Not able to see GPU memory consumption as part of system metrics in aim stack.

[dushyantbehl]

❓Question

I have been using aimstack version 3.17.5 and unable to see any GPU memory consumption when doing aimstack runs.

The dashboard shows GPU %, GPU temprature but not the memory used. Is there a way to track what is going on?

I am happy to share any information about the environment you may have. Thanks in advance.

[dushyantbehl]

Hi @alberttorosyan Thanks for marking this as a bug.
Could I be of any help here to fix things or dig deeper. Please let me know, I'll be happy to see if I can help.

[alberttorosyan]

@dushyantbehl, here's the code snippet which extracts the GPU information before passing it to Aim tracking methods:

                gpu_info = dict()
                handle = nvml.nvmlDeviceGetHandleByIndex(i)
                try:
                    util = nvml.nvmlDeviceGetUtilizationRates(handle)
                    # GPU utilization percent
                    gpu_info["gpu"] = round10e5(util.gpu)
                except nvml.NVMLError_NotSupported:
                    pass
                try:
                    # Get device memory
                    memory = nvml.nvmlDeviceGetMemoryInfo(handle)
                    # Device memory usage
                    # 'memory_used': round10e5(memory.used / 1024 / 1024),
                    gpu_info["gpu_memory_percent"] = round10e5(memory.used * 100 / memory.total)
                except nvml.NVMLError_NotSupported:
                    pass
                try:
                    # Get device temperature
                    nvml_tmp = nvml.NVML_TEMPERATURE_GPU
                    temp = nvml.nvmlDeviceGetTemperature(handle, nvml_tmp)
                    # Device temperature
                    gpu_info["gpu_temp"] = round10e5(temp)
                except nvml.NVMLError_NotSupported:
                    pass
                try:
                    # Compute power usage in watts and percent
                    power_watts = nvml.nvmlDeviceGetPowerUsage(handle) / 1000
                    power_cap = nvml.nvmlDeviceGetEnforcedPowerLimit(handle)
                    power_cap_watts = power_cap / 1000
                    power_watts / power_cap_watts * 100
                    # Power usage in watts and percent
                    gpu_info["gpu_power_watts"]: round10e5(power_watts)
                    # gpu_info["power_percent"] = round10e5(power_usage)
                except nvml.NVMLError_NotSupported:
                    pass

Each call to nmvl API is wrapped with try/except block. If you see the power consumption, temperature, etc. that means that the specific call has failed, due to device support.

[ChanderG]

@alberttorosyan It was not a device support problem since directly using nmvl APIs worked. After some debugging, I found the cause. Have opened a PR here: #3044

[dushyantbehl]

Fix merged here - #3044