New pytorch presets

build.sbt

@@ -29,7 +29,8 @@ val cudaVersion = "12.1-8.9"
 val openblasVersion = "0.3.23"
 val mklVersion = "2023.1"
 ThisBuild / scalaVersion := "3.3.0"
-ThisBuild / javaCppVersion := "1.5.9"
+ThisBuild / javaCppVersion := "1.5.10-SNAPSHOT"
+ThisBuild / resolvers ++= Resolver.sonatypeOssRepos("snapshots")
 
 ThisBuild / githubWorkflowJavaVersions := Seq(JavaSpec.temurin("11"))
 
@@ -40,8 +41,7 @@ ThisBuild / enableGPU := false
 lazy val commonSettings = Seq(
   Compile / doc / scalacOptions ++= Seq("-groups", "-snippet-compiler:compile"),
   javaCppVersion := (ThisBuild / javaCppVersion).value,
-  javaCppPlatform := Seq(),
-  resolvers ++= Resolver.sonatypeOssRepos("snapshots")
+  javaCppPlatform := Seq()
   // This is a hack to avoid depending on the native libs when publishing
   // but conveniently have them on the classpath during development.
   // There's probably a cleaner way to do this.
@@ -75,8 +75,7 @@ lazy val core = project
       (if (enableGPU.value) "pytorch-gpu" else "pytorch") -> pytorchVersion,
       "mkl" -> mklVersion,
       "openblas" -> openblasVersion
-      // TODO remove cuda (not cuda-redist) once https://github.com/bytedeco/javacpp-presets/issues/1376 is fixed
-    ) ++ (if (enableGPU.value) Seq("cuda-redist" -> cudaVersion, "cuda" -> cudaVersion) else Seq()),
+    ) ++ (if (enableGPU.value) Seq("cuda-redist" -> cudaVersion) else Seq()),
     javaCppPlatform := org.bytedeco.sbt.javacpp.Platform.current,
     fork := true,
     Test / fork := true,

core/src/main/scala/torch/Tensor.scala

@@ -53,7 +53,6 @@ import spire.math.{Complex, UByte}
 import scala.reflect.Typeable
 import internal.NativeConverters
 import internal.NativeConverters.toArray
-import internal.LoadCusolver
 import Device.CPU
 import Layout.Strided
 import org.bytedeco.pytorch.ByteArrayRef
@@ -342,7 +341,7 @@ sealed abstract class Tensor[D <: DType]( /* private[torch]  */ val native: pyto
     import ScalarType.*
     val out = native.dtype().toScalarType().intern() match
       case Byte        => UByte(native.item_int())
-      case Char        => native.item_byte()
+      case Char        => native.item_char()
       case Short       => native.item_short()
       case Int         => native.item_int()
       case Long        => native.item_long()
@@ -357,7 +356,7 @@ sealed abstract class Tensor[D <: DType]( /* private[torch]  */ val native: pyto
         val b = native.contiguous.createBuffer[DoubleBuffer]
         Complex(b.get(), b.get())
       case Bool                   => native.item().toBool
-      case QInt8                  => native.item_byte()
+      case QInt8                  => native.item_char()
       case QUInt8                 => native.item_short()
       case QInt32                 => native.item_int()
       case BFloat16               => native.item().toBFloat16.asFloat()
@@ -802,7 +801,6 @@ type IntTensor = UInt8Tensor | Int8Tensor | Int16Tensor | Int32Tensor | Int64Ten
 type ComplexTensor = Complex32Tensor | Complex64Tensor | Complex128Tensor
 
 object Tensor:
-  LoadCusolver // TODO workaround for https://github.com/bytedeco/javacpp-presets/issues/1376
 
   def apply[D <: DType](native: pytorch.Tensor): Tensor[D] = (native.scalar_type().intern() match
     case ScalarType.Byte          => new UInt8Tensor(native)

core/src/main/scala/torch/cuda/package.scala

@@ -17,13 +17,11 @@
 package torch
 
 import org.bytedeco.pytorch.global.torch as torchNative
-import torch.internal.LoadCusolver
 
 /** This package adds support for CUDA tensor types, that implement the same function as CPU
   * tensors, but they utilize GPUs for computation.
   */
 package object cuda {
-  LoadCusolver
 
   /** Returns a Boolean indicating if CUDA is currently available. */
   def isAvailable: Boolean = torchNative.cuda_is_available()

core/src/main/scala/torch/internal/NativeConverters.scala

@@ -35,12 +35,9 @@ import org.bytedeco.pytorch.GenericDictIterator
 import spire.math.Complex
 import spire.math.UByte
 import scala.annotation.targetName
-import internal.LoadCusolver
 
 private[torch] object NativeConverters:
 
-  LoadCusolver // TODO workaround for https://github.com/bytedeco/javacpp-presets/issues/1376
-
   inline def convertToOptional[T, U <: T | Option[T], V >: Null](i: U, f: T => V): V = i match
     case i: Option[T] => i.map(f(_)).orNull
     case i: T         => f(i)

core/src/main/scala/torch/nn/functional/package.scala

@@ -18,7 +18,6 @@ package torch
 package nn
 
 import functional.*
-import torch.internal.LoadCusolver
 
 /** @groupname nn_conv Convolution functions
   * @groupname nn_pooling Pooling functions
@@ -38,6 +37,4 @@ package object functional
     with Linear
     with Loss
     with Pooling
-    with Sparse {
-  LoadCusolver
-}
+    with Sparse

core/src/main/scala/torch/nn/init.scala

@@ -21,7 +21,7 @@ import org.bytedeco.pytorch.global.torch as torchNative
 import org.bytedeco.pytorch.FanModeType
 import org.bytedeco.pytorch.kFanIn
 import org.bytedeco.pytorch.kFanOut
-import org.bytedeco.pytorch.NonlinearityType
+import org.bytedeco.pytorch.Nonlinearity as NonlinearityNative
 import org.bytedeco.pytorch.kLinear
 import org.bytedeco.pytorch.kConv1D
 import org.bytedeco.pytorch.kConv2D
@@ -33,11 +33,9 @@ import org.bytedeco.pytorch.kSigmoid
 import org.bytedeco.pytorch.kReLU
 import org.bytedeco.pytorch.kLeakyReLU
 import org.bytedeco.pytorch.Scalar
-import torch.internal.LoadCusolver
 
 // TODO implement remaining init functions
 object init:
-  LoadCusolver
   def kaimingNormal_(
       t: Tensor[?],
       a: Double = 0,
@@ -56,7 +54,7 @@ object init:
   enum NonLinearity:
     case Linear, Conv1D, Conv2D, Conv3D, ConvTranspose1D, ConvTranspose2D, ConvTranspose3D, Sigmoid,
       ReLU, LeakyReLU
-    private[torch] def toNative: NonlinearityType = NonlinearityType(this match
+    private[torch] def toNative: NonlinearityNative = NonlinearityNative(this match
       case NonLinearity.Linear          => kLinear()
       case NonLinearity.Conv1D          => kConv1D()
       case NonLinearity.Conv2D          => kConv2D()

core/src/main/scala/torch/nn/modules/Module.scala

@@ -72,7 +72,7 @@ abstract class Module {
 
   def copy(): this.type =
     val clone = super.clone().asInstanceOf[Module]
-    clone._nativeModule = _nativeModule.clone()
+    clone._nativeModule = _nativeModule.clone(null)
     clone.asInstanceOf[this.type]
 
   protected[torch] def registerWithParent[T <: pytorch.Module](parent: T)(using
@@ -100,15 +100,15 @@ abstract class Module {
 
   def to(device: Device): this.type =
     // val nativeCopy = nativeModule.clone()
-    nativeModule.asModule.to(device.toNative)
+    nativeModule.asModule.to(device.toNative, false)
     // copy
     // val clone: this.type = copy()
     // clone.nativeModule = nativeCopy
     this
 
   def to(dtype: DType, nonBlocking: Boolean = false): this.type =
-    val nativeCopy = nativeModule.clone()
-    nativeCopy.asModule.to(dtype.toScalarType)
+    val nativeCopy = nativeModule.clone(null)
+    nativeCopy.asModule.to(dtype.toScalarType, false)
     this
 
   def save(outputArchive: OutputArchive) = nativeModule.save(outputArchive)

core/src/main/scala/torch/nn/modules/batchnorm/BatchNorm2d.scala

@@ -119,7 +119,7 @@ final class BatchNorm2d[ParamType <: FloatNN | ComplexNN: Default](
   options.track_running_stats().put(trackRunningStats)
 
   override private[torch] val nativeModule: BatchNorm2dImpl = BatchNorm2dImpl(options)
-  nativeModule.asModule.to(paramType.toScalarType)
+  nativeModule.asModule.to(paramType.toScalarType, false)
 
   override def registerWithParent[M <: pytorch.Module](parent: M)(using
       name: sourcecode.Name

core/src/main/scala/torch/nn/modules/conv/Conv2d.scala

@@ -59,7 +59,7 @@ final class Conv2d[ParamType <: FloatNN | ComplexNN: Default](
   options.padding_mode().put(paddingModeNative)
 
   override private[torch] val nativeModule: Conv2dImpl = Conv2dImpl(options)
-  nativeModule.asModule.to(paramType.toScalarType)
+  nativeModule.asModule.to(paramType.toScalarType, false)
 
   override def registerWithParent[M <: pytorch.Module](parent: M)(using
       name: sourcecode.Name

core/src/main/scala/torch/nn/modules/linear/Linear.scala

@@ -28,7 +28,7 @@ import torch.nn.modules.{HasParams, TensorModule}
   *
   * This module supports `TensorFloat32<tf32_on_ampere>`.
   *
-  * * Example:
+  * Example:
   *
   * ```scala sc:nocompile
   * import torch.*
@@ -57,7 +57,7 @@ final class Linear[ParamType <: FloatNN: Default](
   private val options = new LinearOptions(inFeatures, outFeatures)
   options.bias().put(bias)
   override private[torch] val nativeModule: LinearImpl = new LinearImpl(options)
-  nativeModule.asModule.to(paramType.toScalarType)
+  nativeModule.asModule.to(paramType.toScalarType, false)
 
   override def registerWithParent[T <: pytorch.Module](parent: T)(using
       name: sourcecode.Name

core/src/main/scala/torch/nn/modules/pooling/MaxPool2d.scala

@@ -44,7 +44,7 @@ final class MaxPool2d[ParamType <: BFloat16 | Float32 | Float64: Default](
   options.ceil_mode().put(ceilMode)
 
   override private[torch] val nativeModule: MaxPool2dImpl = MaxPool2dImpl(options)
-  nativeModule.asModule.to(paramType.toScalarType)
+  nativeModule.asModule.to(paramType.toScalarType, false)
 
   override def registerWithParent[M <: pytorch.Module](parent: M)(using
       name: sourcecode.Name

core/src/main/scala/torch/nn/package.scala

@@ -16,16 +16,13 @@
 
 package torch
 
-import torch.internal.LoadCusolver
-
 /** These are the basic building blocks for graphs.
   *
   * @groupname nn_conv Convolution Layers
   * @groupname nn_linear Linear Layers
   * @groupname nn_utilities Utilities
   */
 package object nn {
-  LoadCusolver
 
   export modules.Module
   export modules.Default

core/src/main/scala/torch/nn/utils.scala

@@ -19,10 +19,8 @@ package nn
 
 import org.bytedeco.pytorch.global.torch as torchNative
 import org.bytedeco.pytorch.TensorVector
-import torch.internal.LoadCusolver
 
 object utils:
-  LoadCusolver
   def clipGradNorm_(
       parameters: Seq[Tensor[?]],
       max_norm: Double,

core/src/main/scala/torch/ops/IndexingSlicingJoiningOps.scala

@@ -20,14 +20,18 @@ package ops
 import internal.NativeConverters.*
 
 import org.bytedeco.pytorch.global.torch as torchNative
-import org.bytedeco.pytorch.{TensorArrayRef, TensorVector}
+import org.bytedeco.pytorch.TensorArrayRef
+import org.bytedeco.pytorch.TensorVector
 
 /** Indexing, Slicing, Joining, Mutating Ops
   *
   * https://pytorch.org/docs/stable/torch.html#indexing-slicing-joining-mutating-ops
   */
 private[torch] trait IndexingSlicingJoiningOps {
 
+  private def toArrayRef(tensors: Seq[Tensor[?]]): TensorArrayRef =
+    new TensorArrayRef(new TensorVector(tensors.map(_.native)*))
+
   /** Returns a view of the tensor conjugated and with the last two dimensions transposed.
     *
     * `x.adjoint()` is equivalent to `x.transpose(-2, -1).conj()` for complex tensors and to
@@ -120,7 +124,7 @@ private[torch] trait IndexingSlicingJoiningOps {
     * @group indexing_slicing_joining_mutating_ops
     */
   def cat[D <: DType](tensors: Seq[Tensor[D]], dim: Int = 0): Tensor[D] = Tensor(
-    torchNative.cat(new TensorArrayRef(new TensorVector(tensors.map(_.native)*)), dim.toLong)
+    torchNative.cat(toArrayRef(tensors), dim.toLong)
   )
 
   /** Returns a view of `input` with a flipped conjugate bit. If `input` has a non-complex dtype,
@@ -292,9 +296,8 @@ private[torch] trait IndexingSlicingJoiningOps {
     *
     * @group indexing_slicing_joining_mutating_ops
     */
-  def columnStack[D <: DType](tensors: Seq[Tensor[D]]): Tensor[D] = Tensor(
-    torchNative.column_stack(new TensorArrayRef(new TensorVector(tensors.map(_.native)*)))
-  )
+  def columnStack[D <: DType](tensors: Seq[Tensor[D]]): Tensor[D] =
+    Tensor(torchNative.column_stack(toArrayRef(tensors)))
 
   /** Stack tensors in sequence depthwise (along third axis).
     *
@@ -323,7 +326,7 @@ private[torch] trait IndexingSlicingJoiningOps {
     * @group indexing_slicing_joining_mutating_ops
     */
   def dstack[D <: DType](tensors: Seq[Tensor[D]]): Tensor[D] = Tensor(
-    torchNative.dstack(new TensorArrayRef(new TensorVector(tensors.map(_.native)*)))
+    torchNative.dstack(toArrayRef(tensors))
   )
 
   /** Gathers values along an axis specified by `dim`.
@@ -427,9 +430,8 @@ private[torch] trait IndexingSlicingJoiningOps {
     *
     * @group indexing_slicing_joining_mutating_ops
     */
-  def hstack[D <: DType](tensors: Seq[Tensor[D]]): Tensor[D] = Tensor(
-    torchNative.hstack(new TensorArrayRef(new TensorVector(tensors.map(_.native)*)))
-  )
+  def hstack[D <: DType](tensors: Seq[Tensor[D]]): Tensor[D] =
+    Tensor(torchNative.hstack(toArrayRef(tensors)))
 
   /** Accumulate the elements of `source` into the `input` tensor by adding to the indices in the
     * order given in `index`.
@@ -1080,7 +1082,7 @@ private[torch] trait IndexingSlicingJoiningOps {
     *   tensors (inclusive)
     */
   def stack[D <: DType](tensors: Seq[Tensor[D]], dim: Int = 0): Tensor[D] = Tensor(
-    torchNative.stack(new TensorArrayRef(new TensorVector(tensors.map(_.native)*)), dim)
+    torchNative.stack(toArrayRef(tensors), dim)
   )
 
   /** Alias for `torch.transpose`.
@@ -1491,7 +1493,7 @@ private[torch] trait IndexingSlicingJoiningOps {
     * @group indexing_slicing_joining_mutating_ops
     */
   def vstack[D <: DType](tensors: Seq[Tensor[D]]): Tensor[D] = Tensor(
-    torchNative.vstack(new TensorArrayRef(new TensorVector(tensors.map(_.native)*)))
+    torchNative.vstack(toArrayRef(tensors))
   )
 
   /** Return a tensor of elements selected from either `input` or `other`, depending on `condition`.

core/src/main/scala/torch/ops/package.scala

@@ -19,7 +19,7 @@ package torch
 import internal.NativeConverters
 import org.bytedeco.pytorch.global.torch as torchNative
 import org.bytedeco.pytorch
-import org.bytedeco.pytorch.MemoryFormatOptional
+import org.bytedeco.pytorch.{MemoryFormatOptional, TensorArrayRef, TensorVector}
 
 package object ops {
 

core/src/main/scala/torch/package.scala

@@ -14,8 +14,6 @@
  * limitations under the License.
  */
 
-import torch.internal.LoadCusolver
-
 import scala.util.Using
 
 /** The torch package contains data structures for multi-dimensional tensors and defines
@@ -37,7 +35,6 @@ package object torch
     with ops.PointwiseOps
     with ops.RandomSamplingOps
     with ops.ReductionOps {
-  LoadCusolver // TODO workaround for https://github.com/bytedeco/javacpp-presets/issues/1376
 
   /** Disable gradient calculation for [[op]].
     *