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

Fix NNCF not applying for TE only quant

Browse Source
This commit is contained in:
Disty0
2025-04-15 02:16:18 +03:00
parent 3564b34e0d
commit 7a4848dada
2 changed files with 10 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sd_models.py
+1 -1
View File
@@ -600,7 +600,7 @@ def load_diffuser(checkpoint_info=None, already_loaded_state_dict=None, timer=No
prompt_parser_diffusers.cache.clear()
set_diffuser_options(sd_model, vae, op, offload=False)
if 'Model' in shared.opts.nncf_compress_weights and not ('Model' in shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx"):
if shared.opts.nncf_compress_weights and not (shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx"):
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 = model_quant.optimum_quanto_weights(sd_model) # run this before move model so it can be compressed in CPU
modules/sd_models_utils.py
+9
View File
@@ -197,6 +197,15 @@ def apply_function_to_model(sd_model, function, options, op=None):
dtype=torch.float32 if devices.dtype != torch.bfloat16 else torch.bfloat16
)
sd_model.text_encoder_3 = function(sd_model.text_encoder_3, op="text_encoder_3", sd_model=sd_model)
if hasattr(sd_model, 'text_encoder_4') and hasattr(sd_model.text_encoder_4, 'config'):
if op == "nncf" and sd_model.text_encoder_4.__class__.__name__ in {"T5EncoderModel", "UMT5EncoderModel"}:
from modules.sd_hijack import NNCF_T5DenseGatedActDense # T5DenseGatedActDense uses fp32
for i in range(len(sd_model.text_encoder_4.encoder.block)):
sd_model.text_encoder_4.encoder.block[i].layer[1].DenseReluDense = NNCF_T5DenseGatedActDense(
sd_model.text_encoder_4.encoder.block[i].layer[1].DenseReluDense,
dtype=torch.float32 if devices.dtype != torch.bfloat16 else torch.bfloat16
)
sd_model.text_encoder_4 = function(sd_model.text_encoder_4, op="text_encoder_4", sd_model=sd_model)
if hasattr(sd_model, 'prior_pipe') and hasattr(sd_model.prior_pipe, 'text_encoder') and hasattr(sd_model.prior_pipe.text_encoder, 'config'):
sd_model.prior_pipe.text_encoder = function(sd_model.prior_pipe.text_encoder, op="prior_pipe.text_encoder", sd_model=sd_model)
if "VAE" in options: