torch-rb 0.19.0 → 0.20.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 03164cc479d8f8a32f0669d597e8fe5310d91955e6954cfdc0fffdc8983c5768
-  data.tar.gz: 87fc733016b6f4489b38a419a3879cacbdb1e190cfaa5c02397aceb57c012d16
+  metadata.gz: 39fe7c12b647086c46318bb77d1aac80bac1874cf3b71f1a0db4a1671322799e
+  data.tar.gz: 81a1c4d3ad4474b06dac82df2c38c498c7b4f5296e700d3bc98f596df14d10a4
 SHA512:
-  metadata.gz: 6ba0480138a10ba43dff625dc1bcf99e2287f238dc4607ea6813e82914f1e335133f55408fa59579343b161c54850316f744c594cf6687b7f2de64a0d71746d1
-  data.tar.gz: 859015641dd14bf919a7982c6673acb296f858518552b8c924fc7e59b9c1b2a9491aa598c01b019b392c8c2bba7b9f65ff0923f838e6cbde7ddaede9c4b69191
+  metadata.gz: d3677acb4621a6e47194acb81612a369716dcecb86518b1f0840e4d0e3e9fe48e8eaa3c7888cb1adc0c8e4ca78aa920a23c5c452d7a33fde92461fc6fb27b9bc
+  data.tar.gz: a02603711a10d16c6521c55c4389ade96503e60d21748d566552679acc2ab10805ada2c8083813b74ccfbba0742b68adc3cff5461b9dd000761326ee6e1ccffb

data/CHANGELOG.md

@@ -1,3 +1,13 @@
+## 0.20.0 (2025-04-26)
+
+- Updated LibTorch to 2.7.0
+- Added `Normal` distribution
+- Fixed `SystemStackError` with certain tensor comparisons
+
+## 0.19.1 (2025-02-10)
+
+- Fixed error with Rice 4.5
+
 ## 0.19.0 (2025-01-29)
 
 - Updated LibTorch to 2.6.0

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-arm64-2.6.0.zip > libtorch.zip
+curl -L https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.7.0.zip > libtorch.zip
 unzip -q libtorch.zip
 ```
 
@@ -413,6 +413,7 @@ Here’s the list of compatible versions.
 
 Torch.rb | LibTorch
 --- | ---
+0.20.x | 2.7.x
 0.19.x | 2.6.x
 0.18.x | 2.5.x
 0.17.x | 2.4.x

data/codegen/native_functions.yaml

@@ -403,6 +403,7 @@
   variants: function, method
   dispatch:
     CPU, CUDA: angle
+    MPS: angle_mps
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: angle_sparse_csr
   tags: pointwise
 
@@ -410,6 +411,7 @@
   device_check: NoCheck   # TensorIterator
   dispatch:
     CPU, CUDA: angle_out
+    MPS: angle_out_mps
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: angle_sparse_csr_out
   tags: pointwise
 
@@ -1456,8 +1458,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: ceil_out
-    MPS: ceil_out_mps
+    CPU, CUDA, MPS: ceil_out
     SparseCPU, SparseCUDA: ceil_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: ceil_sparse_csr_out
   tags: pointwise
@@ -2576,8 +2577,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: exp_out
-    MPS: exp_out_mps
+    CPU, CUDA, MPS: exp_out
   tags: pointwise
 
 - func: exp2(Tensor self) -> Tensor
@@ -2740,8 +2740,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: floor_out
-    MPS: floor_out_mps
+    CPU, CUDA, MPS: floor_out
     SparseCPU, SparseCUDA: floor_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: floor_sparse_csr_out
   tags: pointwise
@@ -3000,6 +2999,7 @@
     CPU: _fft_r2c_mkl
     CUDA: _fft_r2c_cufft
     MPS: _fft_r2c_mps
+  tags: core
 
 - func: _fft_r2c.out(Tensor self, int[] dim, int normalization, bool onesided, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
@@ -3864,6 +3864,7 @@
   device_guard: False
   dispatch:
     CompositeImplicitAutograd: value_selecting_reduction_backward_symint
+    NestedTensorCPU, NestedTensorCUDA: value_selecting_reduction_backward_nested_symint
 
 - func: amax(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor
   variants: function, method
@@ -4177,6 +4178,14 @@
   dispatch:
     CPU: _weight_int4pack_mm_cpu
 
+- func: _dyn_quant_pack_4bit_weight(Tensor weights, Tensor scales_zeros, Tensor? bias, int block_size, int in_features, int out_features) -> Tensor
+  dispatch:
+    CPU: _dyn_quant_pack_4bit_weight_cpu
+
+- func: _dyn_quant_matmul_4bit(Tensor inp, Tensor packed_weights, int block_size, int in_features, int out_features) -> Tensor
+  dispatch:
+    CPU: _dyn_quant_matmul_4bit_cpu
+
 - func: _weight_int8pack_mm(Tensor self, Tensor mat2, Tensor scales) -> Tensor
   dispatch:
     CPU: _weight_int8pack_mm_cpu
@@ -4989,9 +4998,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU: round_out
-    CUDA: round_out
-    MPS: round_out_mps
+    CPU, CUDA, MPS: round_out
     SparseCPU, SparseCUDA: round_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: round_sparse_csr_out
   tags: pointwise
@@ -5013,8 +5020,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU: round_decimals_out
-    CUDA: round_decimals_out
+    CPU, CUDA, MPS: round_decimals_out
   tags: pointwise
 
 - func: rrelu(Tensor self, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
@@ -5376,7 +5382,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: sinc_out
+    CPU, CUDA, MPS: sinc_out
   tags: pointwise
 
 - func: sinh(Tensor self) -> Tensor
@@ -5747,11 +5753,11 @@
 - func: dstack.out(Tensor[] tensors, *, Tensor(a!) out) -> Tensor(a!)
 
 # Overload without center & pad mode, needed for forward-compatibility
-- func: stft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool normalized=False, bool? onesided=None, bool? return_complex=None) -> Tensor
+- func: stft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool normalized=False, bool? onesided=None, bool? return_complex=None, bool? align_to_window=None) -> Tensor
   variants: function, method
   cpp_no_default_args: ['hop_length', 'win_length', 'window', 'normalized']
 
-- func: stft.center(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, str pad_mode="reflect", bool normalized=False, bool? onesided=None, bool? return_complex=None) -> Tensor
+- func: stft.center(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, str pad_mode="reflect", bool normalized=False, bool? onesided=None, bool? return_complex=None, bool? align_to_window=None) -> Tensor
   variants: function, method
 
 - func: istft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, bool normalized=False, bool? onesided=None, int? length=None, bool return_complex=False) -> Tensor
@@ -5856,8 +5862,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: sqrt_out
-    MPS: sqrt_out_mps
+    CPU, CUDA, MPS: sqrt_out
     SparseCPU, SparseCUDA: sqrt_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sqrt_sparse_csr_out
   tags: pointwise
@@ -6048,8 +6053,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: tanh_out
-    MPS: tanh_out_mps
+    CPU, CUDA, MPS: tanh_out
     SparseCPU, SparseCUDA: tanh_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: tanh_sparse_csr_out
   tags: pointwise
@@ -6340,8 +6344,7 @@
   structured_inherits: TensorIteratorBase
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: trunc_out
-    MPS: trunc_out_mps
+    CPU, CUDA, MPS: trunc_out
     SparseCPU, SparseCUDA: trunc_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: trunc_sparse_csr_out
   tags: pointwise
@@ -7070,6 +7073,12 @@
   dispatch:
     CUDA: _scaled_mm_out_cuda
 
+
+- 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
+  variants: function
+  dispatch:
+    CUDA: _scaled_grouped_mm_cuda
+
 # NOTE [ Sparse: autograd and API ]
 #
 #
@@ -9274,12 +9283,12 @@
     MPS: nonzero_mps
   tags: [dynamic_output_shape, core]
 
-- func: nonzero_static.out(Tensor self, *, int size, int fill_value=-1, Tensor(a!) out) -> Tensor(a!)
+- func: nonzero_static.out(Tensor self, *, SymInt 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
+- func: nonzero_static(Tensor self, *, SymInt size, int fill_value=-1) -> Tensor
   variants: method, function
   dispatch:
     CPU: nonzero_static_cpu
@@ -9428,11 +9437,13 @@
 - func: cholesky.out(Tensor self, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU, CUDA: cholesky_out
+    MPS: cholesky_mps_out
 
 - func: cholesky(Tensor self, bool upper=False) -> Tensor
   variants: method, function
   dispatch:
     CPU, CUDA: cholesky
+    MPS: cholesky_mps
 
 - func: cholesky_solve.out(Tensor self, Tensor input2, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -9506,6 +9517,7 @@
   structured: True
   dispatch:
     CPU, CUDA: lu_unpack_out
+    MPS: lu_unpack_out_mps
 
 # TODO: remove dispatch section when porting TH CUDA to ATen
 - func: multinomial.out(Tensor self, int num_samples, bool replacement=False, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)
@@ -9602,8 +9614,7 @@
   structured: True
   structured_inherits: TensorIteratorBase
   dispatch:
-    CPU, CUDA: erfinv_out
-    MPS: erfinv_out_mps
+    CPU, CUDA, MPS: erfinv_out
     SparseCPU, SparseCUDA: erfinv_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: erfinv_sparse_csr_out
   tags: pointwise
@@ -12711,6 +12722,7 @@
   dispatch:
     CPU: _upsample_bilinear2d_aa_out_cpu
     CUDA: _upsample_bilinear2d_aa_out_cuda
+    MPS: _upsample_bilinear2d_aa_out_mps
 
 - func: _upsample_bilinear2d_aa(Tensor self, SymInt[2] output_size, bool align_corners, float? scales_h=None, float? scales_w=None) -> Tensor
   python_module: nn
@@ -13058,7 +13070,6 @@
   autogen: _slow_conv2d_backward.output_mask_out
 
 - func: _conv_depthwise2d.out(Tensor self, Tensor weight, SymInt[2] kernel_size, Tensor? bias, SymInt[2] stride, SymInt[2] padding, SymInt[2] dilation, *, Tensor(a!) out) -> Tensor(a!)
-  use_const_ref_for_mutable_tensors: True
   python_module: nn
   dispatch:
     CUDA: conv_depthwise2d_cuda_out
@@ -13225,7 +13236,7 @@
   python_module: special
   variants: function
   dispatch:
-    CPU, CUDA: special_entr_out
+    CPU, CUDA, MPS: special_entr_out
   tags: pointwise
 
 - func: special_ndtri(Tensor self) -> Tensor
@@ -13372,7 +13383,7 @@
   python_module: special
   variants: function
   dispatch:
-    CPU, CUDA: special_xlog1py_out
+    CPU, CUDA, MPS: special_xlog1py_out
   tags: pointwise
 
 - func: special_xlog1py.self_scalar_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
@@ -13451,7 +13462,7 @@
   python_module: special
   variants: function
   dispatch:
-    CPU, CUDA: special_zeta_out
+    CPU, CUDA, MPS: special_zeta_out
   tags: pointwise
 
 - func: special_zeta.self_scalar_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
@@ -13744,7 +13755,6 @@
     CompositeImplicitAutograd: fft_hfft2_symint
 
 - func: fft_hfft2.out(Tensor self, SymInt[1]? s=None, int[1] dim=[-2,-1], str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
-  use_const_ref_for_mutable_tensors: True
   python_module: fft
   variants: function
   dispatch:
@@ -13758,7 +13768,6 @@
     CompositeImplicitAutograd: fft_ihfft2_symint
 
 - func: fft_ihfft2.out(Tensor self, SymInt[1]? s=None, int[1] dim=[-2,-1], str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
-  use_const_ref_for_mutable_tensors: True
   python_module: fft
   variants: function
   dispatch:
@@ -13820,7 +13829,6 @@
     CompositeImplicitAutograd: fft_hfftn_symint
 
 - func: fft_hfftn.out(Tensor self, SymInt[1]? s=None, int[1]? dim=None, str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
-  use_const_ref_for_mutable_tensors: True
   python_module: fft
   variants: function
   dispatch:
@@ -13834,7 +13842,6 @@
     CompositeImplicitAutograd: fft_ihfftn_symint
 
 - func: fft_ihfftn.out(Tensor self, SymInt[1]? s=None, int[1]? dim=None, str? norm=None, *, Tensor(a!) out) -> Tensor(a!)
-  use_const_ref_for_mutable_tensors: True
   python_module: fft
   variants: function
   dispatch:
@@ -13891,6 +13898,7 @@
   structured: True
   dispatch:
     CPU, CUDA: linalg_cholesky_ex_out
+    MPS: linalg_cholesky_ex_out_mps
 
 - func: linalg_cholesky(Tensor self, *, bool upper=False) -> Tensor
   python_module: linalg
@@ -13937,6 +13945,7 @@
   structured: True
   dispatch:
     CPU, CUDA: linalg_lu_factor_ex_out
+    MPS: linalg_lu_factor_ex_out_mps
 
 # linalg.lu
 - func: linalg_lu(Tensor A, *, bool pivot=True) -> (Tensor P, Tensor L, Tensor U)
@@ -13971,7 +13980,7 @@
 - func: _linalg_det.result(Tensor A, *, Tensor(a!) result, Tensor(b!) LU, Tensor(c!) pivots) -> (Tensor(a!) result, Tensor(b!) LU, Tensor(c!) pivots)
   structured: True
   dispatch:
-    CPU, CUDA: _linalg_det_out
+    CPU, CUDA, MPS: _linalg_det_out
 
 - func: linalg_det(Tensor A) -> Tensor
   python_module: linalg
@@ -14058,7 +14067,7 @@
 - func: _linalg_slogdet.sign(Tensor A, *, Tensor(a!) sign, Tensor(b!) logabsdet, Tensor(c!) LU, Tensor(d!) pivots) -> (Tensor(a!) sign, Tensor(b!) logabsdet, Tensor(c!) LU, Tensor(d!) pivots)
   structured: True
   dispatch:
-    CPU, CUDA: _linalg_slogdet_out
+    CPU, CUDA, MPS: _linalg_slogdet_out
 
 - func: linalg_slogdet(Tensor A) -> (Tensor sign, Tensor logabsdet)
   python_module: linalg
@@ -14300,6 +14309,7 @@
   structured: True
   dispatch:
     CPU, CUDA: _linalg_solve_ex_out
+    MPS: _linalg_solve_ex_out_mps
 
 - func: linalg_solve_ex(Tensor A, Tensor B, *, bool left=True, bool check_errors=False) -> (Tensor result, Tensor info)
   python_module: linalg
@@ -14837,6 +14847,7 @@
     Meta: _fused_sdp_choice_meta
     CPU, NestedTensorCPU: _fused_sdp_choice_cpp
     CUDA, NestedTensorCUDA: _fused_sdp_choice_cuda
+    XPU: _fused_sdp_choice_xpu
   tags: nondeterministic_seeded
 
 - func: _scaled_dot_product_attention_math(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)
@@ -14848,7 +14859,7 @@
     MPS: _scaled_dot_product_attention_math_mps
   tags: nondeterministic_seeded
 
-- func: _scaled_dot_product_flash_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)
+- func: _scaled_dot_product_flash_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 rng_state, Tensor unused, Tensor debug_attn_mask)
   dispatch:
     CUDA: _scaled_dot_product_flash_attention_cuda
     NestedTensorCUDA: _scaled_dot_product_flash_attention_nestedtensor_cuda
@@ -14862,6 +14873,7 @@
 - func: _scaled_dot_product_fused_attention_overrideable(Tensor query, Tensor key, Tensor value, Tensor? attn_bias=None, 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:
     CompositeExplicitAutograd: _scaled_dot_product_fused_attention_overrideable
+    XPU: _scaled_dot_product_fused_attention_overrideable_xpu
   tags: nondeterministic_seeded
 
 - func: _scaled_dot_product_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 grad_query, Tensor grad_key, Tensor grad_value)
@@ -14898,6 +14910,7 @@
 - func: _scaled_dot_product_cudnn_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, bool compute_log_sumexp, 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
+    NestedTensorCUDA: _scaled_dot_product_cudnn_attention_nestedtensor_cuda
   tags: nondeterministic_seeded
 
 - func: _scaled_dot_product_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)
@@ -14905,13 +14918,13 @@
     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)
+- 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)
   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, SymInt? window_size_left=None, SymInt? window_size_right=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 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
   dispatch:
@@ -14930,6 +14943,11 @@
   dispatch:
     CUDA: _efficient_attention_backward
 
+- func: _cudnn_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, bool compute_log_sumexp, 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: _cudnn_attention_forward
+  tags: nondeterministic_seeded
+
 - func: _triton_scaled_dot_attention(Tensor q, Tensor k, Tensor v, float dropout_p=0.0) -> Tensor
   variants: function
   dispatch:
@@ -15710,7 +15728,7 @@
 
 - func: special_spherical_bessel_j0.out(Tensor x, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
-    CPU, CUDA: special_spherical_bessel_j0_out
+    CPU, CUDA, MPS: special_spherical_bessel_j0_out
   python_module: special
   structured_inherits: TensorIteratorBase
   structured: True

data/ext/torch/templates.h

@@ -46,10 +46,7 @@ namespace Rice::detail
   template<typename T>
   struct Type<c10::complex<T>>
   {
-    static bool verify()
-    {
-      return true;
-    }
+    static bool verify() { return true; }
   };
 
   template<typename T>
@@ -66,6 +63,8 @@ namespace Rice::detail
   class From_Ruby<c10::complex<T>>
   {
   public:
+    Convertible is_convertible(VALUE value) { return Convertible::Cast; }
+
     c10::complex<T> convert(VALUE x)
     {
       VALUE real = rb_funcall(x, rb_intern("real"), 0);
@@ -80,16 +79,15 @@ namespace Rice::detail
   template<>
   struct Type<FanModeType>
   {
-    static bool verify()
-    {
-      return true;
-    }
+    static bool verify() { return true; }
   };
 
   template<>
   class From_Ruby<FanModeType>
   {
   public:
+    Convertible is_convertible(VALUE value) { return Convertible::Cast; }
+
     FanModeType convert(VALUE x)
     {
       auto s = String(x).str();
@@ -106,16 +104,15 @@ namespace Rice::detail
   template<>
   struct Type<NonlinearityType>
   {
-    static bool verify()
-    {
-      return true;
-    }
+    static bool verify() { return true; }
   };
 
   template<>
   class From_Ruby<NonlinearityType>
   {
   public:
+    Convertible is_convertible(VALUE value) { return Convertible::Cast; }
+
     NonlinearityType convert(VALUE x)
     {
       auto s = String(x).str();
@@ -150,16 +147,15 @@ namespace Rice::detail
   template<>
   struct Type<Scalar>
   {
-    static bool verify()
-    {
-      return true;
-    }
+    static bool verify() { return true; }
   };
 
   template<>
   class From_Ruby<Scalar>
   {
   public:
+    Convertible is_convertible(VALUE value) { return Convertible::Cast; }
+
     Scalar convert(VALUE x)
     {
       if (FIXNUM_P(x)) {

data/ext/torch/torch.cpp

@@ -21,7 +21,13 @@ torch::Tensor make_tensor(Rice::Array a, const std::vector<int64_t> &size, const
 }
 
 void init_torch(Rice::Module& m) {
-  register_handler<torch::Error>(handle_global_error);
+  Rice::detail::Registries::instance.handlers.set([]() {
+    try {
+      throw;
+    } catch (const torch::Error& ex) {
+      handle_global_error(ex);
+    }
+  });
   add_torch_functions(m);
   m.define_singleton_function(
       "grad_enabled?",

data/ext/torch/utils.h

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

data/lib/torch/distributions/distribution.rb

@@ -0,0 +1,26 @@
+module Torch
+  module Distributions
+    class Distribution
+      def initialize(batch_shape: [], event_shape: [], validate_args: nil)
+        @batch_shape = batch_shape
+        @event_shape = event_shape
+        if !validate_args.nil?
+          @validate_args = validate_args
+        end
+        if @validate_args
+          raise NotImplementedYet
+        end
+        super()
+      end
+
+      private
+
+      def _extended_shape(sample_shape: [])
+        if !sample_shape.is_a?(Array)
+          sample_shape = sample_shape.to_a
+        end
+        sample_shape + @batch_shape + @event_shape
+      end
+    end
+  end
+end

data/lib/torch/distributions/exponential_family.rb

@@ -0,0 +1,6 @@
+module Torch
+  module Distributions
+    class ExponentialFamily < Distribution
+    end
+  end
+end

data/lib/torch/distributions/normal.rb

@@ -0,0 +1,22 @@
+module Torch
+  module Distributions
+    class Normal < ExponentialFamily
+      def initialize(loc, scale, validate_args: nil)
+        @loc, @scale = Utils.broadcast_all(loc, scale)
+        if loc.is_a?(Numeric) && scale.is_a?(Numeric)
+          batch_shape = []
+        else
+          batch_shape = @loc.size
+        end
+        super(batch_shape:, validate_args:)
+      end
+
+      def sample(sample_shape: [])
+        shape = _extended_shape(sample_shape:)
+        Torch.no_grad do
+          Torch.normal(@loc.expand(shape), @scale.expand(shape))
+        end
+      end
+    end
+  end
+end

data/lib/torch/distributions/utils.rb

@@ -0,0 +1,10 @@
+module Torch
+  module Distributions
+    module Utils
+      def self.broadcast_all(*values)
+        # TODO improve
+        values
+      end
+    end
+  end
+end

data/lib/torch/tensor.rb

@@ -159,6 +159,7 @@ module Torch
 
     # TODO better compare?
     def <=>(other)
+      other = other.item if other.is_a?(Tensor)
       item <=> other
     end
 

data/lib/torch/version.rb

@@ -1,3 +1,3 @@
 module Torch
-  VERSION = "0.19.0"
+  VERSION = "0.20.0"
 end

data/lib/torch.rb

@@ -13,6 +13,12 @@ require_relative "torch/inspector"
 require_relative "torch/tensor"
 require_relative "torch/version"
 
+# distributions
+require_relative "torch/distributions/distribution"
+require_relative "torch/distributions/exponential_family"
+require_relative "torch/distributions/normal"
+require_relative "torch/distributions/utils"
+
 # optim
 require_relative "torch/optim/optimizer"
 require_relative "torch/optim/adadelta"

metadata

@@ -1,13 +1,13 @@
 --- !ruby/object:Gem::Specification
 name: torch-rb
 version: !ruby/object:Gem::Version
-  version: 0.19.0
+  version: 0.20.0
 platform: ruby
 authors:
 - Andrew Kane
 bindir: bin
 cert_chain: []
-date: 2025-01-30 00:00:00.000000000 Z
+date: 1980-01-02 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: rice
@@ -15,14 +15,14 @@ dependencies:
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 4.3.3
+        version: '4.5'
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 4.3.3
+        version: '4.5'
 email: andrew@ankane.org
 executables: []
 extensions:
@@ -64,6 +64,10 @@ files:
 - lib/torch-rb.rb
 - lib/torch.rb
 - lib/torch/device.rb
+- lib/torch/distributions/distribution.rb
+- lib/torch/distributions/exponential_family.rb
+- lib/torch/distributions/normal.rb
+- lib/torch/distributions/utils.rb
 - lib/torch/hub.rb
 - lib/torch/inspector.rb
 - lib/torch/nn/adaptive_avg_pool1d.rb
@@ -237,7 +241,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.6.2
+rubygems_version: 3.6.7
 specification_version: 4
 summary: Deep learning for Ruby, powered by LibTorch
 test_files: []