[Feat] amdsmi bindings integration

[parthraut]

• Added amdsmi to project dependency
• Implemented a non-blocking contructor of AMDGPU using concurrent.futures.ThreadPoolExecutor (AMDGPUs constructor takes 0.5s (polling time) regardless of number of GPUs)
• Right now, it succeeds if the measured value is in 1% of expected value, and waits 0.5s. These can be changed.

A few questions:

1. What do you think about the threshold (1%) and wait time (0.5s)?
2. This is failing pyright check:
   zeus/zeus/device/gpu/amd.py:247:13 - error: Operator "*" not supported for types "c_uint32" and "Literal[1000]" (reportOperatorIssue) zeus/zeus/device/gpu/amd.py:258:9 - error: Method "supportsGetTotalEnergyConsumption" overrides class "GPU" in an incompatible manner Positional parameter count mismatch; base method has 1, but override has 2 (reportIncompatibleMethodOverride) zeus/zeus/device/gpu/amd.py:280:26 - error: Cannot assign to attribute "_supportsGetTotalEnergyConsumption" for class "AMDGPU*" Attribute "_supportsGetTotalEnergyConsumption" is unknown (reportAttributeAccessIssue) zeus/zeus/device/gpu/amd.py:282:26 - error: Cannot assign to attribute "_supportsGetTotalEnergyConsumption" for class "AMDGPU*" Attribute "_supportsGetTotalEnergyConsumption" is unknown (reportAttributeAccessIssue) zeus/zeus/device/gpu/amd.py:293:26 - error: Cannot assign to attribute "_supportsGetTotalEnergyConsumption" for class "AMDGPU*" Attribute "_supportsGetTotalEnergyConsumption" is unknown (reportAttributeAccessIssue) zeus/zeus/device/gpu/amd.py:344:19 - error: Cannot access attribute "value" for class "AmdSmiException" Attribute "value" is unknown (reportAttributeAccessIssue) zeus/zeus/device/gpu/amd.py:346:37 - error: Cannot access attribute "msg" for class "AmdSmiException"
   I can fix these, but do you think this is the right approach? I wanted to make sure before tweaking base class signatures.

[jaywonchung]

Looked at a high level. Thanks for your work!

I think 0.5 seconds is just right, assuming it works.

[jaywonchung]

Thanks for the work! Just a couple nits :)

[jaywonchung]

Do we want to first check whether self._supportsGetTotalEnergyConsumption is not None and return the cached value for future invocations of this method?

[parthraut]

with the current structure it's never called more than once. But that would be better, I'll add it

[jaywonchung]

This method is part of the public API of GPU and GPUs, meaning people can always call it in their code. Thanks!

[jaywonchung]

@parthraut Is this ready for review? Did stuff work properly on the AMD HPC cluster?

[jaywonchung]

Thanks @parthraut! I've made a few specific suggestions, so please take a look and apply as you see fit. After these, I think it LGTM.

[jaywonchung]

I'm confused. There's an int() cast inside getInstantPowerUsage. Wouldn't this raise an error? In any case, this API design isn't clean. A function with a return value of int should always return an int regardless of the circumstances. Otherwise this defeats our purpose of static type checking.

[jaywonchung]

Suggested change

	# set _supportsInstantPowerUsage for all GPUs
	for gpu in self._gpus:
	if gpu.getInstantPowerUsage() == "N/A":
	gpu._supportsInstantPowerUsage = False
	# set _supportsInstantPowerUsage for all GPUs
	for gpu in self._gpus:
	gpu._supportsInstantPowerUsage = isinstance(
	amdsmi.amdsmi_get_power_info(gpu.handle)["current_socket_power"],
	int,
	)

[jaywonchung]

I checked the output of amdsmi.amdsmi_get_power_info, and the value of current/average socket power is already int. So, the int() casting before returning is redundant. Please remove it.

[jaywonchung]

What is the status of this PR?

[jaywonchung]

LGTM! Thanks a lot!