Does Apex really support PyTorch 0.4.1?

In Readme.md it says:

PyTorch 0.4 or newer.

But sample code examples/simple/distributed goes down with error AttributeError: 'Linear' object has no attribute 'named_buffers', because named_buffers method of torch.Module was introduced after 0.4.1 version.

Is there any build, that supports pytorch 0.4.1? If not, at least maybe the readme.md should be changed?

P.S.: I’m using apex with CUDA and C++ libraries, installed with command pip install -v --no-cache-dir .

[WrathOfGrapes]

@ptrblck The code runs fine without amp.

I’m hoping to find some time during this month to localize and summarize the issues for you.

[ptrblck]

@WrathOfGrapes is the IllegalMemoryAccess error only thrown if you are using amp or always using your code in PyTorch 1.1 on a GPU? In the latter case, you could try to run your code using

CUDA_LAUNCH_BLOCKING=1 python script.py args

and check, e.g. if no values are out of bounds.

In the former case, we would need to reproduce this issue in order to debug it, so let us know, if you find some time for debugging and we’ll have a look at it.