add sd35-ipadapter and more balanced offload optimizations

Signed-off-by: Vladimir Mandic <mandic00@live.com>
2026-09-20 09:38:23 +02:00 · 2024-12-20 10:22:42 -05:00
parent 6e0ef69103
commit 20f2554cec
7 changed files with 148 additions and 97 deletions
@@ -4,7 +4,7 @@

 ### Highlights for 2024-12-19

-### SD.Next X-mass edition: *What's new?*
+### SD.Next Xmass edition: *What's new?*

 While we have several new supported models, workflows and tools, this release is primarily about *quality-of-life improvements*:  
 - New memory management engine: list of changes that went into this one is long: changes to GPU offloading, LoRA loader, system memory management, etc.  
@@ -15,18 +15,18 @@ While we have several new supported models, workflows and tools, this release is
 - New settings panel with simplified and streamlined configuration  

 We've also added support for several new models (see [supported models](https://vladmandic.github.io/sdnext-docs/Model-Support/) for full list) such as highly anticipated [NVLabs Sana](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px)  
-And several new video models: [Lightricks LTX-Video](https://huggingface.co/Lightricks/LTX-Video), [Hunyuan Video](https://huggingface.co/tencent/HunyuanVideo) and [Genmo Mochi.1 Preview](https://huggingface.co/genmo/mochi-1-preview)
+And several new SOTA video models: [Lightricks LTX-Video](https://huggingface.co/Lightricks/LTX-Video), [Hunyuan Video](https://huggingface.co/tencent/HunyuanVideo) and [Genmo Mochi.1 Preview](https://huggingface.co/genmo/mochi-1-preview)  

 And a lot of Control and IPAdapter goodies  
- for SDXL there is new [ProMax](https://huggingface.co/xinsir/controlnet-union-sdxl-1.0), improved *Union* and *Tiling*  
- for FLUX.1 there are [Flux Tools](https://blackforestlabs.ai/flux-1-tools/) as well as official *Canny* and *Depth* models and a cool [Redux](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev) model  
- for SD 3.5 there are official *Canny*, *Blur* and *Depth* in addition to existing 3rd party models  
+- for **SDXL** there is new [ProMax](https://huggingface.co/xinsir/controlnet-union-sdxl-1.0), improved *Union* and *Tiling* models 
+- for **FLUX.1** there are [Flux Tools](https://blackforestlabs.ai/flux-1-tools/) as well as official *Canny* and *Depth* models and a cool [Redux](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev) model  
+- for **SD3.5** there are official *Canny*, *Blur* and *Depth* models in addition to existing 3rd party models as well as [InstantX](https://huggingface.co/InstantX/SD3.5-Large-IP-Adapter) IP-adapter  

 Plus couple of new integrated workflows such as [FreeScale](https://github.com/ali-vilab/FreeScale) and [Style Aligned Image Generation](https://style-aligned-gen.github.io/)  

-And it wouldn't be a X-mass edition custom themes: *Snowflake* and *Elf-Green*  
+And it wouldn't be a Xmass edition without couple of custom themes: *Snowflake* and *Elf-Green*!  

-[README](https://github.com/vladmandic/automatic/blob/master/README.md) | [CHANGELOG](https://github.com/vladmandic/automatic/blob/master/CHANGELOG.md) | [Docs](https://vladmandic.github.io/sdnext-docs/) | [WiKi](https://github.com/vladmandic/automatic/wiki) | [Discord](https://discord.com/invite/sd-next-federal-batch-inspectors-1101998836328697867)
+[ReadMe](https://github.com/vladmandic/automatic/blob/master/README.md) | [ChangeLog](https://github.com/vladmandic/automatic/blob/master/CHANGELOG.md) | [Docs](https://vladmandic.github.io/sdnext-docs/) | [WiKi](https://github.com/vladmandic/automatic/wiki) | [Discord](https://discord.com/invite/sd-next-federal-batch-inspectors-1101998836328697867)

 ## Details for 2024-12-19

@@ -136,6 +136,7 @@ And it wouldn't be a X-mass edition custom themes: *Snowflake* and *Elf-Green*
  - balanced offload: add both high and low watermark, defaults as below  
    `0.25` for low-watermark: skip offload if memory usage is below 25%  
    `0.70` high-watermark: must offload if memory usage is above 70%  
+  - balanced offload will attempt to run offload as non-blocking and force gc at the end  
  - change-in-behavior:  
    low-end systems, triggered by either `lowvrwam` or by detection of <=4GB will use *sequential offload*  
    all other systems use *balanced offload* by default (can be changed in settings)  
@@ -198,6 +199,7 @@ And it wouldn't be a X-mass edition custom themes: *Snowflake* and *Elf-Green*
 - taesd limit render to 2024px  
 - taesd downscale preview to 1024px max: configurable in settings -> live preview  
 - uninstall conflicting `wandb` package  
+- dont skip diffusers version check if quick is specified  

 ## Update for 2024-11-21

@@ -457,9 +457,9 @@ def check_python(supported_minors=[9, 10, 11, 12], reason=None):

 # check diffusers version
 def check_diffusers():
-    if args.skip_all or args.skip_requirements:
+    if args.skip_all or args.skip_git:
        return
-    sha = '862a7d5038c1c53641ffcab146a7eeb5ab683656' # diffusers commit hash
+    sha = 'b64ca6c11cbc1644c22f1dae441c8124d588bb14' # diffusers commit hash
    pkg = pkg_resources.working_set.by_key.get('diffusers', None)
    minor = int(pkg.version.split('.')[1] if pkg is not None else 0)
    cur = opts.get('diffusers_version', '') if minor > 0 else ''
@@ -186,7 +186,7 @@ def get_device_for(task): # pylint: disable=unused-argument
    return get_optimal_device()


-def torch_gc(force=False, fast=False):
+def torch_gc(force:bool=False, fast:bool=False, reason:str=None):
    def get_stats():
        mem_dict = memstats.memory_stats()
        gpu_dict = mem_dict.get('gpu', {})
@@ -207,15 +207,21 @@ def torch_gc(force=False, fast=False):
    from modules.shared import cmd_opts

    t0 = time.time()
-    gpu, used_gpu, ram, used_ram, oom = get_stats()
+    gpu, used_gpu, ram, _used_ram, oom = get_stats()
    threshold = 0 if (cmd_opts.lowvram and not cmd_opts.use_zluda) else opts.torch_gc_threshold
    collected = 0
-    if force or threshold == 0 or used_gpu >= threshold or used_ram >= threshold:
+    if reason is None and force:
+        reason='force'
+    if threshold == 0 or used_gpu >= threshold:
        force = True
+        if reason is None:
+            reason = 'threshold'
    if oom > previous_oom:
        previous_oom = oom
        log.warning(f'Torch GPU out-of-memory error: {memstats.memory_stats()}')
        force = True
+        if reason is None:
+            reason = 'oom'
    if force:
        # actual gc
        collected = gc.collect() if not fast else 0 # python gc
@@ -237,10 +243,10 @@ def torch_gc(force=False, fast=False):
    new_gpu, new_used_gpu, new_ram, new_used_ram, oom = get_stats()
    before = { 'gpu': gpu, 'ram': ram }
    after = { 'gpu': new_gpu, 'ram': new_ram, 'oom': oom }
-    utilization = { 'gpu': new_used_gpu, 'ram': new_used_ram, 'threshold': threshold }
-    results = { 'saved': round(gpu - new_gpu, 2), 'collected': collected }
+    utilization = { 'gpu': new_used_gpu, 'ram': new_used_ram }
+    results = { 'gpu': round(gpu - new_gpu, 2), 'py': collected }
    fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access
-    log.debug(f'GC: utilization={utilization} gc={results} before={before} after={after} device={torch.device(get_optimal_device_name())} fn={fn} time={round(t1 - t0, 2)}')
+    log.debug(f'GC: current={after} prev={before} load={utilization} gc={results} fn={fn} why={reason} time={t1-t0:.2f}')
    return new_gpu, new_ram


@@ -9,12 +9,15 @@ import os
 import time
 import json
 from PIL import Image
+import diffusers
+import transformers
 from modules import processing, shared, devices, sd_models


-clip_repo = "h94/IP-Adapter"
 clip_loaded = None
 adapters_loaded = []
+CLIP_ID = "h94/IP-Adapter"
+SIGLIP_ID = 'google/siglip-so400m-patch14-384'
 ADAPTERS_NONE = {
    'None': { 'name': 'none', 'repo': 'none', 'subfolder': 'none' },
 }
@@ -37,11 +40,13 @@ ADAPTERS_SDXL = {
    'Ostris Composition ViT-H SDXL': { 'name': 'ip_plus_composition_sdxl.safetensors', 'repo': 'ostris/ip-composition-adapter', 'subfolder': '' },
 }
 ADAPTERS_SD3 = {
-    'InstantX Large': { 'name': 'ip-adapter.bin', 'repo': 'InstantX/SD3.5-Large-IP-Adapter' },
+    'None': { 'name': 'none', 'repo': 'none', 'subfolder': 'none' },
+    'InstantX Large': { 'name': 'none', 'repo': 'InstantX/SD3.5-Large-IP-Adapter', 'subfolder': 'none', 'revision': 'refs/pr/10' },
 }
 ADAPTERS_F1 = {
-    'XLabs AI v1': { 'name': 'ip_adapter.safetensors', 'repo': 'XLabs-AI/flux-ip-adapter' },
-    'XLabs AI v2': { 'name': 'ip_adapter.safetensors', 'repo': 'XLabs-AI/flux-ip-adapter-v2' },
+    'None': { 'name': 'none', 'repo': 'none', 'subfolder': 'none' },
+    'XLabs AI v1': { 'name': 'ip_adapter.safetensors', 'repo': 'XLabs-AI/flux-ip-adapter', 'subfolder': 'none' },
+    'XLabs AI v2': { 'name': 'ip_adapter.safetensors', 'repo': 'XLabs-AI/flux-ip-adapter-v2', 'subfolder': 'none' },
 }
 ADAPTERS = { **ADAPTERS_SD15, **ADAPTERS_SDXL, **ADAPTERS_SD3, **ADAPTERS_F1 }
 ADAPTERS_ALL = { **ADAPTERS_SD15, **ADAPTERS_SDXL, **ADAPTERS_SD3, **ADAPTERS_F1 }
@@ -126,6 +131,8 @@ def crop_images(images, crops):
                    shared.log.error(f'IP adapter: failed to crop image: source={len(images[i])} faces={len(cropped)}')
    except Exception as e:
        shared.log.error(f'IP adapter: failed to crop image: {e}')
+    if shared.sd_model_type == 'sd3' and len(images) == 1:
+        return images[0]
    return images


@@ -144,27 +151,64 @@ def unapply(pipe): # pylint: disable=arguments-differ
        pass


-def apply(pipe, p: processing.StableDiffusionProcessing, adapter_names=[], adapter_scales=[1.0], adapter_crops=[False], adapter_starts=[0.0], adapter_ends=[1.0], adapter_images=[]):
-    global clip_loaded, adapters_loaded # pylint: disable=global-statement
-    # overrides
-    if hasattr(p, 'ip_adapter_names'):
-        if isinstance(p.ip_adapter_names, str):
-            p.ip_adapter_names = [p.ip_adapter_names]
-        adapters = [ADAPTERS_ALL.get(adapter_name, None) for adapter_name in p.ip_adapter_names if adapter_name is not None and adapter_name.lower() != 'none']
-        adapter_names = p.ip_adapter_names
-    else:
-        if isinstance(adapter_names, str):
-            adapter_names = [adapter_names]
-        adapters = [ADAPTERS.get(adapter, None) for adapter in adapter_names]
-    adapters = [adapter for adapter in adapters if adapter is not None and adapter['name'].lower() != 'none']
-    if len(adapters) == 0:
-        unapply(pipe)
-        if hasattr(p, 'ip_adapter_images'):
-            del p.ip_adapter_images
-        return False
-    if shared.sd_model_type not in ['sd', 'sdxl', 'sd3', 'f1']:
-        shared.log.error(f'IP adapter: model={shared.sd_model_type} class={pipe.__class__.__name__} not supported')
-        return False
+def load_image_encoder(pipe: diffusers.DiffusionPipeline, adapter_names: list[str]):
+    global clip_loaded # pylint: disable=global-statement
+    for adapter_name in adapter_names:
+        # which clip to use
+        clip_repo = CLIP_ID
+        if 'ViT' not in adapter_name: # defaults per model
+            clip_subfolder = 'models/image_encoder' if shared.sd_model_type == 'sd' else 'sdxl_models/image_encoder'
+        if 'ViT-H' in adapter_name:
+            clip_subfolder = 'models/image_encoder' # this is vit-h
+        elif 'ViT-G' in adapter_name:
+            clip_subfolder = 'sdxl_models/image_encoder' # this is vit-g
+        else:
+            if shared.sd_model_type == 'sd':
+                clip_subfolder = 'models/image_encoder'
+            elif shared.sd_model_type == 'sdxl':
+                clip_subfolder = 'sdxl_models/image_encoder'
+            elif shared.sd_model_type == 'sd3':
+                clip_repo = SIGLIP_ID
+                clip_subfolder = None
+            elif shared.sd_model_type == 'f1':
+                shared.log.error(f'IP adapter: adapter={adapter_name} type={shared.sd_model_type} cls={shared.sd_model.__class__.__name__}: unsupported base model')
+                return False
+            else:
+                shared.log.error(f'IP adapter: unknown model type: {adapter_name}')
+                return False
+
+    # load image encoder used by ip adapter
+    if pipe.image_encoder is None or clip_loaded != f'{clip_repo}/{clip_subfolder}':
+        try:
+            if shared.sd_model_type == 'sd3':
+                pipe.image_encoder = transformers.SiglipVisionModel.from_pretrained(clip_repo, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir)
+            else:
+                pipe.image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, subfolder=clip_subfolder, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True)
+            shared.log.debug(f'IP adapter load: encoder="{clip_repo}/{clip_subfolder}" cls={pipe.image_encoder.__class__.__name__}')
+            clip_loaded = f'{clip_repo}/{clip_subfolder}'
+        except Exception as e:
+            shared.log.error(f'IP adapter load: encoder="{clip_repo}/{clip_subfolder}" {e}')
+            return False
+    sd_models.move_model(pipe.image_encoder, devices.device)
+    return True
+
+
+def load_feature_extractor(pipe):
+    # load feature extractor used by ip adapter
+    if pipe.feature_extractor is None:
+        try:
+            if shared.sd_model_type == 'sd3':
+                pipe.feature_extractor = transformers.SiglipImageProcessor.from_pretrained(SIGLIP_ID, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir)
+            else:
+                pipe.feature_extractor = transformers.CLIPImageProcessor()
+            shared.log.debug(f'IP adapter load: extractor={pipe.feature_extractor.__class__.__name__}')
+        except Exception as e:
+            shared.log.error(f'IP adapter load: extractor {e}')
+            return False
+    return True
+
+
+def parse_params(p: processing.StableDiffusionProcessing, adapters: list, adapter_scales: list[float], adapter_crops: list[bool], adapter_starts: list[float], adapter_ends: list[float], adapter_images: list):
    if hasattr(p, 'ip_adapter_scales'):
        adapter_scales = p.ip_adapter_scales
    if hasattr(p, 'ip_adapter_crops'):
@@ -205,6 +249,33 @@ def apply(pipe, p: processing.StableDiffusionProcessing, adapter_names=[], adapt
    p.ip_adapter_starts = adapter_starts.copy()
    adapter_ends = get_scales(adapter_ends, adapter_images)
    p.ip_adapter_ends = adapter_ends.copy()
+    return adapter_images, adapter_masks, adapter_scales, adapter_crops, adapter_starts, adapter_ends
+
+
+def apply(pipe, p: processing.StableDiffusionProcessing, adapter_names=[], adapter_scales=[1.0], adapter_crops=[False], adapter_starts=[0.0], adapter_ends=[1.0], adapter_images=[]):
+    global adapters_loaded # pylint: disable=global-statement
+    # overrides
+    if hasattr(p, 'ip_adapter_names'):
+        if isinstance(p.ip_adapter_names, str):
+            p.ip_adapter_names = [p.ip_adapter_names]
+        adapters = [ADAPTERS_ALL.get(adapter_name, None) for adapter_name in p.ip_adapter_names if adapter_name is not None and adapter_name.lower() != 'none']
+        adapter_names = p.ip_adapter_names
+    else:
+        if isinstance(adapter_names, str):
+            adapter_names = [adapter_names]
+        adapters = [ADAPTERS.get(adapter_name, None) for adapter_name in adapter_names if adapter_name.lower() != 'none']
+
+    if len(adapters) == 0:
+        unapply(pipe)
+        if hasattr(p, 'ip_adapter_images'):
+            del p.ip_adapter_images
+        return False
+    if shared.sd_model_type not in ['sd', 'sdxl', 'sd3', 'f1']:
+        shared.log.error(f'IP adapter: model={shared.sd_model_type} class={pipe.__class__.__name__} not supported')
+        return False
+
+    adapter_images, adapter_masks, adapter_scales, adapter_crops, adapter_starts, adapter_ends = parse_params(p, adapters, adapter_scales, adapter_crops, adapter_starts, adapter_ends, adapter_images)
+
    # init code
    if pipe is None:
        return False
@@ -223,61 +294,29 @@ def apply(pipe, p: processing.StableDiffusionProcessing, adapter_names=[], adapt
        shared.log.error(f'IP adapter: pipeline not supported: {pipe.__class__.__name__}')
        return False

-    for adapter_name in adapter_names:
-        # which clip to use
-        if 'ViT' not in adapter_name: # defaults per model
-            if shared.sd_model_type == 'sd':
-                clip_subfolder = 'models/image_encoder'
-            else:
-                clip_subfolder = 'sdxl_models/image_encoder'
-        if 'ViT-H' in adapter_name:
-            clip_subfolder = 'models/image_encoder' # this is vit-h
-        elif 'ViT-G' in adapter_name:
-            clip_subfolder = 'sdxl_models/image_encoder' # this is vit-g
-        else:
-            if shared.sd_model_type == 'sd':
-                clip_subfolder = 'models/image_encoder'
-            elif shared.sd_model_type == 'sdxl':
-                clip_subfolder = 'sdxl_models/image_encoder'
-            elif shared.sd_model_type == 'sd3':
-                shared.log.error(f'IP adapter: adapter={adapter_name} type={shared.sd_model_type} cls={shared.sd_model.__class__.__name__}: unsupported base model')
-                return False
-            elif shared.sd_model_type == 'f1':
-                shared.log.error(f'IP adapter: adapter={adapter_name} type={shared.sd_model_type} cls={shared.sd_model.__class__.__name__}: unsupported base model')
-                return False
-            else:
-                shared.log.error(f'IP adapter: unknown model type: {adapter_name}')
-                return False
+    if not load_image_encoder(pipe, adapter_names):
+        return False

-    # load feature extractor used by ip adapter
-    if pipe.feature_extractor is None:
-        try:
-            from transformers import CLIPImageProcessor
-            shared.log.debug('IP adapter load: feature extractor')
-            pipe.feature_extractor = CLIPImageProcessor()
-        except Exception as e:
-            shared.log.error(f'IP adapter load: feature extractor {e}')
-            return False
-
-    # load image encoder used by ip adapter
-    if pipe.image_encoder is None or clip_loaded != f'{clip_repo}/{clip_subfolder}':
-        try:
-            from transformers import CLIPVisionModelWithProjection
-            shared.log.debug(f'IP adapter load: image encoder="{clip_repo}/{clip_subfolder}"')
-            pipe.image_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_repo, subfolder=clip_subfolder, torch_dtype=devices.dtype, cache_dir=shared.opts.diffusers_dir, use_safetensors=True)
-            clip_loaded = f'{clip_repo}/{clip_subfolder}'
-        except Exception as e:
-            shared.log.error(f'IP adapter load: image encoder="{clip_repo}/{clip_subfolder}" {e}')
-            return False
-    sd_models.move_model(pipe.image_encoder, devices.device)
+    if not load_feature_extractor(pipe):
+        return False

    # main code
    try:
        t0 = time.time()
-        repos = [adapter['repo'] for adapter in adapters]
-        subfolders = [adapter['subfolder'] for adapter in adapters]
-        names = [adapter['name'] for adapter in adapters]
-        pipe.load_ip_adapter(repos, subfolder=subfolders, weight_name=names)
+        repos = [adapter.get('repo', None) for adapter in adapters if adapter.get('repo', 'none') != 'none']
+        subfolders = [adapter.get('subfolder', None) for adapter in adapters if adapter.get('subfolder', 'none') != 'none']
+        names = [adapter.get('name', None) for adapter in adapters if adapter.get('name', 'none') != 'none']
+        revisions = [adapter.get('revision', None) for adapter in adapters if adapter.get('revision', 'none') != 'none']
+        kwargs = {}
+        if len(repos) == 1:
+            repos = repos[0]
+        if len(subfolders) > 0:
+            kwargs['subfolder'] = subfolders if len(subfolders) > 1 else subfolders[0]
+        if len(names) > 0:
+            kwargs['weight_name'] = names if len(names) > 1 else names[0]
+        if len(revisions) > 0:
+            kwargs['revision'] = revisions[0]
+        pipe.load_ip_adapter(repos, **kwargs)
        adapters_loaded = names
        if hasattr(p, 'ip_adapter_layers'):
            pipe.set_ip_adapter_scale(p.ip_adapter_layers)
@@ -286,8 +325,8 @@ def apply(pipe, p: processing.StableDiffusionProcessing, adapter_names=[], adapt
            for i in range(len(adapter_scales)):
                if adapter_starts[i] > 0:
                    adapter_scales[i] = 0.00
-            pipe.set_ip_adapter_scale(adapter_scales)
-            ip_str =  [f'{os.path.splitext(adapter)[0]}:{scale}:{start}:{end}' for adapter, scale, start, end in zip(adapter_names, adapter_scales, adapter_starts, adapter_ends)]
+            pipe.set_ip_adapter_scale(adapter_scales if len(adapter_scales) > 1 else adapter_scales[0])
+            ip_str =  [f'{os.path.splitext(adapter)[0]}:{scale}:{start}:{end}:{crop}' for adapter, scale, start, end, crop in zip(adapter_names, adapter_scales, adapter_starts, adapter_ends, adapter_crops)]
        p.task_args['ip_adapter_image'] = crop_images(adapter_images, adapter_crops)
        if len(adapter_masks) > 0:
            p.cross_attention_kwargs = { 'ip_adapter_masks': adapter_masks }
@@ -480,5 +480,5 @@ 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)
+    devices.torch_gc(force=True, reason='final')
    return processed
@@ -496,7 +496,11 @@ def apply_balanced_offload(sd_model, exclude=[]):
                    used_gpu -= module_size
                debug_move(f'Offload: type=balanced op={"move" if do_offload else "skip"} gpu={prev_gpu:.3f}:{used_gpu:.3f} perc={perc_gpu:.2f} ram={used_ram:.3f} current={module.device} dtype={module.dtype} component={module.__class__.__name__} size={module_size:.3f}')
            except Exception as e:
-                if 'bitsandbytes' not in str(e):
+                if 'out of memory' in str(e):
+                    devices.torch_gc(fast=True, force=True, reason='oom')
+                elif 'bitsandbytes' in str(e):
+                    pass
+                else:
                    shared.log.error(f'Offload: type=balanced op=apply module={module_name} {e}')
                if os.environ.get('SD_MOVE_DEBUG', None):
                    errors.display(e, f'Offload: type=balanced op=apply module={module_name}')
@@ -508,7 +512,7 @@ def apply_balanced_offload(sd_model, exclude=[]):
            if device_map and max_memory:
                module.balanced_offload_device_map = device_map
                module.balanced_offload_max_memory = max_memory
-        devices.torch_gc(fast=True, force=True)
+        devices.torch_gc(fast=True, force=True, reason='offload')

    apply_balanced_offload_to_module(sd_model)
    if hasattr(sd_model, "pipe"):
@@ -518,7 +522,6 @@ def apply_balanced_offload(sd_model, exclude=[]):
    if hasattr(sd_model, "decoder_pipe"):
        apply_balanced_offload_to_module(sd_model.decoder_pipe)
    set_accelerate(sd_model)
-    devices.torch_gc(fast=True)
    t = time.time() - t0
    process_timer.add('offload', t)
    fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access
@@ -7,7 +7,8 @@ from modules import scripts, processing, shared, devices, sd_models
 from installer import install


-redux_pipe: diffusers.FluxPriorReduxPipeline = None
+# redux_pipe: diffusers.FluxPriorReduxPipeline = None
+redux_pipe = None
 processor_canny = None
 processor_depth = None
 title = 'Flux Tools'