cuda.bindings.nvml.init_v2#

cuda.bindings.nvml.init_v2()#

Initialize the NVML Library lazily, without allocating any device state.

This will initialize the NVML Library state without enumerating any discovered devices. This will allow NVML to communicate with a device, even if other devices are in an unstable or bad state. Enumeration of a device can be done by obtaining the device handle via the nvmlDeviceGetHandleBy* class of APIs. This method needs to be called once before any usage of NVML Library APIs. For all products.

Note

A reference count of the number of initializations is maintained, and a corresponding call to shutdown() needs to be issued once usage of the NVML Library is complete. Shutdown will only occur after the reference count reaches zero.