wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-12-15 01:07:03 +08:00
Code Issues Actions 12 Packages Projects Releases Wiki Activity

Fix regression when text encoder loaded directly on GPU. (#11129)

Browse Source
This commit is contained in:
comfyanonymous 2025-12-05 12:33:16 -08:00 committed by GitHub
parent 43071e3de3
commit 6fd463aec9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 26 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
comfy/ops.py
View File

@ -552,6 +552,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
weight_scale_key = f"{prefix}weight_scale"
scale = state_dict.pop(weight_scale_key, None)
if scale is not None:
scale = scale.to(device)
layout_params = {
'scale': scale,
'orig_dtype': MixedPrecisionOps._compute_dtype,

44
comfy/sd.py
View File

@ -98,7 +98,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
class CLIP:
def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}):
def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}):
if no_init:
return
params = target.params.copy()
@ -129,6 +129,27 @@ class CLIP:
self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
self.patcher.is_clip = True
self.apply_hooks_to_conds = None
if len(state_dict) > 0:
if isinstance(state_dict, list):
for c in state_dict:
m, u = self.load_sd(c)
if len(m) > 0:
logging.warning("clip missing: {}".format(m))
if len(u) > 0:
logging.debug("clip unexpected: {}".format(u))
else:
m, u = self.load_sd(state_dict, full_model=True)
if len(m) > 0:
m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))
if len(m_filter) > 0:
logging.warning("clip missing: {}".format(m))
else:
logging.debug("clip missing: {}".format(m))
if len(u) > 0:
logging.debug("clip unexpected {}:".format(u))
if params['device'] == load_device:
model_management.load_models_gpu([self.patcher], force_full_load=True)
self.layer_idx = None
@ -1225,14 +1246,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
parameters += comfy.utils.calculate_parameters(c)
tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options)
clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, model_options=model_options)
for c in clip_data:
m, u = clip.load_sd(c)
if len(m) > 0:
logging.warning("clip missing: {}".format(m))
if len(u) > 0:
logging.debug("clip unexpected: {}".format(u))
clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, state_dict=clip_data, model_options=model_options)
return clip
def load_gligen(ckpt_path):
@ -1362,17 +1376,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
clip_sd = model_config.process_clip_state_dict(sd)
if len(clip_sd) > 0:
parameters = comfy.utils.calculate_parameters(clip_sd)
clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options)
m, u = clip.load_sd(clip_sd, full_model=True)
if len(m) > 0:
m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))
if len(m_filter) > 0:
logging.warning("clip missing: {}".format(m))
else:
logging.debug("clip missing: {}".format(m))
if len(u) > 0:
logging.debug("clip unexpected {}:".format(u))
clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, state_dict=clip_sd, model_options=te_model_options)
else:
logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.")