I’m trying to compile a specific version of PyTorch (2.1.2) from source in Linux, but I’m facing the following issue:
[1861/1881] Building CXX object caffe2/torch/CMakeFiles/torch_python.dir/csrc/utils/tensor_numpy.cpp.o
FAILED: caffe2/torch/CMakeFiles/torch_python.dir/csrc/utils/tensor_numpy.cpp.o
/src/pytorch/torch/csrc/utils/tensor_numpy.cpp: In function ‘at::Tensor torch::utils::tensor_from_cuda_array_interface(PyObject*)’:
/src/pytorch/torch/csrc/utils/tensor_numpy.cpp:404:34: error: ‘PyArray_Descr’ has no member named ‘elsize’
404 | dtype_size_in_bytes = descr->elsize;
| ^~~~~~
Maybe the cause is an incompatible version of NumPy that is being pulled in.
ps: my environment is actually a bit different. It’s a slightly patched PyTorch being built on riscv64 using qemu. But maybe the error is not related.