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

Move quant functions to model_quant.py

Browse Source
This commit is contained in:
Disty0
2025-01-23 21:50:26 +03:00
parent 5a7c1f50c1
commit 9b579bfd96
5 changed files with 318 additions and 292 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sd_models.py
+5 -5
View File
@@ -16,7 +16,7 @@ from modules.modeldata import model_data
from modules.sd_checkpoint import CheckpointInfo, select_checkpoint, list_models, checkpoints_list, checkpoint_titles, get_closet_checkpoint_match, model_hash, update_model_hashes, setup_model, write_metadata, read_metadata_from_safetensors # pylint: disable=unused-import
from modules.sd_offload import disable_offload, set_diffuser_offload, apply_balanced_offload, set_accelerate # pylint: disable=unused-import
from modules.sd_models_legacy import get_checkpoint_state_dict, load_model_weights, load_model, repair_config # pylint: disable=unused-import
from modules.sd_models_utils import NoWatermark, get_signature, get_call, path_to_repo, patch_diffuser_config, convert_to_faketensors, read_state_dict, get_state_dict_from_checkpoint # pylint: disable=unused-import
from modules.sd_models_utils import NoWatermark, get_signature, get_call, path_to_repo, patch_diffuser_config, convert_to_faketensors, read_state_dict, get_state_dict_from_checkpoint, apply_function_to_model # pylint: disable=unused-import
model_dir = "Stable-diffusion"
@@ -130,9 +130,9 @@ def set_diffuser_options(sd_model, vae=None, op:str='model', offload:bool=True,
model.requires_grad_(False)
model.eval()
return model
sd_model = sd_models_compile.apply_compile_to_model(sd_model, eval_model, ["Model", "VAE", "Text Encoder"], op="eval")
sd_model = apply_function_to_model(sd_model, eval_model, ["Model", "VAE", "Text Encoder"], op="eval")
if len(shared.opts.torchao_quantization) > 0 and shared.opts.torchao_quantization_mode == 'post':
sd_model = sd_models_compile.torchao_quantization(sd_model)
sd_model = model_quant.torchao_quantization(sd_model)
if shared.opts.opt_channelslast and hasattr(sd_model, 'unet'):
shared.log.quiet(quiet, f'Setting {op}: channels-last=True')
@@ -567,9 +567,9 @@ def load_diffuser(checkpoint_info=None, already_loaded_state_dict=None, timer=No
set_diffuser_options(sd_model, vae, op, offload=False)
if shared.opts.nncf_compress_weights and not ('Model' in shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx"):
sd_model = sd_models_compile.nncf_compress_weights(sd_model) # run this before move model so it can be compressed in CPU
sd_model = model_quant.nncf_compress_weights(sd_model) # run this before move model so it can be compressed in CPU
if shared.opts.optimum_quanto_weights:
sd_model = sd_models_compile.optimum_quanto_weights(sd_model) # run this before move model so it can be compressed in CPU
sd_model = model_quant.optimum_quanto_weights(sd_model) # run this before move model so it can be compressed in CPU
if shared.opts.layerwise_quantization:
model_quant.apply_layerwise(sd_model)
timer.record("options")