@tqchen Hi! What should I do next for this PR? For the release of DLPack, here are the implements of TVM, PyTorch and Cupy.
[TVM src/runtime/ndarray.cc#L148](https://github.com/dmlc/tvm/blob/master/src/runtime/ndarray.cc#L148) ```c++ NDArray NDArray::FromDLPack(DLManagedTensor* tensor) { NDArray::Container* data = new NDArray::Container(); data->deleter = Internal::DLPackDeleter; data->manager_ctx = tensor; data->dl_tensor = tensor->dl_tensor; return NDArray(data); } ``` [PyTorch aten/src/ATen/DLConvertor.cpp#L170](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/DLConvertor.cpp#L170) ```C++ Tensor fromDLPack(const DLManagedTensor* src) { Backend backend = getATenBackend(src->dl_tensor.ctx); ScalarType stype = toScalarType(src->dl_tensor.dtype); auto deleter = [src](void * self) { src->deleter(const_cast<DLManagedTensor*>(src)); }; return getType(backend, stype).tensorFromBlob( src->dl_tensor.data, IntList(src->dl_tensor.shape, src->dl_tensor.ndim), IntList(src->dl_tensor.strides, src->dl_tensor.ndim), deleter); } ``` [CuPy cupy/core/dlpack.pyx#L231](https://github.com/cupy/cupy/blob/master/cupy/core/dlpack.pyx#L231) ```python mem = DLPackMemory(dltensor) mem_ptr = memory.MemoryPointer(mem, mem.dlm_tensor.dl_tensor.byte_offset) cupy_array = ndarray(shape_vec, cp_dtype, mem_ptr) ``` The all NDArray save the deleter of dlpack or DLManagedTensor. So I added a new member `dlpack` for mx.nd.NDArray (Python) for store of dlpack. When the NDArray release, the dlpack will call the deleter. Dose NDArray(C++) manage the release of dltensor? [ Full content available at: https://github.com/apache/incubator-mxnet/pull/12047 ] This message was relayed via gitbox.apache.org for [email protected]
