `arange.start_out` if following the naming convention. However that
 # name is already taken. Disabled because of CI job failures.
@@ -781,6 +782,7 @@
 - func: arange.start_step(Scalar start, Scalar end, Scalar step=1, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:
     CompositeExplicitAutograd: arange
+    Meta: arange_meta
   cpp_no_default_args: ['step']
   tags: core
 
@@ -1293,8 +1295,7 @@
 - func: logical_xor.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA, XPU: logical_xor_out
-    MPS: logical_xor_out_mps
+    CPU, CUDA, MPS, MTIA, XPU: logical_xor_out
   tags: pointwise
 
 - func: logical_and(Tensor self, Tensor other) -> Tensor
@@ -1314,8 +1315,7 @@
 - func: logical_and.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA, XPU: logical_and_out
-    MPS: logical_and_out_mps
+    CPU, CUDA, MPS, MTIA, XPU: logical_and_out
   tags: pointwise
 
 - func: logical_or(Tensor self, Tensor other) -> Tensor
@@ -1335,8 +1335,7 @@
 - func: logical_or.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA, XPU: logical_or_out
-    MPS: logical_or_out_mps
+    CPU, CUDA, MPS, MTIA, XPU: logical_or_out
   tags: pointwise
 
 - func: blackman_window(int window_length, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
@@ -3689,6 +3688,7 @@
   dispatch:
     CPU, Meta: logspace_out
     CUDA: logspace_cuda_out
+    MPS: logspace_out_mps
     XPU: logspace_xpu_out
 
 - func: logspace.Tensor_Tensor_out(Tensor start, Tensor end, int steps, float base=10.0, *, Tensor(a!) out) -> Tensor(a!)
@@ -4169,12 +4169,14 @@
   dispatch:
     CPU: _int_mm_cpu
     CUDA: _int_mm_cuda
+    MPS: _int_mm_mps
     XPU: _int_mm_xpu
 
 - func: _int_mm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU: _int_mm_out_cpu
     CUDA: _int_mm_out_cuda
+    MPS: _int_mm_out_mps
     XPU: _int_mm_out_xpu
 
 - func: _convert_weight_to_int4pack(Tensor self, int innerKTiles) -> Tensor
@@ -5085,8 +5087,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, XPU: relu
-    MPS: relu_mps
+    CPU, CUDA, MPS, XPU: relu
     MTIA: relu_mtia
     MkldnnCPU: mkldnn_relu
     QuantizedCPU: relu_quantized_cpu
@@ -5100,8 +5101,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, XPU: relu_
-    MPS: relu_mps_
+    CPU, CUDA, MPS, XPU: relu_
     MTIA: relu_mtia_
     MkldnnCPU: mkldnn_relu_
     QuantizedCPU: relu_quantized_cpu_
@@ -5970,8 +5970,7 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA, XPU: prod
-    MPS: prod_mps
+    CPU, CUDA, MPS, XPU: prod
   autogen: prod.out
   tags: [core, reduction]
 
@@ -5985,8 +5984,7 @@
   structured: True
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, XPU: prod_out
-    MPS: prod_out_mps
+    CPU, CUDA, MPS, XPU: prod_out
   tags: reduction
 
 - func: t(Tensor(a) self) -> Tensor(a)
@@ -6154,8 +6152,7 @@
 - func: flip(Tensor self, int[] dims) -> Tensor
   variants: function, method
   dispatch:
-    CPU, QuantizedCPU, CUDA, QuantizedCUDA, XPU: flip
-    MPS: flip_mps
+    CPU, QuantizedCPU, CUDA, QuantizedCUDA, XPU, MPS: flip
   autogen: flip.out
   tags: core
 
@@ -6638,6 +6635,13 @@
     CUDA: _philox_uniform_cuda_
   autogen: _philox_uniform, _philox_uniform.out
 
+- func: _philox_randint_(Tensor(a!) self, Tensor key, int? low=None, int? high=None) -> Tensor(a!)
+  variants: function, method
+  dispatch:
+    CPU: _philox_randint_cpu_
+    CUDA: _philox_randint_cuda_
+  autogen: _philox_randint, _philox_randint.out
+
 - func: _dirichlet_grad(Tensor x, Tensor alpha, Tensor total) -> Tensor
   dispatch:
     CPU: _dirichlet_grad_cpu
@@ -6855,7 +6859,7 @@
 
 - func: frexp.Tensor_out(Tensor self, *, Tensor(a!) mantissa, Tensor(b!) exponent) -> (Tensor(a!) mantissa, Tensor(b!) exponent)
   dispatch:
-    CPU, CUDA, XPU: frexp_out
+    CPU, CUDA, MPS, XPU: frexp_out
   tags: pointwise
 
 # Deprecated (v.1.12)
@@ -6918,7 +6922,8 @@
   device_check: NoCheck   # TensorIterator
   variants: method, function
   dispatch:
-    CPU, CUDA, MPS, XPU: zero_
+    CPU, MPS, XPU: zero_
+    CUDA: zero_cuda_
     Meta: zero_meta_
     SparseCPU, SparseCUDA, SparseMPS, SparseMeta, SparseXPU: zero_sparse_
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta, SparseCsrXPU: zero_sparse_csr_
@@ -6931,8 +6936,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA, XPU: sub_out
-    MPS: sub_out_mps
+    CPU, CUDA, MPS, XPU: sub_out
     MTIA: sub_out_mtia
     SparseCPU, SparseCUDA, SparseMPS, SparseXPU: sub_out_sparse
   tags: pointwise
@@ -6999,7 +7003,7 @@
   structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, XPU: heaviside_out
+    CPU, CUDA, MPS, XPU: heaviside_out
   tags: pointwise
 
 - func: heaviside(Tensor self, Tensor values) -> Tensor
@@ -7131,7 +7135,6 @@
 - 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
   structured_delegate: _scaled_mm_v2.out
   variants: function
-  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!)
   structured: True
@@ -7139,7 +7142,24 @@
   dispatch:
     CPU: _scaled_mm_cpu_v2_out
     CUDA: _scaled_mm_cuda_v2_out
+    MPS: _scaled_mm_mps_v2_out
     XPU: _scaled_mm_xpu_v2_out
+
+- func: _scaled_addmm(Tensor self, Tensor mat1, Tensor mat2, Tensor[] scale_a, int[] recipe_a, int[] swizzle_a, Tensor[] scale_b, int[] recipe_b, int[] swizzle_b, int[] contraction_dim=[], *, Scalar beta=1, Scalar alpha=1, bool use_fast_accum=False) -> Tensor
+  structured_delegate: _scaled_addmm.out
+  variants: function
+  tags: needs_exact_strides
+
+- func: _scaled_addmm_(Tensor(a!) self, Tensor mat1, Tensor mat2, Tensor[] scale_a, int[] recipe_a, int[] swizzle_a, Tensor[] scale_b, int[] recipe_b, int[] swizzle_b, int[] contraction_dim=[], *, Scalar beta=1, Scalar alpha=1, bool use_fast_accum=False) -> Tensor(a!)
+  structured_delegate: _scaled_addmm.out
+  variants: function
+  tags: needs_exact_strides
+
+- func: _scaled_addmm.out(Tensor self, Tensor mat1, Tensor mat2, Tensor[] scale_a, int[] recipe_a, int[] swizzle_a, Tensor[] scale_b, int[] recipe_b, int[] swizzle_b, int[] contraction_dim=[], *, Scalar beta=1, Scalar alpha=1, bool use_fast_accum=False, Tensor(a!) out) -> Tensor(a!)
+  structured: True
+  variants: function
+  dispatch:
+    CUDA: _scaled_addmm_cuda_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
@@ -7165,6 +7185,7 @@
   dispatch:
     CompositeExplicitAutograd: _grouped_mm
     CUDA: _grouped_mm_cuda
+    MPS: _grouped_mm_mps
 
 # NOTE [ Sparse: autograd and API ]
 #
@@ -8207,7 +8228,7 @@
 - func: put_(Tensor(a!) self, Tensor index, Tensor source, bool accumulate=False) -> Tensor(a!)
   variants: method
   dispatch:
-    CPU, CUDA, XPU: put_
+    CPU, CUDA, MPS, XPU: put_
   autogen: put.out
 
 - func: put(Tensor self, Tensor index, Tensor source, bool accumulate=False) -> Tensor
@@ -9259,12 +9280,12 @@
 
 - func: take.out(Tensor self, Tensor index, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU, CUDA, XPU: take_out
+    CPU, CUDA, MPS, XPU: take_out
 
 - func: take(Tensor self, Tensor index) -> Tensor
   variants: method, function
   dispatch:
-    CPU, CUDA, XPU: take
+    CPU, CUDA, MPS, XPU: take
 
 - func: take_along_dim.out(Tensor self, Tensor indices, int? dim=None, *, Tensor(a!) out) -> Tensor(a!)
 
@@ -9383,8 +9404,7 @@
   structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA, XPU: addcmul_out
-    MPS: addcmul_out_mps
+    CPU, CUDA, MPS, MTIA, XPU: addcmul_out
   tags: pointwise
 
 - func: addcmul(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor
@@ -9404,8 +9424,7 @@
   structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MTIA, XPU: addcdiv_out
-    MPS: addcdiv_out_mps
+    CPU, CUDA, MPS, MTIA, XPU: addcdiv_out
   tags: pointwise
 
 - func: addcdiv(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor
@@ -9517,12 +9536,12 @@
 
 - func: geqrf.a(Tensor self, *, Tensor(a!) a, Tensor(b!) tau) -> (Tensor(a!) a, Tensor(b!) tau)
   dispatch:
-    CPU, CUDA, MPS: geqrf_out
+    CPU, CUDA, MPS, XPU: geqrf_out
 
 - func: geqrf(Tensor self) -> (Tensor a, Tensor tau)
   variants: method, function
   dispatch:
-    CPU, CUDA, MPS: geqrf
+    CPU, CUDA, MPS, XPU: geqrf
 
 # orgqr, alias for linalg_householder_product
 - func: orgqr(Tensor self, Tensor input2) -> Tensor
@@ -10251,8 +10270,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA, XPU: pow_Tensor_Tensor_out
-    MPS: pow_tensor_tensor_out_mps
+    CPU, CUDA, MPS, XPU: pow_Tensor_Tensor_out
   tags: pointwise
 
 - func: pow.Tensor_Tensor(Tensor self, Tensor exponent) -> Tensor
@@ -11245,7 +11263,6 @@
     CompositeExplicitAutograd: foreach_tensor_ternary_lerp_slow
     CUDA: foreach_tensor_lerp_ternary_cuda
     XPU: foreach_tensor_lerp_ternary_xpu
-  autogen: _foreach_lerp.List_out
 
 - func: _foreach_lerp_.List(Tensor(a!)[] self, Tensor[] tensors1, Tensor[] weights) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
@@ -11263,7 +11280,6 @@
     CompositeExplicitAutograd: foreach_tensor_lerp_list_kernel_slow
     CUDA: foreach_tensor_lerp_list_cuda
     XPU: foreach_tensor_lerp_list_xpu
-  autogen: _foreach_lerp.Scalar_out
 
 - func: _foreach_lerp_.Scalar(Tensor(a!)[] self, Tensor[] tensors1, Scalar weight) -> ()
   device_check: NoCheck   # foreach kernels fall back to slow path when tensors are on different devices
@@ -11281,7 +11297,6 @@
     CompositeExplicitAutograd: foreach_tensor_lerp_scalarlist_kernel_slow
     CUDA: foreach_tensor_lerp_scalarlist_cuda
     XPU: foreach_tensor_lerp_scalarlist_xpu
-  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
@@ -11420,6 +11435,7 @@
   dispatch:
     CompositeExplicitAutograd: foreach_tensor_powsum_slow
     CUDA: foreach_tensor_powsum_cuda
+    XPU: foreach_tensor_powsum_xpu
   autogen: _foreach_powsum.Scalar_out
 
 - func: _foreach_pow.List(Tensor[] self, Tensor[] exponent) -> Tensor[]
@@ -12300,8 +12316,7 @@
   device_check: NoCheck   # TensorIterator
   python_module: nn
   dispatch:
-    CPU, CUDA, XPU: softplus_out
-    MPS: softplus_out_mps
+    CPU, CUDA, MPS, XPU: softplus_out
 
 - func: softplus(Tensor self, Scalar beta=1, Scalar threshold=20) -> Tensor
   structured_delegate: softplus.out
@@ -12314,8 +12329,7 @@
   structured_inherits: TensorIteratorBase
   python_module: nn
   dispatch:
-    CPU, CUDA, XPU: softplus_backward_out
-    MPS: softplus_backward_out_mps
+    CPU, CUDA, MPS, XPU: softplus_backward_out
 
 - func: softplus_backward(Tensor grad_output, Tensor self, Scalar beta, Scalar threshold) -> Tensor
   structured_delegate: softplus_backward.grad_input
@@ -13318,8 +13332,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA, XPU: logit_backward_out
-    MPS: logit_backward_out_mps
+    CPU, CUDA, MPS, XPU: logit_backward_out
   tags: pointwise
 
 - func: logit_backward(Tensor grad_output, Tensor self, float? eps=None) -> Tensor
@@ -13332,8 +13345,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA, MTIA, XPU: tanh_backward_out
-    MPS: tanh_backward_out_mps
+    CPU, CUDA, MPS, MTIA, XPU: tanh_backward_out
   tags: pointwise
 
 - func: tanh_backward(Tensor grad_output, Tensor output) -> Tensor
@@ -13519,6 +13531,7 @@
   variants: method
   dispatch:
     CUDA: record_stream_cuda
+    MPS: record_stream_mps
     XPU: record_stream_xpu
 
 - func: isposinf(Tensor self) -> Tensor
@@ -14485,12 +14498,12 @@
   python_module: linalg
   variants: function
   dispatch:
-    CPU, CUDA, MPS: linalg_householder_product
+    CPU, CUDA, MPS, XPU: linalg_householder_product
 
 - func: linalg_householder_product.out(Tensor input, Tensor tau, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
   dispatch:
-    CPU, CUDA, MPS: linalg_householder_product_out
+    CPU, CUDA, MPS, XPU: linalg_householder_product_out
 
 - func: linalg_inv_ex(Tensor A, *, bool check_errors=False) -> (Tensor inverse, Tensor info)
   python_module: linalg
@@ -14551,7 +14564,7 @@
   variants: function
   structured_delegate: linalg_vector_norm.out
   dispatch:
-    SparseCPU, SparseCUDA, SparseMPS: linalg_vector_norm_sparse
+    SparseCPU, SparseCUDA, SparseMPS, SparseXPU: linalg_vector_norm_sparse
   tags: reduction
 
 - func: linalg_vector_norm.out(Tensor self, Scalar ord=2, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
@@ -14569,7 +14582,7 @@
   variants: function
   dispatch:
     CompositeExplicitAutograd: linalg__powsum_slow
-    CPU, CUDA: linalg__powsum
+    CPU, CUDA, XPU: linalg__powsum
   tags: reduction
 
 - func: linalg_matrix_norm(Tensor self, Scalar ord, int[] dim=[-2,-1], bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
@@ -14689,6 +14702,7 @@
   python_module: sparse
   dispatch:
     SparseCsrCUDA: _sparse_csr_linear_solve
+    SparseCsrXPU: _sparse_csr_linear_solve_xpu
 
 - func: linalg_solve.out(Tensor A, Tensor B, *, bool left=True, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
@@ -14718,7 +14732,7 @@
   python_module: linalg
   structured: True
   dispatch:
-    CPU, CUDA, MPS: linalg_qr_out
+    CPU, CUDA, MPS, XPU: linalg_qr_out
 
 - func: linalg_polar(Tensor A) -> (Tensor U, Tensor H)
   python_module: linalg
@@ -14729,7 +14743,7 @@
   python_module: linalg
   structured: True
   dispatch:
-    CPU, CUDA, MPS: linalg_polar_out
+    CPU, CUDA, MPS, XPU: linalg_polar_out
 
 - func: linalg_matrix_power(Tensor self, int n) -> Tensor
   python_module: linalg
@@ -15352,6 +15366,12 @@
     CUDA: _cudnn_attention_forward
   tags: nondeterministic_seeded
 
+- func: _cudnn_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, bool is_causal=False, *, float? scale=None, Tensor? seqused_k=None, Tensor? block_table=None) -> Tensor logsumexp
+  variants: function
+  dispatch:
+    CUDA: _cudnn_attention_forward_no_dropout_inplace
+  tags: nondeterministic_seeded
+
 - func: _cudnn_attention_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor out, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor attn_bias, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, *, float? scale=None) -> (Tensor, Tensor, Tensor)
   dispatch:
     CUDA: _cudnn_attention_backward
@@ -15771,7 +15791,7 @@
 - func: special_laguerre_polynomial_l.out(Tensor x, Tensor n, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck
   dispatch:
-    CPU, CUDA, XPU: special_laguerre_polynomial_l_out
+    CPU, CUDA, MPS, XPU: special_laguerre_polynomial_l_out
   python_module: special
   structured_inherits: TensorIteratorBase
   structured: True

data/ext/torch/utils.h

@@ -8,7 +8,7 @@
 #include <rice/stl.hpp>
 
 static_assert(
-  TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 14,
+  TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 15,
   "Incompatible LibTorch version"
 );
 

data/lib/torch/nn/module.rb

@@ -337,7 +337,7 @@ module Torch
           if vars.any?
             vars.map(&:inspect)
           else
-            options.map { |k, v| [k, v.inspect] }.to_h
+            options.to_h { |k, v| [k, v.inspect] }
           end
         str % vars
       end
@@ -346,7 +346,7 @@ module Torch
       # remove tensors for performance
       # so we can skip call to inspect
       def dict
-        instance_variables.reject { |k| instance_variable_get(k).is_a?(Tensor) }.map { |k| [k[1..-1].to_sym, instance_variable_get(k)] }.to_h
+        instance_variables.reject { |k| instance_variable_get(k).is_a?(Tensor) }.to_h { |k| [k[1..-1].to_sym, instance_variable_get(k)] }
       end
 
       def load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)

data/lib/torch/optim/lr_scheduler/multi_step_lr.rb

@@ -3,7 +3,7 @@ module Torch
     module LRScheduler
       class MultiStepLR < LRScheduler
         def initialize(optimizer, milestones, gamma: 0.1, last_epoch: -1)
-          @milestones = milestones.map.with_index.map { |v, i| [v, i + 1] }.to_h
+          @milestones = milestones.map.with_index.to_h { |v, i| [v, i + 1] }
           @gamma = gamma
           super(optimizer, last_epoch)
         end

data/lib/torch/optim/optimizer.rb

@@ -35,13 +35,13 @@ module Torch
         raise NotImplementedYet
 
         pack_group = lambda do |group|
-          packed = group.select { |k, _| k != :params }.map { |k, v| [k.to_s, v] }.to_h
+          packed = group.select { |k, _| k != :params }.to_h { |k, v| [k.to_s, v] }
           packed["params"] = group[:params].map { |p| p.object_id }
           packed
         end
 
         param_groups = @param_groups.map { |g| pack_group.call(g) }
-        packed_state = @state.map { |k, v| [k.is_a?(Tensor) ? k.object_id : k, v] }.to_h
+        packed_state = @state.to_h { |k, v| [k.is_a?(Tensor) ? k.object_id : k, v] }
 
         {
           "state" => packed_state,

data/lib/torch/version.rb

@@ -1,3 +1,3 @@
 module Torch
-  VERSION = "0.26.0"
+  VERSION = "0.27.0"
 end

data/lib/torch.rb

@@ -272,7 +272,7 @@ module Torch
     qint32: 14,
     bfloat16: 15
   }
-  ENUM_TO_DTYPE = DTYPE_TO_ENUM.map(&:reverse).to_h
+  ENUM_TO_DTYPE = DTYPE_TO_ENUM.invert
 
   TENSOR_TYPE_CLASSES = []
 
@@ -408,7 +408,7 @@ module Torch
 
     def tensor(data, **options)
       if options[:dtype].nil? && defined?(Numo::NArray) && data.is_a?(Numo::NArray)
-        numo_to_dtype = _dtype_to_numo.map(&:reverse).to_h
+        numo_to_dtype = _dtype_to_numo.invert
         options[:dtype] = numo_to_dtype[data.class]
       end
 

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: torch-rb
 version: !ruby/object:Gem::Version
-  version: 0.26.0
+  version: 0.27.0
 platform: ruby
 authors:
 - Andrew Kane
@@ -241,7 +241,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 4.0.14
+rubygems_version: 4.0.20
 specification_version: 4
 summary: Deep learning for Ruby, powered by LibTorch
 test_files: []