The Hogwild example doesn't use my GPU when executing

#1

I implemented the Hogwild example program from here. My PC has a GPU, but when I execute the program it does not use the GPU but all the CPU cores, why? How can I indicate I want it to be run with the GPU instead?

#2

You can use the GPU by passing the --cuda argument as seen here.

#3

Thanks, but I get an error: RuntimeError: Cannot pickle CUDA storage; try pickling a CUDA tensor instead, do you know how I can fix it?

#4

Could you try to move these two lines right after the args = parser.parse_args() call?
I needed to move them as I got an error as:

RuntimeError: context has already been set

I’ll create an issue with a fix if I can reproduce it using the nightly (my current PyTorch build is a source build from a few weeks ago).

#5

I did the moving of those lines and after running the program I got the context error, were you able to fix it?

#6

I got the before changing anything. After I moved the lines of code it worked.
Are you only seeing the error after moving?

#7

Hi, the context error displays when moving the lines as you told me, the pickle error occurs when running the posted code as is.