Issue with multiple GPUs

[shermanlo77]

Thanks for the code

We were experiencing issues running this code with multiple GPUs with the -p argument, where the stream does not match the device and the program deadlocks if it does match

I suggest the following:

• The stream should be initialised after the device
• pynvml.nvml.Init() should only be called once, rather than at every thread
• Avoid race conditions for stop_flag

I've made those fixes on my fork. Let me know what you think and I'll be happy to make changes or a PR