SAM3: moving the text encoder with "Select CLIP Device" fails with 'NoneType' object has no attribute 'keys' #16675

Expected Behavior

SelectCLIPDevice can move the CLIP of a SAM3 / SAM 3.1 checkpoint (CheckpointLoaderSimple) to a second GPU, like any other CLIP.

Actual Behavior

When the selected device differs from the load device, the node fails:

File "comfy_extras/nodes_multigpu.py", line 276, in execute
    clip.patcher = _apply_patcher_device(clip.patcher, resolved)
File "comfy/model_patcher.py", line 521, in deepclone_multigpu
    temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1])
File "comfy/sd.py", line 2179, in load_checkpoint_clip_patcher
    _, clip, _, _ = load_checkpoint_guess_config(...)
File "comfy/sd.py", line 2309, in load_state_dict_guess_config
    clip_sd = model_config.process_clip_state_dict(sd)
File "comfy/supported_models.py", line 2464, in process_clip_state_dict
    clip_keys = utils.state_dict_prefix_replace(clip_keys, {...}, filter_keys=True)
File "comfy/utils.py", line 245, in state_dict_prefix_replace
AttributeError: 'NoneType' object has no attribute 'keys'

Cause

SAM3.process_unet_state_dict moves the language_backbone keys into self._clip_stash, and SAM3.process_clip_state_dict reads them back with getattr(self, "_clip_stash", {}). The CLIP-only reload (load_checkpoint_clip_patcher, output_model=False) never calls process_unet_state_dict, so the attribute is missing. The default {} is never used, because supported_models_base.BASE.__getattr__ returns None for every unknown attribute. (With {} the reload would silently build an empty text encoder instead.)

Minimal reproduction (core only)

import comfy.sd, folder_paths
ckpt = folder_paths.get_full_path_or_raise("checkpoints", "sam3.1_multiplex_fp16.safetensors")
comfy.sd.load_checkpoint_clip_patcher(ckpt)   # AttributeError: 'NoneType' object has no attribute 'keys'

Suggested fix

Collect the stash from the checkpoint when it is missing, as process_unet_state_dict does:

def process_clip_state_dict(self, state_dict):
    if self.__dict__.get("_clip_stash") is None:  # CLIP-only reload
        self._clip_stash = {k: state_dict.pop(k) for k in list(state_dict.keys())
                            if "language_backbone" in k and "resizer" not in k}
    clip_keys = self._clip_stash
    ...

With this change the reload above returns the SAM3 text encoder (SAM3ClipModelWrapper, 354M parameters), and the normal load path is unchanged.

Environment

ComfyUI 0.37.0 (a716932; code unchanged on master 8cfe5e1), frontend 1.53.6, Windows 11, PyTorch 2.13 + ROCm, two AMD GPUs (Radeon AI PRO R9700 + RX 9070 XT). Custom nodes are not involved: the reproduction above uses core code only.