[WIP] Fix redundant nvml init by cached handles

[Stonesjtu]

Fix for #54

[Stonesjtu]

The error catching workflow is complicated for me, can we just throw the raw error messages instead of the current one.

[wookayin]

For example, if an error happens during the loop, we can simply throw it out.

[Stonesjtu]

The error catching requires all the NVML API calls to be wrapped with a long try-catch block.

[wookayin]

Good job and thanks. I would like a better style of handling loops (maybe involving a bit of refactoring on main function; please see the comments on code). BTW does it indeed bring improvement on the performance?

[wookayin]

We should not remove this static method this is already being used (as a public API, sort of) in other applications. Since you are lengthening the lifecycle of GPUStatCollection and making it update-able, we can think of supporting both constructor and factory method.

[Stonesjtu]

I thought the wrapper function gpustat.new_query() was the real API for users, instead of the implementation itself.

It's ok to retain the legacy API, but I prefer a factory function like

@classmethod
def new_query(cls):
    return cls()

[wookayin]

It makes less sense to take gpu_stat as an argument for this function. I was thinking of a style like (i) we first create a GPUStatCollection instance here and (ii) then print outputs using it. As the same instance should be reused, we can consider running a wait loop inside this function rather than calling it multiple times.

[Stonesjtu]

What about making this function a generator, so we can keep the GPUStatCollection object in the closure.

[wookayin]

For example, if an error happens during the loop, we can simply throw it out.

[wookayin]

Superceded by #166