Fix dlpack deletion (#47310)

* fix dlpack deletion

* add unittest

* fix unittest

paddle/fluid/framework/tensor_util.cc

@@ -1108,6 +1108,48 @@ void TensorFromDLPack(const ::DLTensor& dl_tensor, phi::DenseTensor* dst) {
 #endif
 }
 
+void TensorFromDLPack(const DLManagedTensor* src, phi::DenseTensor* dst) {
+  std::vector<int64_t> vec;
+  std::copy(src->dl_tensor.shape,
+            src->dl_tensor.shape + src->dl_tensor.ndim,
+            std::back_inserter(vec));
+
+  framework::DDim vddim = phi::make_ddim(vec);
+  dst->Resize(vddim);
+  ::DLDataType type = src->dl_tensor.dtype;
+
+  auto src_ptr = static_cast<const void*>(src->dl_tensor.data);
+  auto size = phi::product(vddim) * type.bits / 8;
+
+  if (src->dl_tensor.device.device_type == kDLCPU) {
+    platform::CPUPlace dst_place = platform::CPUPlace();
+    platform::CPUPlace src_place = platform::CPUPlace();
+    void* dst_ptr = GetDstPtrByDLDataType(type, dst, dst_place);
+    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
+  }
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+  if (src->dl_tensor.device.device_type == kDLGPU) {
+    platform::CUDAPlace dst_place =
+        platform::CUDAPlace(src->dl_tensor.device.device_id);
+    platform::CUDAPlace src_place =
+        platform::CUDAPlace(src->dl_tensor.device.device_id);
+    void* dst_ptr = GetDstPtrByDLDataType(type, dst, dst_place);
+    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(dst_place);
+    // Fix copy by share allocation.
+    memory::Copy(dst_place,
+                 dst_ptr,
+                 src_place,
+                 src_ptr,
+                 size,
+                 reinterpret_cast<const phi::GPUContext&>(*ctx).stream());
+  }
+#endif
+  src->deleter(const_cast<DLManagedTensor*>(src));
+#ifdef PADDLE_WITH_XPU
+  PADDLE_THROW(platform::errors::Unimplemented("XPUPlace is not supported"));
+#endif
+}
+
 template <typename T>
 std::string format_tensor(const phi::DenseTensor& tensor) {
   // TODO(zhiqiu): use the print option to format tensor.

paddle/fluid/framework/tensor_util.h

@@ -112,7 +112,9 @@ template <typename T>
 void TesnorToVector(const phi::DenseTensor& src, std::vector<T>* dst);
 
 // convert dlpack's DLTensor to tensor
+
 void TensorFromDLPack(const ::DLTensor& dl_tensor, phi::DenseTensor* dst);
+void TensorFromDLPack(const DLManagedTensor* src, phi::DenseTensor* dst);
 
 //
 // The implementation of template functions.

paddle/fluid/pybind/pybind.cc

@@ -703,11 +703,11 @@ PYBIND11_MODULE(libpaddle, m) {
     phi::DenseTensor tensor;
 
     if (dl.device.device_type == kDLCPU) {
-      paddle::framework::TensorFromDLPack(dl, &tensor);
+      paddle::framework::TensorFromDLPack(dmt, &tensor);
     }
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
     if (dl.device.device_type == kDLGPU) {
-      paddle::framework::TensorFromDLPack(dl, &tensor);
+      paddle::framework::TensorFromDLPack(dmt, &tensor);
     }
 #endif
     return tensor;

python/paddle/tests/test_dlpack.py

@@ -123,6 +123,14 @@ class TestDLPack(unittest.TestCase):
             self.func_test_dlpack_dtype_conversion()
         self.func_test_dlpack_dtype_conversion()
 
+    def test_dlpack_deletion(self):
+        # See Paddle issue 47171
+        if paddle.is_compiled_with_cuda():
+            for i in range(80):
+                a = paddle.rand(shape=[1024 * 128, 1024], dtype="float32")
+                dlpack = paddle.utils.dlpack.to_dlpack(a)
+                b = paddle.utils.dlpack.from_dlpack(dlpack)
+
 
 class TestRaiseError(unittest.TestCase):
     def func_test_from_dlpack_raise_type_error(self):