torch-rb 0.15.0 → 0.16.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 5cb7d8bf760b3f2a52976b0fda929cd7227a89ed7475b19a87cd9ce53ab5fe4e
-  data.tar.gz: 9977b9740c8490b4be11682dbd700bfe1175cf1f109e81412a694c4a4fd4a043
+  metadata.gz: 8442fc0f85d6f2465258a54e5aefbe03d23a7c0e58753e06855bfebd2f4de802
+  data.tar.gz: ac0efb89f9b6d413498bfb1c2e84336aa728047dd013d00fa736449e5be82617
 SHA512:
-  metadata.gz: fb7283603fff1ad25fd234926abaeaafada2f0def3b5a52051242f3348445a5e4b615680f414d09c72a502c355fade701e79867444dd848c8dcd989f42359f19
-  data.tar.gz: bb41a341f271fb5bce47ad6e1a6ef223b6d7c01165a86de13681d25720c96d5a17de81e549439feedcf482c4e539083e3e04e97a768999438b939bc235fb652d
+  metadata.gz: 6830efe74de98fc8a8d23e7308795a60ee60fff72b3f82fa7cb92815f4efe52fdf3637e0821490f5e3e8c2c8731043f52f5aff20cfb01db1340be0962fed18db
+  data.tar.gz: 3e50976e5add37b4158956c76e3c922167911492acda9e171af42ad39d5abe946c36427e545d9fc820a2800e3df0523b0068ce76b804d0c05a6f1e2ad495de01

data/CHANGELOG.md

@@ -1,3 +1,9 @@
+## 0.16.0 (2024-06-12)
+
+- Updated LibTorch to 2.3.0
+- Added `ELU` and `GELU` classes
+- Dropped support for Ruby < 3.1
+
 ## 0.15.0 (2024-02-28)
 
 - Updated LibTorch to 2.2.0

data/README.md

@@ -410,6 +410,7 @@ Here’s the list of compatible versions.
 
 Torch.rb | LibTorch
 --- | ---
+0.16.x | 2.3.x
 0.15.x | 2.2.x
 0.14.x | 2.1.x
 0.13.x | 2.0.x

data/codegen/native_functions.yaml

@@ -134,7 +134,7 @@
   autogen: _new_zeros_with_same_feature_meta.out
 
 # This function compares the storage numel of self with that of other, where
-# storage numel is cumputed as: `other.storage().nbytes() / other.itemsize()`.
+# storage numel is computed as: `other.storage().nbytes() / other.itemsize()`.
 # We create this function for composite compliance purposes. The batching rule
 # always returns true because vmapped as_strided does not support accessing
 # storage locations not indexable by the input tensor.
@@ -175,12 +175,24 @@
     CPU: _assert_async_msg_cpu
     CUDA: _assert_async_msg_cuda
 
+- func: _assert_scalar(Scalar self, str assert_msg) -> ()
+  dispatch:
+    CompositeExplicitAutograd: _assert_scalar
+
+- func: _functional_assert_scalar(Scalar self, str assert_msg, Tensor dep_token) -> Tensor
+  dispatch:
+    CompositeExplicitAutograd: _functional_assert_scalar
+
 - func: _functional_assert_async.msg(Tensor self, str assert_msg, Tensor dep_token) -> Tensor
   dispatch:
     CPU: _functional_assert_async_msg_cpu
 
 - func: _assert_tensor_metadata(Tensor a, SymInt[]? size=None, SymInt[]? stride=None, ScalarType? dtype=None) -> ()
 
+- func: _print(str s) -> ()
+  dispatch:
+    CompositeExplicitAutograd: _print
+
 - func: sym_constrain_range(Scalar size, *, int? min=None, int? max=None) -> ()
   dispatch:
     CompositeExplicitAutograd: sym_constrain_range
@@ -470,6 +482,7 @@
 - func: conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU, CUDA: conj_physical_out
+    MPS: conj_physical_out_mps
     SparseCPU, SparseCUDA: conj_physical_out_sparse
     SparseCsrCPU, SparseCsrCUDA: conj_physical_sparse_csr_out
   tags: pointwise
@@ -564,8 +577,8 @@
   dispatch:
     SparseCPU: add_out_sparse_cpu
     SparseCUDA: add_out_sparse_cuda
-    SparseCsrCPU: add_out_sparse_csr_cpu
-    SparseCsrCUDA: add_out_sparse_csr_cuda
+    SparseCsrCPU: add_out_sparse_compressed_cpu
+    SparseCsrCUDA: add_out_sparse_compressed_cuda
     MkldnnCPU: mkldnn_add_out
     MPS: add_out_mps
   tags: pointwise
@@ -763,7 +776,7 @@
   dispatch:
     CompositeExplicitAutograd: arange
 
-# This operator should be named `aragne.start_out` if following the naming convention. However that
+# 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.
 # FIXME: enable this
 #- func: arange.start_out_(Scalar start, Scalar end, *, Tensor(a!) out) -> Tensor(a!)
@@ -1220,6 +1233,13 @@
     CompositeExplicitAutograd: copysign_out
   tags: pointwise
 
+- func: _lazy_clone(Tensor self) -> Tensor
+  # Like clone, but the copy takes place lazily, only if either the
+  # input or the output are written.
+  variants: function, method
+  dispatch:
+    CompositeExplicitAutograd: _lazy_clone
+
 - func: logical_not(Tensor self) -> Tensor
   device_check: NoCheck   # TensorIterator
   variants: function, method
@@ -1621,6 +1641,7 @@
 - func: complex.out(Tensor real, Tensor imag, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU, CUDA: complex_out
+    MPS: complex_out_mps
 
 - func: polar(Tensor abs, Tensor angle) -> Tensor
   variants: function
@@ -1847,7 +1868,10 @@
 - func: cudnn_convolution(Tensor self, Tensor weight, SymInt[] padding, SymInt[] stride, SymInt[] dilation, SymInt groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
   dispatch:
     CUDA: cudnn_convolution
-  autogen: cudnn_convolution.out
+
+- func: cudnn_convolution.out(Tensor self, Tensor weight, SymInt[] padding, SymInt[] stride, SymInt[] dilation, SymInt groups, bool benchmark, bool deterministic, bool allow_tf32, *, Tensor(a!) out) -> Tensor(a!)
+  dispatch:
+    CUDA: cudnn_convolution_out
 
 - func: cudnn_convolution_transpose(Tensor self, Tensor weight, SymInt[] padding, SymInt[] output_padding, SymInt[] stride, SymInt[] dilation, SymInt groups, bool benchmark, bool deterministic, bool allow_tf32) -> Tensor
   dispatch:
@@ -2346,7 +2370,7 @@
     Meta: empty_meta_symint
     MkldnnCPU: empty_mkldnn
     SparseCPU, SparseCUDA, SparseMeta: empty_sparse
-    SparseCsrCPU, SparseCsrCUDA: empty_sparse_compressed
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: empty_sparse_compressed
     QuantizedCPU, QuantizedCUDA, QuantizedMeta: empty_unknown_quantized
   tags: core
 
@@ -2452,7 +2476,7 @@
     CompositeExplicitAutograd: empty_like
     QuantizedCPU, QuantizedCUDA: empty_like_quantized
     SparseCPU, SparseCUDA, SparseMeta: empty_like_sparse_coo
-    SparseCsrCPU, SparseCsrCUDA: empty_like_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: empty_like_sparse_csr
     NestedTensorCPU, NestedTensorCUDA: empty_like_nested
   autogen: empty_like.out
 
@@ -2954,12 +2978,14 @@
   dispatch:
     CPU: _fft_r2c_mkl
     CUDA: _fft_r2c_cufft
+    MPS: _fft_r2c_mps
 
 - func: _fft_r2c.out(Tensor self, int[] dim, int normalization, bool onesided, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
     CPU: _fft_r2c_mkl_out
     CUDA: _fft_r2c_cufft_out
+    MPS: _fft_r2c_mps_out
 
 # Complex to real inverse FFT
 - func: _fft_c2r(Tensor self, int[] dim, int normalization, SymInt last_dim_size) -> Tensor
@@ -2967,12 +2993,14 @@
   dispatch:
     CPU: _fft_c2r_mkl
     CUDA: _fft_c2r_cufft
+    MPS: _fft_c2r_mps
 
 - func: _fft_c2r.out(Tensor self, int[] dim, int normalization, SymInt last_dim_size, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
     CPU: _fft_c2r_mkl_out
     CUDA: _fft_c2r_cufft_out
+    MPS: _fft_c2r_mps_out
 
 # Standard complex to complex FFT (forward or backward)
 - func: _fft_c2c(Tensor self, SymInt[] dim, int normalization, bool forward) -> Tensor
@@ -2980,12 +3008,14 @@
   dispatch:
     CPU: _fft_c2c_mkl
     CUDA: _fft_c2c_cufft
+    MPS: _fft_c2c_mps
 
 - func: _fft_c2c.out(Tensor self, SymInt[] dim, int normalization, bool forward, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   dispatch:
     CPU: _fft_c2c_mkl_out
     CUDA: _fft_c2c_cufft_out
+    MPS: _fft_c2c_mps_out
 
 - func: _validate_compressed_sparse_indices(bool is_crow, Tensor compressed_idx, Tensor plain_idx, int cdim, int dim, int nnz) -> ()
   device_check: NoCheck
@@ -3302,11 +3332,15 @@
   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) -> 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) -> Tensor
   dispatch:
     CUDA: _cslt_sparse_mm
 
-- func: _sparse_semi_structured_linear(Tensor input, Tensor weight, Tensor meta, *, Tensor? bias=None, str? activation=None) -> Tensor
+- 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:
+    CUDA: _cslt_sparse_mm_search
+
+- func: _sparse_semi_structured_linear(Tensor input, Tensor weight, Tensor meta, *, Tensor? bias=None, str? activation=None, ScalarType? out_dtype=None) -> Tensor
   dispatch:
     CUDA: _sparse_semi_structured_linear
 
@@ -4058,12 +4092,18 @@
 
 - func: _convert_weight_to_int4pack(Tensor self, int innerKTiles) -> Tensor
   dispatch:
+    CPU: _convert_weight_to_int4pack_cpu
     CUDA: _convert_weight_to_int4pack_cuda
 
 - func: _weight_int4pack_mm(Tensor self, Tensor mat2, int qGroupSize, Tensor qScaleAndZeros) -> Tensor
   dispatch:
+    CPU: _weight_int4pack_mm_cpu
     CUDA: _weight_int4pack_mm_cuda
 
+- func: _weight_int8pack_mm(Tensor self, Tensor mat2, Tensor scales) -> Tensor
+  dispatch:
+    CPU: _weight_int8pack_mm_cpu
+
 - func: _sparse_mm(Tensor sparse, Tensor dense) -> Tensor
   python_module: sparse
 
@@ -4439,7 +4479,6 @@
     MPS: pixel_shuffle_mps
     CompositeExplicitAutogradNonFunctional: math_pixel_shuffle
   autogen: pixel_shuffle.out
-  tags: core
 
 - func: pixel_unshuffle(Tensor self, int downscale_factor) -> Tensor
   dispatch:
@@ -4810,7 +4849,7 @@
   device_guard: False
   dispatch:
     CompositeImplicitAutograd: reshape_symint
-    CompositeImplicitAutogradNestedTensor: reshape_nested
+    CompositeImplicitAutogradNestedTensor: reshape_nested_symint
 
 - func: _reshape_copy(Tensor self, SymInt[] size) -> Tensor
   variants: function
@@ -4969,6 +5008,7 @@
   device_check: NoCheck   # TensorIterator
   python_module: nn
   dispatch:
+    QuantizedCPU: gelu_quantized_cpu_
     NestedTensorCPU, NestedTensorCUDA: NestedTensor_gelu_
 
 - func: gelu(Tensor self, *, str approximate='none') -> Tensor
@@ -5356,6 +5396,21 @@
     CompositeExplicitAutograd: slice_backward
   autogen: slice_backward.out
 
+# NB: This op exists to back the implementation of reverse view_funcs for various views (chunk,
+# slice.Tensor, split_with_sizes, et. al.). Currently, these are only used during fake-ification
+# of PT2 graph input subclass instances that are views. This means:
+# * This op shouldn't really show up in eager mode (so e.g. XLA shouldn't have to implement it)
+# * This op shouldn't show up in a PT2 graph (so a PT2 backend shouldn't have to implement it)
+# * A subclass will have to implement this to work in PT2 if a subclass view is used as a graph
+#   input AND the view utilizes this op in its inverse. The idea is that slice_inverse() is
+#   easier to implement for a subclass than as_strided()
+- func: slice_inverse(Tensor(a) self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a)
+  variants: function, method
+  device_check: NoCheck
+  device_guard: False
+  dispatch:
+    CompositeExplicitAutograd: slice_inverse_symint
+
 - func: slice_scatter(Tensor self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor
   variants: function, method
   device_check: NoCheck
@@ -5363,7 +5418,7 @@
   dispatch:
     CompositeExplicitAutogradNonFunctional: slice_scatter
   autogen: slice_scatter.out
-  tags: core
+  tags: [core, view_copy]
 
 - func: select_scatter(Tensor self, Tensor src, int dim, SymInt index) -> Tensor
   variants: function, method
@@ -5562,6 +5617,14 @@
     SparseCPU: _sspaddmm_out_cpu
     SparseCUDA: _sspaddmm_out_cuda
 
+- func: _chunk_cat(Tensor[] tensors, int dim, int num_chunks) -> Tensor
+  dispatch:
+    CompositeExplicitAutograd: _chunk_cat
+
+- func: _chunk_cat.out(Tensor[] tensors, int dim, int num_chunks, *, Tensor(a!) out) -> Tensor(a!)
+  dispatch:
+    CompositeExplicitAutograd: _chunk_cat_out
+
 - func: stack(Tensor[] tensors, int dim=0) -> Tensor
   dispatch:
     CompositeExplicitAutograd: stack
@@ -5753,6 +5816,7 @@
   variants: function
   dispatch:
     CPU, CUDA: std_mean
+    MPS: std_mean_mps
   autogen: std_mean.correction_out
 
 - func: std_mean.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> (Tensor, Tensor)
@@ -6008,7 +6072,6 @@
     CPU, MPS: roll
     CUDA: roll_cuda
   autogen: roll.out
-  tags: core
 
 # default int[] value [0,1] should not add space after comma, since codegen parser uses ', ' to split args
 
@@ -6091,6 +6154,52 @@
     CompositeExplicitAutogradNonFunctional: _nested_view_from_buffer_copy
   autogen: _nested_view_from_buffer_copy.out
 
+- func: _nested_view_from_jagged(Tensor(a) self, Tensor offsets, Tensor dummy, Tensor? lengths=None, int ragged_idx=1) -> Tensor(a)
+  variants: function
+  device_check: NoCheck
+  dispatch: {}
+
+- func: _nested_view_from_jagged_copy(Tensor self, Tensor offsets, Tensor dummy, Tensor? lengths=None, int ragged_idx=1) -> Tensor
+  variants: function
+  device_check: NoCheck
+  tags: view_copy
+  dispatch:
+    CompositeExplicitAutogradNonFunctional: _nested_view_from_jagged_copy
+  autogen: _nested_view_from_jagged_copy.out
+
+- func: _nested_get_values(Tensor(a) self) -> Tensor(a)
+  variants: function
+  device_check: NoCheck
+  dispatch: {}
+
+- func: _nested_get_values_copy(Tensor self) -> Tensor
+  variants: function
+  device_check: NoCheck
+  tags: view_copy
+  dispatch:
+    CompositeExplicitAutogradNonFunctional: _nested_get_values_copy
+  autogen: _nested_get_values_copy.out
+
+- func: _nested_get_offsets(Tensor self) -> Tensor
+  variants: function
+  device_check: NoCheck
+  dispatch: {}
+
+# returns undefined Tensor if no lengths present
+- func: _nested_get_lengths(Tensor self) -> Tensor
+  variants: function
+  device_check: NoCheck
+  dispatch: {}
+
+- func: _nested_get_ragged_idx(Tensor self) -> int
+  variants: function
+  device_check: NoCheck
+  dispatch: {}
+
+- func: _nested_get_jagged_dummy(Tensor any) -> Tensor
+  category_override: dummy
+  dispatch: {}
+
 - func: _trilinear(Tensor i1, Tensor i2, Tensor i3, int[] expand1, int[] expand2, int[] expand3, int[] sumdim, int unroll_dim=1) -> Tensor
   dispatch:
     # calls unsqueeze
@@ -6275,6 +6384,7 @@
   variants: function
   dispatch:
     CPU, CUDA: var_mean
+    MPS: var_mean_mps
   autogen: var_mean.correction_out
 
 - func: var_mean.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> (Tensor, Tensor)
@@ -6295,15 +6405,13 @@
   device_check: NoCheck   # TensorIterator
   variants: function, method
   dispatch:
-    CPU, CUDA: where
-    MPS: where_mps
+    CPU, CUDA, MPS: where
   tags: [core, pointwise]
 
 - func: where.self_out(Tensor condition, Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA: where_self_out
-    MPS: where_self_out_mps
+    CPU, CUDA, MPS: where_self_out
 
 - func: where.ScalarSelf(Tensor condition, Scalar self, Tensor other) -> Tensor
   variants: function
@@ -6644,7 +6752,7 @@
     MPS: zero_mps_
     Meta: zero_meta_
     SparseCPU, SparseCUDA, SparseMeta: zero_sparse_
-    SparseCsrCPU, SparseCsrCUDA: zero_sparse_csr_
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: zero_sparse_csr_
     MkldnnCPU: mkldnn_zero_
     NestedTensorCPU, NestedTensorCUDA: zero_nested_
   autogen: zero, zero.out
@@ -6934,7 +7042,7 @@
 # FIXME: would be nicer if TensorOptions was optional based; not adding default arguments for options given
 # the default would never make sense.
 
-- func: sparse_compressed_tensor.comp_plain_value_size(Tensor compressed_indices, Tensor plain_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
+- func: sparse_compressed_tensor.comp_plain_value_size(Tensor compressed_indices, Tensor plain_indices, Tensor values, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
   dispatch:
     CompositeExplicitAutograd: sparse_compressed_tensor
 
@@ -6951,7 +7059,10 @@
 - func: sparse_bsr_tensor.crow_col_value(Tensor crow_indices, Tensor col_indices, Tensor values, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
 - func: sparse_bsc_tensor.ccol_row_value(Tensor ccol_indices, Tensor row_indices, Tensor values, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=False) -> Tensor
 
-- func: _sparse_compressed_tensor_unsafe(Tensor compressed_indices, Tensor plain_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
+- func: _sparse_compressed_tensor_unsafe(Tensor compressed_indices, Tensor plain_indices, Tensor values, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
+  dispatch:
+    CompositeImplicitAutograd: _sparse_compressed_tensor_unsafe_symint
+
 - func: _sparse_csr_tensor_unsafe(Tensor crow_indices, Tensor col_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 - func: _sparse_csc_tensor_unsafe(Tensor ccol_indices, Tensor row_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
 - func: _sparse_bsr_tensor_unsafe(Tensor crow_indices, Tensor col_indices, Tensor values, int[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
@@ -7037,7 +7148,7 @@
   dispatch:
     CPU, CUDA: sparse_dim_strided
     SparseCPU, SparseCUDA, SparseMeta: sparse_dim_sparse
-    SparseCsrCPU, SparseCsrCUDA: sparse_dim_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sparse_dim_sparse_csr
   device_check: NoCheck
   device_guard: False
 
@@ -7054,7 +7165,7 @@
   dispatch:
     CPU, CUDA: dense_dim_strided
     SparseCPU, SparseCUDA, SparseMeta: dense_dim_sparse
-    SparseCsrCPU, SparseCsrCUDA: dense_dim_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: dense_dim_sparse_csr
   device_check: NoCheck
   device_guard: False
 
@@ -7070,7 +7181,7 @@
   variants: method
   dispatch:
     SparseCPU, SparseCUDA, SparseMeta: _nnz_sparse
-    SparseCsrCPU, SparseCsrCUDA: _nnz_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: _nnz_sparse_csr
   device_check: NoCheck
   device_guard: False
 
@@ -7133,7 +7244,7 @@
   variants: method
   dispatch:
     SparseCPU, SparseCUDA, SparseMeta: values_sparse
-    SparseCsrCPU, SparseCsrCUDA: values_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: values_sparse_csr
     NestedTensorCPU, NestedTensorCUDA: values_nested
     CompositeExplicitAutograd: values_default
   device_check: NoCheck
@@ -7142,7 +7253,7 @@
 - func: crow_indices(Tensor(a) self) -> Tensor(a)
   variants: method
   dispatch:
-    SparseCsrCPU, SparseCsrCUDA: crow_indices_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: crow_indices_sparse_csr
     CompositeExplicitAutograd: crow_indices_default
   device_check: NoCheck
   device_guard: False
@@ -7150,7 +7261,7 @@
 - func: col_indices(Tensor(a) self) -> Tensor(a)
   variants: method
   dispatch:
-    SparseCsrCPU, SparseCsrCUDA: col_indices_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: col_indices_sparse_csr
     CompositeExplicitAutograd: col_indices_default
   device_check: NoCheck
   device_guard: False
@@ -7158,7 +7269,7 @@
 - func: ccol_indices(Tensor(a) self) -> Tensor(a)
   variants: method
   dispatch:
-    SparseCsrCPU, SparseCsrCUDA: ccol_indices_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: ccol_indices_sparse_csr
     CompositeExplicitAutograd: ccol_indices_default
   device_check: NoCheck
   device_guard: False
@@ -7166,7 +7277,7 @@
 - func: row_indices(Tensor(a) self) -> Tensor(a)
   variants: method
   dispatch:
-    SparseCsrCPU, SparseCsrCUDA: row_indices_sparse_csr
+    SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: row_indices_sparse_csr
     CompositeExplicitAutograd: row_indices_default
   device_check: NoCheck
   device_guard: False
@@ -7675,6 +7786,7 @@
   dispatch:
     CPU, CUDA, Meta, MPS: set_
   autogen: set.source_Storage, set.source_Storage_out
+  tags: inplace_view
 
 - func: set_.source_Storage_storage_offset(Tensor(a!) self, Storage source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!)
   variants: method
@@ -7687,6 +7799,7 @@
     MPS: set_storage_mps_
     QuantizedCPU, QuantizedCUDA: set_storage_quantized_
   autogen: set.source_Storage_storage_offset, set.source_Storage_storage_offset_out
+  tags: inplace_view
 
 - func: set_.source_Tensor_storage_offset(Tensor(a!) self, Tensor source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!)
   variants: method
@@ -7694,6 +7807,7 @@
   device_guard: False
   dispatch:
     CompositeImplicitAutograd: set__symint
+  tags: inplace_view
 
 - func: set_.source_Tensor(Tensor(a!) self, Tensor source) -> Tensor(a!)
   variants: method
@@ -7702,6 +7816,7 @@
   dispatch:
     CPU, CUDA, Meta, MPS: set_tensor_
   autogen: set.source_Tensor, set.source_Tensor_out
+  tags: inplace_view
 
 - func: set_(Tensor(a!) self) -> Tensor(a!)
   variants: method
@@ -7711,6 +7826,7 @@
     Meta: set_meta_
     MPS: set_mps_
   autogen: set, set.out
+  tags: inplace_view
 
 # Not making it CompositeImplicitAutograd because lift
 # should be a primitive w.r.t. functorch
@@ -10112,12 +10228,14 @@
   variants: function
   dispatch:
     CUDA: _amp_foreach_non_finite_check_and_unscale_cuda_
+    CPU: _amp_foreach_non_finite_check_and_unscale_cpu_
   autogen: _amp_foreach_non_finite_check_and_unscale, _amp_foreach_non_finite_check_and_unscale.out
 
 - func: _amp_update_scale_(Tensor(a!) self, Tensor(b!) growth_tracker, Tensor found_inf, float scale_growth_factor, float scale_backoff_factor, int growth_interval) -> Tensor(a!)
   variants: function
   dispatch:
     CUDA: _amp_update_scale_cuda_
+    CPU: _amp_update_scale_cpu_
   autogen: _amp_update_scale, _amp_update_scale.out
 
     #- func: _cat(Tensor[] tensors, int dim=0) -> Tensor
@@ -12341,6 +12459,7 @@
   dispatch:
     CPU: upsample_linear1d_out_cpu
     CUDA: upsample_linear1d_out_cuda
+    MPS: upsample_linear1d_out_mps
 
 - func: upsample_linear1d(Tensor self, SymInt[1] output_size, bool align_corners, float? scales=None) -> Tensor
   python_module: nn
@@ -12352,6 +12471,7 @@
   dispatch:
     CPU: upsample_linear1d_backward_out_cpu
     CUDA: upsample_linear1d_backward_out_cuda
+    MPS: upsample_linear1d_backward_out_mps
 
 - func: upsample_linear1d_backward(Tensor grad_output, SymInt[1] output_size, SymInt[3] input_size, bool align_corners, float? scales=None) -> Tensor
   python_module: nn
@@ -12824,7 +12944,7 @@
     SparseMeta: isinf_sparse_meta
     SparseCsrCPU, SparseCsrCUDA: isinf_sparse_csr
   autogen: isinf.out
-  tags: core
+  tags: [core, pointwise]
 
 - func: record_stream(Tensor(a!) self, Stream s) -> ()
   variants: method
@@ -13750,11 +13870,18 @@
   dispatch:
     CPU, CUDA: linalg_eig_out
 
+- func: _linalg_eigvals(Tensor self) -> Tensor
+  python_module: linalg
+  dispatch:
+    CPU, CUDA: _linalg_eigvals
+
 - func: linalg_eigvals(Tensor self) -> Tensor
   python_module: linalg
 
 - func: linalg_eigvals.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   python_module: linalg
+  dispatch:
+    CPU, CUDA: linalg_eigvals_out
 
 # This function is exposes the `compute_v` flag, which is then used to implement `linalg.eigh` and
 # `linalg.eigvalsh` as composite functions that call this one
@@ -14058,6 +14185,12 @@
 # It is undocumented and should not be used outside of tests.
 - func: _test_serialization_subcmul(Tensor self, Tensor other, Scalar alpha=1) -> Tensor
 
+# Note: for testing COW materialization within `at::parallel_for` loop function
+- func: _test_parallel_materialize(Tensor self, int num_parallel, bool skip_first=False) -> Tensor
+  variants: function
+  dispatch:
+    CompositeExplicitAutograd: _test_parallel_materialize
+
 # Note: this function is only for testing.
 - func: _test_optional_intlist(Tensor values, int[]? addends) -> Tensor
   python_module: nn
@@ -14392,6 +14525,7 @@
   variants: function
   dispatch:
     CompositeExplicitAutograd: split_with_sizes_copy_out
+    CUDA: split_with_sizes_copy_out_cuda
 
 - func: view_copy(Tensor self, SymInt[] size) -> Tensor
   variants: function
@@ -14468,19 +14602,28 @@
 
 - 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)
   dispatch:
-    CPU: _scaled_dot_product_flash_attention_cpu
     CUDA: _scaled_dot_product_flash_attention_cuda
     NestedTensorCUDA: _scaled_dot_product_flash_attention_nestedtensor_cuda
   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
+  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)
   device_check: NoCheck
   variants: function
   dispatch:
-    CPU: _scaled_dot_product_flash_attention_backward_cpu
     CUDA: _scaled_dot_product_flash_attention_backward_cuda
     NestedTensorCUDA: _scaled_dot_product_flash_attention_backward_nested
 
+- func: _scaled_dot_product_flash_attention_for_cpu_backward(Tensor grad_out, Tensor query, Tensor key, Tensor value, Tensor out, Tensor logsumexp, float dropout_p, bool is_causal, *, Tensor? attn_mask=None, float? scale=None) -> (Tensor grad_query, Tensor grad_key, Tensor grad_value)
+  device_check: NoCheck
+  variants: function
+  dispatch:
+    CPU: _scaled_dot_product_flash_attention_cpu_backward
+
 - func: _scaled_dot_product_efficient_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, bool compute_log_sumexp, float dropout_p=0.0, bool is_causal=False, *, float? scale=None) -> (Tensor output, Tensor log_sumexp, Tensor philox_seed, Tensor philox_offset)
   dispatch:
     CUDA: _scaled_dot_product_efficient_attention_cuda
@@ -14493,6 +14636,11 @@
     CUDA: _scaled_dot_product_efficient_attention_backward_cuda
   tags: nondeterministic_seeded
 
+- func: _scaled_dot_product_cudnn_attention(Tensor query, Tensor key, Tensor value, float dropout_p=0.0, bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset)
+  dispatch:
+    CUDA: _scaled_dot_product_cudnn_attention_cuda
+  tags: nondeterministic_seeded
+
 - func: _flash_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? cum_seq_q, Tensor? cum_seq_k, SymInt max_q, SymInt max_k, float dropout_p, bool is_causal, bool return_debug_mask, *, float? scale=None) -> (Tensor output, Tensor softmax_logsumexp, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
   variants: function
   dispatch:
@@ -14505,8 +14653,8 @@
   dispatch:
     CUDA: _flash_attention_backward
 
-# Returns ouput, logsumexp if compute_logsumexp
-- func: _efficient_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? bias, Tensor? cu_seqlens_q, Tensor? cu_seqlens_k, int? max_seqlen_q, float dropout_p, int custom_mask_type, bool compute_log_sumexp=False, *, float? scale=None, Tensor? causal_diagonal=None, Tensor? seqlen_k=None) -> (Tensor output, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset, SymInt max_seqlen_batch_q, SymInt max_seqlen_batch_k)
+# Returns output, logsumexp if compute_logsumexp
+- func: _efficient_attention_forward(Tensor query, Tensor key, Tensor value, Tensor? bias, Tensor? cu_seqlens_q, Tensor? cu_seqlens_k, int? max_seqlen_q, int? max_seqlen_k, float dropout_p, int custom_mask_type, bool compute_log_sumexp=False, *, float? scale=None, Tensor? causal_diagonal=None, Tensor? seqlen_k=None) -> (Tensor output, Tensor logsumexp, Tensor philox_seed, Tensor philox_offset, SymInt max_seqlen_batch_q, SymInt max_seqlen_batch_k)
   variants: function
   dispatch:
     CUDA: _efficient_attention_forward
@@ -15345,6 +15493,22 @@
     CUDA: _fused_adamw_kernel_cuda_
   autogen: _fused_adamw.tensor_lr, _fused_adamw.tensor_lr_out
 
+- func: _fused_sgd_(Tensor(a!)[] self, Tensor(b!)[] grads, Tensor(c!)[] momentum_buffer_list, *, float weight_decay, float momentum, float lr, float dampening, bool nesterov, bool maximize, bool is_first_step, Tensor? grad_scale=None, Tensor? found_inf=None) -> ()
+  # Unlike "foreach" functions, lists of tensors should be guaranteed to be on the same device (for now).
+  variants: function
+  dispatch:
+    CUDA: _fused_sgd_kernel_cuda_
+  autogen: _fused_sgd, _fused_sgd.out
+
+- func: _fused_sgd_.tensor_lr(Tensor(a!)[] self, Tensor(b!)[] grads, Tensor(c!)[] momentum_buffer_list, *, float weight_decay, float momentum, Tensor lr, float dampening, bool nesterov, bool maximize, bool is_first_step, Tensor? grad_scale=None, Tensor? found_inf=None) -> ()
+  # Unlike "foreach" functions, lists of tensors should be guaranteed to be on the same device (for now).
+  # but still skip the device check as the Tensor LR can be on CPU
+  device_check: NoCheck
+  variants: function
+  dispatch:
+    CUDA: _fused_sgd_kernel_cuda_
+  autogen: _fused_sgd.tensor_lr, _fused_sgd.tensor_lr_out
+
 # This op is ONLY used by pytorch/XLA in functionalization, and should never show up in vanilla eager mode or in any pytorch tracing contexts.
 - func: _propagate_xla_data(Tensor input, Tensor output) -> ()
   variants: function

data/ext/torch/utils.h

@@ -6,7 +6,7 @@
 #include <rice/stl.hpp>
 
 static_assert(
-  TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 2,
+  TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR == 3,
   "Incompatible LibTorch version"
 );
 

data/lib/torch/nn/elu.rb

@@ -0,0 +1,20 @@
+module Torch
+  module NN
+    class ELU < Module
+      def initialize(alpha: 1, inplace: false)
+        super()
+        @alpha = alpha
+        @inplace = inplace
+      end
+
+      def forward(input)
+        F.elu(input, alpha: @alpha, inplace: @inplace)
+      end
+
+      def extra_inspect
+        inplace_str = @inplace ? ", inplace: true" : ""
+        format("alpha: %s", @alpha) + inplace_str
+      end
+    end
+  end
+end

data/lib/torch/nn/functional.rb

@@ -174,6 +174,18 @@ module Torch
 
         # activation layers
 
+        def elu(input, alpha: 1, inplace: false)
+          if inplace
+            NN.elu!(input, alpha)
+          else
+            NN.elu(input, alpha)
+          end
+        end
+
+        def gelu(input, approximate: 'none')
+          NN.gelu(input, approximate: approximate)
+        end
+
         def hardshrink(input, lambd = 0.5)
           Torch.hardshrink(input, lambd)
         end

data/lib/torch/nn/gelu.rb

@@ -0,0 +1,18 @@
+module Torch
+  module NN
+    class GELU < Module
+      def initialize(approximate: 'none')
+        super()
+        @approximate = approximate
+      end
+
+      def forward(input)
+        F.gelu(input, approximate: @approximate)
+      end
+
+      def extra_inspect
+        "approximate: #{@approximate.inspect}"
+      end
+    end
+  end
+end

data/lib/torch/nn/leaky_relu.rb

@@ -13,7 +13,7 @@ module Torch
 
       def extra_inspect
         inplace_str = @inplace ? ", inplace: true" : ""
-        format("negative_slope: %s%s", @negative_slope, inplace_str)
+        format("negative_slope: %s", @negative_slope) + inplace_str
       end
     end
   end

data/lib/torch/version.rb

@@ -1,3 +1,3 @@
 module Torch
-  VERSION = "0.15.0"
+  VERSION = "0.16.0"
 end

data/lib/torch.rb

@@ -123,6 +123,8 @@ require_relative "torch/nn/dropout3d"
 require_relative "torch/nn/feature_alpha_dropout"
 
 # nn activations
+require_relative "torch/nn/elu"
+require_relative "torch/nn/gelu"
 require_relative "torch/nn/hardshrink"
 require_relative "torch/nn/leaky_relu"
 require_relative "torch/nn/log_sigmoid"

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: torch-rb
 version: !ruby/object:Gem::Version
-  version: 0.15.0
+  version: 0.16.0
 platform: ruby
 authors:
 - Andrew Kane
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2024-02-29 00:00:00.000000000 Z
+date: 2024-06-13 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: rice
@@ -43,24 +43,17 @@ files:
 - ext/torch/ext.cpp
 - ext/torch/extconf.rb
 - ext/torch/fft.cpp
-- ext/torch/fft_functions.h
 - ext/torch/generator.cpp
 - ext/torch/ivalue.cpp
 - ext/torch/linalg.cpp
-- ext/torch/linalg_functions.h
 - ext/torch/nn.cpp
-- ext/torch/nn_functions.h
 - ext/torch/random.cpp
 - ext/torch/ruby_arg_parser.cpp
 - ext/torch/ruby_arg_parser.h
-- ext/torch/sparse_functions.h
 - ext/torch/special.cpp
-- ext/torch/special_functions.h
 - ext/torch/templates.h
 - ext/torch/tensor.cpp
-- ext/torch/tensor_functions.h
 - ext/torch/torch.cpp
-- ext/torch/torch_functions.h
 - ext/torch/utils.h
 - ext/torch/wrap_outputs.h
 - lib/torch-rb.rb
@@ -103,12 +96,14 @@ files:
 - lib/torch/nn/dropout2d.rb
 - lib/torch/nn/dropout3d.rb
 - lib/torch/nn/dropoutnd.rb
+- lib/torch/nn/elu.rb
 - lib/torch/nn/embedding.rb
 - lib/torch/nn/embedding_bag.rb
 - lib/torch/nn/feature_alpha_dropout.rb
 - lib/torch/nn/fold.rb
 - lib/torch/nn/functional.rb
 - lib/torch/nn/functional_attention.rb
+- lib/torch/nn/gelu.rb
 - lib/torch/nn/group_norm.rb
 - lib/torch/nn/gru.rb
 - lib/torch/nn/hardshrink.rb
@@ -230,14 +225,14 @@ required_ruby_version: !ruby/object:Gem::Requirement
   requirements:
   - - ">="
     - !ruby/object:Gem::Version
-      version: '3'
+      version: '3.1'
 required_rubygems_version: !ruby/object:Gem::Requirement
   requirements:
   - - ">="
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.5.3
+rubygems_version: 3.5.11
 signing_key:
 specification_version: 4
 summary: Deep learning for Ruby, powered by LibTorch

data/ext/torch/fft_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_fft_functions(Rice::Module& m);

data/ext/torch/linalg_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_linalg_functions(Rice::Module& m);

data/ext/torch/nn_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_nn_functions(Rice::Module& m);

data/ext/torch/sparse_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_sparse_functions(Rice::Module& m);

data/ext/torch/special_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_special_functions(Rice::Module& m);

data/ext/torch/tensor_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_tensor_functions(Rice::Module& m);

data/ext/torch/torch_functions.h

@@ -1,6 +0,0 @@
-// generated by rake generate:functions
-// do not edit by hand
-
-#pragma once
-
-void add_torch_functions(Rice::Module& m);