But even with this change, it still needs to use delayed creation
because BMC side kernel doesn't know how many DIMMs are populated on
a remote server before the remote server completes its memory
training and testing in BIOS, but it needs to check the remote
server's CPU temperature as immediate as possible to make
appropriate thermal control based on the remote CPU's temperature to
avoid any critical thermal issue. What would be a better solution in
this case?

You could change this driver so that it supports one DIMM.  Move the
'hotplug' part into another driver which creates and destroys
instances of the hwmon DIMM device as the DIMMS come and go.

Also, do you need to handle CPU hotplug? You could split the CPU
temperature part into a separate hwmon driver? And again create and
destroy devices as CPUs come and go?

	Andrew