Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 01:04:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

reimplement torchao quantization

Browse Source 
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2024-10-18 09:34:01 -04:00
parent 1d51ae3baa
commit ae4591ac0b
6 changed files with 89 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sd_models.py
+3 -3
View File
@@ -737,8 +737,8 @@ def set_diffuser_options(sd_model, vae = None, op: str = 'model', offload=True):
model.eval()
return model
sd_model = sd_models_compile.apply_compile_to_model(sd_model, eval_model, ["Model", "VAE", "Text Encoder"], op="eval")
if shared.opts.diffusers_quantization:
sd_model = sd_models_compile.dynamic_quantization(sd_model)
if len(shared.opts.torchao_quantization) > 0:
sd_model = sd_models_compile.torchao_quantization(sd_model)
if shared.opts.opt_channelslast and hasattr(sd_model, 'unet'):
shared.log.debug(f'Setting {op}: channels-last=True')
@@ -1193,7 +1193,7 @@ def load_diffuser_file(model_type, pipeline, checkpoint_info, diffusers_load_con
from diffusers.utils import import_utils
import_utils._accelerate_available = False # pylint: disable=protected-access
if shared.opts.diffusers_to_gpu and model_type.startswith('Stable Diffusion'):
shared.log.debug(f'Diffusers accelerate: hijack={shared.opts.diffusers_to_gpu}')
shared.log.debug(f'Diffusers accelerate: direct={shared.opts.diffusers_to_gpu}')
sd_hijack_accelerate.hijack_accelerate()
else:
sd_hijack_accelerate.restore_accelerate()