Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ fix T5 Loading

Browse Source
This commit is contained in:
Disty0
2025-10-25 20:03:12 +03:00
parent eb3fcac55a
commit 01c4ce18b6
1 changed files with 10 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
pipelines/generic.py
+10 -12
View File
@@ -114,34 +114,32 @@ def load_text_encoder(repo_id, cls_name, load_config={}, subfolder="text_encoder
text_encoder = model_quant.do_post_load_quant(text_encoder, allow=quant_type is not None)
# load from local file safetensors
elif local_file is not None and local_file.lower().endswith('.safetensors'):
shared.log.debug(f'Load model: text_encoder="{local_file}" cls={cls_name.__name__} quant="{quant_type}"')
shared.log.debug(f'Load model: text_encoder="{local_file}" cls={cls_name.__name__} quant="{quant_type}" args={load_args}')
from modules import model_te
text_encoder = model_te.load_t5(local_file)
text_encoder = model_quant.do_post_load_quant(text_encoder, allow=quant_type is not None)
# use shared t5 if possible
elif cls_name == transformers.T5EncoderModel and allow_shared:
with open(os.path.join('configs', 'flux', 'text_encoder_2', 'config.json'), encoding='utf8') as f:
load_args['config'] = transformers.T5Config(**json.load(f))
if model_quant.check_nunchaku('TE'):
import nunchaku
repo_id = 'nunchaku-tech/nunchaku-t5/awq-int4-flux.1-t5xxl.safetensors'
cls_name = nunchaku.NunchakuT5EncoderModel
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="SVDQuant"')
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="SVDQuant" args={load_args}')
text_encoder = nunchaku.NunchakuT5EncoderModel.from_pretrained(
repo_id,
torch_dtype=dtype,
)
text_encoder.quantization_method = 'SVDQuant'
elif shared.opts.te_shared_t5:
"""
if 'sdnq-' in repo_id.lower():
if 'sdnq-uint4-svd' in repo_id.lower():
repo_id = 'Disty0/FLUX.1-dev-SDNQ-uint4-svd-r32'
load_args['subfolder'] = 'text_encoder_2'
quant_args = {}
else:
"""
repo_id = 'Disty0/t5-xxl'
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" shared={shared.opts.te_shared_t5}')
repo_id = 'Disty0/t5-xxl'
if 'sdnq-' not in repo_id.lower():
with open(os.path.join('configs', 'flux', 'text_encoder_2', 'config.json'), encoding='utf8') as f:
load_args['config'] = transformers.T5Config(**json.load(f))
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" shared={shared.opts.te_shared_t5} args={load_args}')
if dtype is not None:
load_args['torch_dtype'] = dtype
text_encoder = cls_name.from_pretrained(
@@ -154,7 +152,7 @@ def load_text_encoder(repo_id, cls_name, load_config={}, subfolder="text_encoder
if shared.opts.te_shared_t5:
repo_id = 'hunyuanvideo-community/HunyuanImage-2.1-Diffusers'
subfolder = 'text_encoder'
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" shared={shared.opts.te_shared_t5}')
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" shared={shared.opts.te_shared_t5} args={load_args}')
if dtype is not None:
load_args['torch_dtype'] = dtype
text_encoder = cls_name.from_pretrained(
@@ -167,7 +165,7 @@ def load_text_encoder(repo_id, cls_name, load_config={}, subfolder="text_encoder
# load from repo
if text_encoder is None:
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" shared={shared.opts.te_shared_t5}')
shared.log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" shared={shared.opts.te_shared_t5} args={load_args}')
if dtype is not None:
load_args['torch_dtype'] = dtype
if subfolder is not None: