Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 09:14:35 +02:00
Code Issues Packages Projects Releases Wiki Activity

NNCF compress Text Encoder and Lora support

Browse Source
This commit is contained in:
Disty0
2024-01-10 02:10:52 +03:00
parent 395e2fd23e
commit 3352be21a6
3 changed files with 68 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sd_models_compile.py
+44 -11
View File
@@ -34,12 +34,24 @@ def ipex_optimize(sd_model):
sd_model.unet = ipex.optimize(sd_model.unet, dtype=devices.dtype_unet, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
else:
shared.log.warning('IPEX Optimize enabled but model has no Unet')
if hasattr(sd_model, 'vae'):
sd_model.vae.training = False
sd_model.vae = ipex.optimize(sd_model.vae, dtype=devices.dtype_vae, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
if hasattr(sd_model, 'movq'):
sd_model.movq.training = False
sd_model.movq = ipex.optimize(sd_model.movq, dtype=devices.dtype_vae, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
if shared.opts.ipex_optimize_vae:
if hasattr(sd_model, 'vae'):
sd_model.vae.training = False
sd_model.vae = ipex.optimize(sd_model.vae, dtype=devices.dtype_vae, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
elif hasattr(sd_model, 'movq'):
sd_model.movq.training = False
sd_model.movq = ipex.optimize(sd_model.movq, dtype=devices.dtype_vae, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
else:
shared.log.warning('Compress VAE Weights enabled but model has no VAE')
if shared.opts.ipex_optimize_text_encoder:
if hasattr(sd_model, 'text_encoder'):
sd_model.text_encoder.training = False
sd_model.text_encoder = ipex.optimize(sd_model.text_encoder, dtype=devices.dtype_unet, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
if hasattr(sd_model, 'text_encoder_2'):
sd_model.text_encoder_2.training = False
sd_model.text_encoder_2 = ipex.optimize(sd_model.text_encoder_2, dtype=devices.dtype_unet, inplace=True, weights_prepack=False) # pylint: disable=attribute-defined-outside-init
else:
shared.log.warning('IPEX Optimize Text Encoder Weights enabled but model has no Text Encoder')
t1 = time.time()
shared.log.info(f"IPEX Optimize: time={t1-t0:.2f}")
return sd_model
@@ -50,15 +62,36 @@ def nncf_compress_weights(sd_model):
try:
t0 = time.time()
import nncf
if hasattr(sd_model, 'unet'):
sd_model.unet = nncf.compress_weights(sd_model.unet)
if shared.compiled_model_state is None:
shared.compiled_model_state = CompiledModelState()
else:
shared.log.warning('Compress Weights enabled but model has no Unet')
shared.compiled_model_state.compiled_cache.clear()
shared.compiled_model_state.partitioned_modules.clear()
backup_lora_model = []
if shared.compiled_model_state.lora_compile:
backup_lora_model = shared.compiled_model_state.lora_model
shared.compiled_model_state = CompiledModelState()
shared.compiled_model_state.lora_model = backup_lora_model
if shared.opts.nncf_compress_weights:
if hasattr(sd_model, 'unet'):
sd_model.unet = nncf.compress_weights(sd_model.unet)
else:
shared.log.warning('Compress Weights enabled but model has no Unet')
if shared.opts.nncf_compress_vae_weights:
if hasattr(sd_model, 'vae'):
sd_model.vae = nncf.compress_weights(sd_model.vae)
if hasattr(sd_model, 'movq'):
elif hasattr(sd_model, 'movq'):
sd_model.movq = nncf.compress_weights(sd_model.movq)
else:
shared.log.warning('Compress VAE Weights enabled but model has no VAE')
if shared.opts.nncf_compress_text_encoder_weights:
if hasattr(sd_model, 'text_encoder'):
sd_model.text_encoder = nncf.compress_weights(sd_model.text_encoder)
if hasattr(sd_model, 'text_encoder_2'):
sd_model.text_encoder_2 = nncf.compress_weights(sd_model.text_encoder_2)
else:
shared.log.warning('Compress VAE Text Encoder Weights enabled but model has no Text Encoder')
t1 = time.time()
shared.log.info(f"Compress Weights: time={t1-t0:.2f}")
return sd_model
@@ -181,7 +214,7 @@ def compile_torch(sd_model):
def compile_diffusers(sd_model):
if shared.opts.ipex_optimize:
sd_model = ipex_optimize(sd_model)
if shared.opts.nncf_compress_weights and not (shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx"):
if not (shared.opts.cuda_compile and shared.opts.cuda_compile_backend == "openvino_fx"):
sd_model = nncf_compress_weights(sd_model)
if not (shared.opts.cuda_compile or shared.opts.cuda_compile_vae or shared.opts.cuda_compile_upscaler):
return sd_model
modules/shared.py
+12 -3
View File
@@ -345,14 +345,23 @@ options_templates.update(options_section(('cuda', "Compute Settings"), {
"cuda_compile_verbose": OptionInfo(False, "Model compile verbose mode"),
"cuda_compile_errors": OptionInfo(True, "Model compile suppress errors"),
"diffusers_quantization": OptionInfo(False, "Enable dynamic quantization with torchao"),
"nncf_sep": OptionInfo("<h2>NNCF</h2>", "", gr.HTML),
"nncf_compress_weights": OptionInfo(False, "Compress Model weights with NNCF"),
"nncf_compress_vae_weights": OptionInfo(False, "Compress VAE weights with NNCF"),
"nncf_compress_text_encoder_weights": OptionInfo(False, "Compress Text Encoder weights with NNCF"),
"ipex_sep": OptionInfo("<h2>IPEX, DirectML and OpenVINO</h2>", "", gr.HTML),
"ipex_optimize": OptionInfo(False if not devices.backend == "ipex" else True, "Enable IPEX Optimize for Intel GPUs"),
"ipex_optimize_upscaler": OptionInfo(False if not devices.backend == "ipex" else True, "Enable IPEX Optimize for Intel GPUs with Upscalers"),
"directml_sep": OptionInfo("<h2>DirectML</h2>", "", gr.HTML),
"directml_memory_provider": OptionInfo(default_memory_provider, 'DirectML memory stats provider', gr.Radio, {"choices": memory_providers}),
"directml_catch_nan": OptionInfo(False, "DirectML retry specific operation when NaN is produced if possible. (makes generation slower)"),
"ipex_sep": OptionInfo("<h2>IPEX</h2>", "", gr.HTML),
"ipex_optimize": OptionInfo(False if not devices.backend == "ipex" else True, "Enable IPEX Optimize for Intel GPUs with UNet"),
"ipex_optimize_vae": OptionInfo(False if not devices.backend == "ipex" else True, "Enable IPEX Optimize for Intel GPUs with VAE"),
"ipex_optimize_text_encoder": OptionInfo(False if not devices.backend == "ipex" else True, "Enable IPEX Optimize for Intel GPUs with Text Encoder"),
"ipex_optimize_upscaler": OptionInfo(False if not devices.backend == "ipex" else True, "Enable IPEX Optimize for Intel GPUs with Upscalers"),
"openvino_sep": OptionInfo("<h2>OpenVINO</h2>", "", gr.HTML),
"openvino_disable_model_caching": OptionInfo(False, "OpenVINO disable model caching"),
"openvino_hetero_gpu": OptionInfo(False, "OpenVINO use Hetero Device for single inference with multiple devices"),
"openvino_remove_cpu_from_hetero": OptionInfo(False, "OpenVINO remove CPU from Hetero Device"),