fix pytorch cuda mkl openblas new release version code compat

Author: Muller Zhang

build.sbt

@@ -56,7 +56,7 @@ lazy val commonSettings = 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.
-  tlJdkRelease := Some(11)
+  tlJdkRelease := Some(21)
 ) ++ tlReplaceCommandAlias(
   "tlReleaseLocal",
   List(

project/plugins.sbt

@@ -8,8 +8,8 @@ addSbtPlugin("com.github.sbt" % "sbt-release" % "1.4.0")
 
 addSbtPlugin("org.bytedeco" % "sbt-javacpp" % "1.17")
 addSbtPlugin("org.scalameta" % "sbt-mdoc" % "2.5.2")
-addSbtPlugin("org.typelevel" % "sbt-typelevel" % "0.6.5")
-addSbtPlugin("org.typelevel" % "sbt-typelevel-site" % "0.6.5")
+addSbtPlugin("org.typelevel" % "sbt-typelevel" % "0.7.7")
+addSbtPlugin("org.typelevel" % "sbt-typelevel-site" % "0.7.7")
 addSbtPlugin("org.xerial.sbt" % "sbt-sonatype" % "3.12.2")
 addSbtPlugin("org.scalameta" % "sbt-native-image" % "0.3.4")
 if (sys.env.isDefinedAt("GITHUB_ACTION")) {

storch_core/src/main/scala/torch/Tensor.scala

@@ -17,47 +17,11 @@
 package torch
 
 import torch.numpy.matrix.NDArray
-import org.bytedeco.javacpp.{
-  Pointer,
-  BoolPointer,
-  BytePointer,
-  DoublePointer,
-  FloatPointer,
-  IntPointer,
-  LongPointer,
-  ShortPointer
-}
+import org.bytedeco.javacpp.{BoolPointer, BytePointer, DoublePointer, FloatPointer, IntPointer, LongPointer, Pointer, ShortPointer}
 import org.bytedeco.pytorch
-import org.bytedeco.pytorch.{
-  Tensor as NativeTensor,
-  BoolOptional,
-  TensorOptions,
-  TensorBase,
-  DoubleOptional,
-  EllipsisIndexType,
-  Generator,
-  GeneratorOptional,
-  LongOptional,
-  NamedTensorMeta,
-  Node,
-  Quantizer,
-  ScalarOptional,
-  ScalarTypeOptional,
-  Storage,
-  SymInt,
-  SymIntOptional,
-  TensorArrayRefOptional,
-  TensorIndex,
-  TensorIndexArrayRef,
-  TensorIndexVector,
-  TensorOptional,
-  TensorOptionalList,
-  TensorTensorHook,
-  TensorVector,
-  VoidTensorHook
-}
+import org.bytedeco.pytorch.{BoolOptional, DoubleOptional, EllipsisIndexType, Generator, GeneratorOptional, LongOptional, NamedTensorMeta, Node, Quantizer, ScalarOptional, ScalarTypeOptional, Storage, SymInt, SymIntOptional, TensorArrayRefOptional, TensorBase, TensorIndex, TensorIndexArrayRef, TensorIndexVector, TensorOptional, TensorOptionalList, TensorOptions, TensorTensorHook, TensorVector, VoidTensorHook, Tensor as NativeTensor}
 import org.bytedeco.pytorch.global.torch as torchNative
-import org.bytedeco.pytorch.global.torch.ScalarType
+import org.bytedeco.pytorch.global.torch.{ScalarType,Backend}
 
 import java.nio.{Buffer, ByteBuffer, DoubleBuffer, FloatBuffer, IntBuffer, LongBuffer, ShortBuffer}
 import scala.collection.immutable.ArraySeq
@@ -1897,7 +1861,7 @@ sealed abstract class Tensor[D <: DType]( /* private[torch]  */ val native: pyto
 
   def put[D2 <: DType](x: Tensor[D2]): Tensor[Promoted[D, D2]] = fromNative(native.put(x.native))
 
-  def toBackend(b: Int): Tensor[D] = fromNative(native.toBackend(b))
+  def toBackend(b: Backend): Tensor[D] = fromNative(native.toBackend(b))
 
   def not(un_used: Int*): Tensor[D] = fromNative(native.not())
 

storch_core/src/main/scala/torch/amp/autocast.scala

@@ -1,3 +1,17 @@
 package torch.amp
 
-class autocast {}
+import org.bytedeco.pytorch.global.torch
+import org.bytedeco.pytorch.global.torch.DeviceType
+
+object autocast {
+  
+  def is_autocast_enabled(device_type: DeviceType): Boolean ={
+//    torch.set_autocast_enabled()
+//    torch.is_autocast_cache_enabled()
+//    torch.is_autocast_eligible()
+    torch.is_autocast_enabled(device_type)
+  }
+//  torch.is_autocast_enabled(DeviceType.CUDA)
+  
+  
+}

storch_core/src/main/scala/torch/cuda/TorchCUDAAllocator.scala

@@ -115,10 +115,11 @@ class TorchCUDAAllocator(val nativeAllocator: CUDAAllocator) extends Closeable {
     * @param device
     *   设备索引
     */
-  def ensureExistsAndIncrefPool(
+  def  createOrIncrefPool(
+    
       device: Byte,
       mempool_id: DeviceAssertionsDataVectorCUDAKernelLaunchInfoVectorPair
-  ): Unit = nativeAllocator.ensureExistsAndIncrefPool(device, mempool_id)
+  ): Unit = nativeAllocator.createOrIncrefPool(device, mempool_id)
 
   def checkPoolLiveAllocations(
       device: Byte,
@@ -132,19 +133,35 @@ class TorchCUDAAllocator(val nativeAllocator: CUDAAllocator) extends Closeable {
 
   def isHistoryEnabled = nativeAllocator.isHistoryEnabled
 
+  //  public native @Cast("bool") boolean isHistoryEnabled();
+  //  public native void recordHistory(
+  //        @Cast("bool") boolean enabled,
+  //        @ByVal @Cast("c10::cuda::CUDACachingAllocator::CreateContextFn*") Pointer context_recorder,
+  //        @Cast("size_t") long alloc_trace_max_entries,
+  //        RecordContext when,
+  //        @Cast("bool") boolean clearHistory);
+  //  public native void recordHistory(
+  //        @Cast("bool") boolean enabled,
+  //        @ByVal @Cast("c10::cuda::CUDACachingAllocator::CreateContextFn*") Pointer context_recorder,
+  //        @Cast("size_t") long alloc_trace_max_entries,
+  //        @Cast("c10::cuda::CUDACachingAllocator::RecordContext") int when,
+  //        @Cast("bool") boolean clearHistory);
+  
   def recordHistory(
       enabled: Boolean,
       context_recorder: Pointer,
       alloc_trace_max_entries: Long,
-      when: torch_cuda.RecordContext
-  ) = nativeAllocator.recordHistory(enabled, context_recorder, alloc_trace_max_entries, when)
+      when: torch_cuda.RecordContext,
+      clearHistory: Boolean
+  ) = nativeAllocator.recordHistory(enabled, context_recorder, alloc_trace_max_entries, when,clearHistory)
 
   def recordHistory(
       enabled: Boolean,
       context_recorder: Pointer,
       alloc_trace_max_entries: Long,
-      when: Int
-  ) = nativeAllocator.recordHistory(enabled, context_recorder, alloc_trace_max_entries, when)
+      when: Int,
+      clearHistory: Boolean
+  ) = nativeAllocator.recordHistory(enabled, context_recorder, alloc_trace_max_entries, when,clearHistory)
 
   def recordAnnotation(md: StringPair) = nativeAllocator.recordAnnotation(md)