support group offloading under auto offloading (#14358)
Let group offloaded models take part in auto offloading
`ComponentsManager.enable_auto_cpu_offload` and `enable_group_offload` are two
independent hook systems — accelerate's `_hf_hook` and diffusers' `HookRegistry`
— and neither noticed the other. Enabling both raised nothing and appeared to
work, but auto offloading frees memory by calling `.to()`, which a group
offloaded module refuses and only warns about. Every offload the manager thought
it performed was a no-op: it recorded memory as freed that never was, and it
charged a group offloaded model's whole weight against the device although only
one group is ever resident.
A group offloaded model now takes part but places itself. It still makes room by
moving other models aside, since its `pre_forward` consults the strategy as
before; it is never chosen as the thing to move, because moving it does nothing;
and the manager no longer pretends to offload it. Either order works, group
offload before or after enabling.
Deciding *what* to move then has to come from somewhere other than memory
estimates, so `enable_auto_cpu_offload` takes an `offload_strategy` and
`set_offload_strategy` can replace it later. The default `AutoOffloadStrategy`
sizes its decisions from model memory footprints, which do not describe a model
holding one group at a time, so it warns when it meets group offloading — only
when no strategy was passed.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
