mirror of https://github.com/vladmandic/automatic
39 lines
1.4 KiB
Python
39 lines
1.4 KiB
Python
import transformers
|
|
import diffusers
|
|
from modules import shared, devices, sd_models, model_quant, sd_hijack_te
|
|
from pipelines import generic
|
|
|
|
|
|
def load_qwen(checkpoint_info, diffusers_load_config={}):
|
|
repo_id = sd_models.path_to_repo(checkpoint_info)
|
|
sd_models.hf_auth_check(checkpoint_info)
|
|
|
|
load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config, module='Model')
|
|
shared.log.debug(f'Load model: type=Qwen model="{checkpoint_info.name}" repo="{repo_id}" offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
|
|
|
|
transformer = generic.load_transformer(repo_id, cls_name=diffusers.QwenImageTransformer2DModel, load_config=diffusers_load_config)
|
|
repo_te = 'Qwen/Qwen-Image' if 'Qwen-Lightning' in repo_id else repo_id
|
|
text_encoder = generic.load_text_encoder(repo_te, cls_name=transformers.Qwen2_5_VLForConditionalGeneration, load_config=diffusers_load_config)
|
|
|
|
pipe = diffusers.QwenImagePipeline.from_pretrained(
|
|
repo_id,
|
|
transformer=transformer,
|
|
text_encoder=text_encoder,
|
|
cache_dir=shared.opts.diffusers_dir,
|
|
**load_args,
|
|
)
|
|
pipe.task_args = {
|
|
'output_type': 'np',
|
|
}
|
|
|
|
del text_encoder
|
|
del transformer
|
|
|
|
sd_hijack_te.init_hijack(pipe)
|
|
from modules.video_models import video_vae
|
|
pipe.vae.orig_decode = pipe.vae.decode
|
|
pipe.vae.decode = video_vae.hijack_vae_decode
|
|
|
|
devices.torch_gc()
|
|
return pipe
|