Change in Multiprocessing between 1.0 alpha and 1.0 dev

Hey PyTorch Community,

I tried rerunning my existing code base, which is working with an “older” version , but does not run with the current stable or nightly release. The error occurs in combination with multiprocessing. When I retrieve the network state dict and write it to a variable, I get an RuntimeError (see Stacktrace below).

Most interestingly the code works with PyTorch ‘1.0.0a0+95ca667’ but fails when using ‘1.0.0.dev20190123’. I couldn’t spot any changes within the release notes that could generate this behavior. Is anybody aware of changes between the alpha and the dev version, which could cause this issue?

Thanks!

Stacktrace:
File “/home/xxx/xxxxxx.py”, line 249, in _run_train
_ = output_queue.get_nowait()
File “/home/xxxxx/miniconda3/lib/python3.7/multiprocessing/queues.py”, line 126, in get_nowait
return self.get(False)
File “/home/xxx/miniconda3/lib/python3.7/multiprocessing/queues.py”, line 113, in get
return _ForkingPickler.loads(res)
File “/home/xxxxx/miniconda3/lib/python3.7/site-packages/torch/multiprocessing/reductions.py”, line 102, in rebuild_cuda_tensor
storage_offset_bytes)
RuntimeError: cuda runtime error (10) : invalid device ordinal at /opt/conda/conda-bld/pytorch-nightly_1548239643329/work/aten/src/THC/THCCachingAllocator.cpp:637