svdquant and others stuff

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -27,6 +27,7 @@
     comma-separate list of regex patterns to skip  
   - ui display reference models with subdued color  
   - xyz grid support bool  
+  - do not force gc at end of processing  
 - **Wiki**  
   - new Nunchaku page  
   - updated HiDream, Quantization, NNCF pages  
@@ -37,6 +38,7 @@
   - SD35 InstantX IP-adapter  
   - **HunyuanVideo-I2V** with latest transformers  
   - trace logging  
+  - xyz grid restore settings  
 
 ## Update for 2025-04-12
 

modules/model_flux.py

@@ -122,8 +122,9 @@ def load_quants(kwargs, repo_id, cache_dir, allow_quant):
             else:
                 shared.log.error(f'Load module: quant=Nunchaku module=transformer repo="{repo_id}" unsupported')
             if nunchaku_repo is not None:
-                shared.log.debug(f'Load module: quant=Nunchaku module=transformer repo="{nunchaku_repo}" precision={nunchaku_precision} attention={shared.opts.nunchaku_attention}')
-                kwargs['transformer'] = nunchaku.NunchakuFluxTransformer2dModel.from_pretrained(nunchaku_repo, torch_dtype=devices.dtype)
+                shared.log.debug(f'Load module: quant=Nunchaku module=transformer repo="{nunchaku_repo}" precision={nunchaku_precision} offload={shared.opts.nunchaku_offload} attention={shared.opts.nunchaku_attention}')
+                kwargs['transformer'] = nunchaku.NunchakuFluxTransformer2dModel.from_pretrained(nunchaku_repo, offload=shared.opts.nunchaku_offload, torch_dtype=devices.dtype)
+                kwargs['transformer'].quantization_method = 'SVDQuant'
                 if shared.opts.nunchaku_attention:
                     kwargs['transformer'].set_attention_impl("nunchaku-fp16")
         elif 'transformer' not in kwargs and model_quant.check_quant('Transformer'):

modules/processing.py

@@ -502,5 +502,5 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
     if not p.disable_extra_networks:
         shared.log.info(f'Processed: images={len(output_images)} its={(p.steps * len(output_images)) / (t1 - t0):.2f} time={t1-t0:.2f} timers={timer.process.dct()} memory={memstats.memory_stats()}')
 
-    devices.torch_gc(force=True, reason='final')
+    devices.torch_gc(force=False, reason='final')
     return processed

modules/processing_args.py

@@ -153,7 +153,7 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:t
             shared.log.error(f'Prompt parser encode: {e}')
             if os.environ.get('SD_PROMPT_DEBUG', None) is not None:
                 errors.display(e, 'Prompt parser encode')
-        timer.process.record('encode', reset=False)
+        timer.process.record('prompt', reset=False)
     else:
         prompt_parser_diffusers.embedder = None
 

modules/processing_vae.py

@@ -2,7 +2,6 @@ import os
 import time
 import numpy as np
 import torch
-import torchvision.transforms.functional as TF
 from modules import shared, devices, sd_models, sd_vae, sd_vae_taesd, errors
 
 
@@ -316,6 +315,7 @@ def vae_decode(latents, model, output_type='np', vae_type='Full', width=None, he
 
 
 def vae_encode(image, model, vae_type='Full'): # pylint: disable=unused-variable
+    import torchvision.transforms.functional as f
     if shared.state.interrupted or shared.state.skipped:
         return []
     if not hasattr(model, 'vae') and hasattr(model, 'pipe'):
@@ -323,7 +323,7 @@ def vae_encode(image, model, vae_type='Full'): # pylint: disable=unused-variable
     if not hasattr(model, 'vae'):
         shared.log.error('VAE not found in model')
         return []
-    tensor = TF.to_tensor(image.convert("RGB")).unsqueeze(0).to(devices.device, devices.dtype_vae)
+    tensor = f.to_tensor(image.convert("RGB")).unsqueeze(0).to(devices.device, devices.dtype_vae)
     if vae_type == 'Full':
         tensor = tensor * 2 - 1
         latents = full_vae_encode(image=tensor, model=shared.sd_model)

modules/shared.py

@@ -552,6 +552,7 @@ options_templates.update(options_section(('quantization', "Quantization Settings
     "nunchaku_sep": OptionInfo("<h2>Nunchaku Engine</h2>", "", gr.HTML),
     "nunchaku_quantization": OptionInfo([], "SVDQuant enabled", gr.CheckboxGroup, {"choices": ["Model", "Transformer", "VAE", "TE", "Video", "LLM", "ControlNet"], "visible": native}),
     "nunchaku_attention": OptionInfo(False, "Nunchaku attention", gr.Checkbox, {"visible": native}),
+    "nunchaku_offload": OptionInfo(False, "Nunchaku offloading", gr.Checkbox, {"visible": native}),
 }))
 
 options_templates.update(options_section(('advanced', "Pipeline Modifiers"), {

scripts/xyz_grid_classes.py

@@ -36,6 +36,7 @@ class SharedSettingsStackHelper(object):
     freeu_b2 = None
     freeu_s1 = None
     freeu_s2 = None
+    cfgzero_enabled = None
     schedulers_sigma_adjust = None
     schedulers_beta_schedule = None
     schedulers_beta_start = None
@@ -53,6 +54,7 @@ class SharedSettingsStackHelper(object):
     eta_noise_seed_delta = None
     tome_ratio = None
     todo_ratio = None
+    teacache_thresh = None
     extra_networks_default_multiplier = None
     disable_weights_auto_swap = None
 
@@ -75,6 +77,7 @@ class SharedSettingsStackHelper(object):
         self.freeu_b2 = shared.opts.freeu_b2
         self.freeu_s1 = shared.opts.freeu_s1
         self.freeu_s2 = shared.opts.freeu_s2
+        self.cfgzero_enabled = shared.opts.cfgzero_enabled
         self.sd_model_checkpoint = shared.opts.sd_model_checkpoint
         self.sd_model_refiner = shared.opts.sd_model_refiner
         self.sd_model_dict = shared.opts.sd_model_dict
@@ -83,6 +86,7 @@ class SharedSettingsStackHelper(object):
         self.sd_text_encoder = shared.opts.sd_text_encoder
         self.extra_networks_default_multiplier = shared.opts.extra_networks_default_multiplier
         self.disable_weights_auto_swap = shared.opts.disable_weights_auto_swap
+        self.teacache_thresh = shared.opts.teacache_thresh
         shared.opts.data["disable_weights_auto_swap"] = False
 
     def __exit__(self, exc_type, exc_value, tb):
@@ -100,12 +104,14 @@ class SharedSettingsStackHelper(object):
         shared.opts.data["schedulers_shift"] = self.schedulers_shift
         shared.opts.data["scheduler_eta"] = self.scheduler_eta
         shared.opts.data["eta_noise_seed_delta"] = self.eta_noise_seed_delta
+        shared.opts.data["cfgzero_enabled"] = self.cfgzero_enabled
         shared.opts.data["freeu_b1"] = self.freeu_b1
         shared.opts.data["freeu_b2"] = self.freeu_b2
         shared.opts.data["freeu_s1"] = self.freeu_s1
         shared.opts.data["freeu_s2"] = self.freeu_s2
         shared.opts.data["tome_ratio"] = self.tome_ratio
         shared.opts.data["todo_ratio"] = self.todo_ratio
+        shared.opts.data["teacache_thresh"] = self.teacache_thresh
 
         if self.sd_model_checkpoint != shared.opts.sd_model_checkpoint:
             shared.opts.data["sd_model_checkpoint"] = self.sd_model_checkpoint