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

jumbo update: add wan22-vace, improve offloading, add offloading-using-streams, change logging-levels, refactor some default packages

Browse Source 
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2025-09-17 12:32:27 -04:00
parent 6a021e7743
commit cc6101ecb2
17 changed files with 156 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/processing_args.py
+10
View File
@@ -128,6 +128,8 @@ def task_specific_kwargs(p, model):
}
if model.__class__.__name__ == 'WanImageToVideoPipeline' and hasattr(p, 'init_images') and len(p.init_images) > 0:
task_args['image'] = p.init_images[0]
if model.__class__.__name__ == 'WanVACEPipeline' and hasattr(p, 'init_images') and len(p.init_images) > 0:
task_args['reference_images'] = p.init_images
if debug_enabled:
debug_log(f'Process task specific args: {task_args}')
@@ -307,6 +309,14 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:t
args['control_strength'] = p.denoising_strength
args['width'] = p.width
args['height'] = p.height
if 'WanVACEPipeline' in model.__class__.__name__:
if isinstance(args['prompt'], list):
args['prompt'] = args['prompt'][0] if len(args['prompt']) > 0 else ''
if isinstance(args.get('negative_prompt', None), list):
args['negative_prompt'] = args['negative_prompt'][0] if len(args['negative_prompt']) > 0 else ''
if isinstance(args['generator'], list) and len(args['generator']) > 0:
args['generator'] = args['generator'][0]
# set callbacks
if 'prior_callback_steps' in possible: # Wuerstchen / Cascade
args['prior_callback_steps'] = 1
modules/processing_diffusers.py
+1 -1
View File
@@ -487,7 +487,7 @@ def validate_pipeline(p: processing.StableDiffusionProcessing):
if m.repo_cls is not None:
models_cls.append(m.repo_cls.__name__)
is_video_model = shared.sd_model.__class__.__name__ in models_cls
override_video_pipelines = ['WanPipeline', 'WanImageToVideoPipeline']
override_video_pipelines = ['WanPipeline', 'WanImageToVideoPipeline', 'WanVACEPipeline']
is_video_pipeline = ('video' in p.__class__.__name__.lower()) or (shared.sd_model.__class__.__name__ in override_video_pipelines)
if is_video_model and not is_video_pipeline:
shared.log.error(f'Mismatch: type={shared.sd_model_type} cls={shared.sd_model.__class__.__name__} request={p.__class__.__name__} video model with non-video pipeline')
modules/processing_helpers.py
+3 -3
View File
@@ -372,7 +372,7 @@ def calculate_base_steps(p, use_denoise_start, use_refiner_start):
if len(getattr(p, 'timesteps', [])) > 0:
return None
cls = shared.sd_model.__class__.__name__
if 'Flex' in cls or 'Kontext' in cls or 'Edit' in cls:
if 'Flex' in cls or 'Kontext' in cls or 'Edit' in cls or 'Wan' in cls:
steps = p.steps
elif not is_txt2img():
if cls in sd_models.i2i_pipes:
@@ -393,7 +393,7 @@ def calculate_base_steps(p, use_denoise_start, use_refiner_start):
def calculate_hires_steps(p):
cls = shared.sd_model.__class__.__name__
if 'Flex' in cls or 'HiDreamImageEditingPipeline' in cls or 'Kontext' in cls:
if 'Flex' in cls or 'Kontext' in cls or 'Edit' in cls or 'Wan' in cls:
steps = p.steps
elif p.hr_second_pass_steps > 0:
steps = (p.hr_second_pass_steps // p.denoising_strength) + 1
@@ -407,7 +407,7 @@ def calculate_hires_steps(p):
def calculate_refiner_steps(p):
cls = shared.sd_model.__class__.__name__
if 'Flex' in cls or 'HiDreamImageEditingPipeline' in cls or 'Kontext' in cls:
if 'Flex' in cls or 'Kontext' in cls or 'Edit' in cls or 'Wan' in cls:
steps = p.steps
elif "StableDiffusionXL" in shared.sd_refiner.__class__.__name__:
if p.refiner_start > 0 and p.refiner_start < 1:
modules/sd_offload.py
+29 -9
View File
@@ -18,6 +18,7 @@ offload_post = ['h1']
offload_hook_instance = None
balanced_offload_exclude = ['CogView4Pipeline', 'MeissonicPipeline']
accelerate_dtype_byte_size = None
move_stream = None
def dtype_byte_size(dtype: torch.dtype):
@@ -176,7 +177,7 @@ class OffloadHook(accelerate.hooks.ModelHook):
self.last_post = None
self.last_cls = None
gpu = f'{(shared.gpu_memory * shared.opts.diffusers_offload_min_gpu_memory):.2f}-{(shared.gpu_memory * shared.opts.diffusers_offload_max_gpu_memory):.2f}:{shared.gpu_memory:.2f}'
shared.log.info(f'Offload: type=balanced op=init watermark={self.min_watermark}-{self.max_watermark} gpu={gpu} cpu={shared.cpu_memory:.3f} limit={shared.opts.cuda_mem_fraction:.2f} always={self.offload_always} never={self.offload_never} pre={shared.opts.diffusers_offload_pre}')
shared.log.info(f'Offload: type=balanced op=init watermark={self.min_watermark}-{self.max_watermark} gpu={gpu} cpu={shared.cpu_memory:.3f} limit={shared.opts.cuda_mem_fraction:.2f} always={self.offload_always} never={self.offload_never} pre={shared.opts.diffusers_offload_pre} streams={shared.opts.diffusers_offload_streams}')
self.validate()
super().__init__()
@@ -210,8 +211,12 @@ class OffloadHook(accelerate.hooks.ModelHook):
def pre_forward(self, module, *args, **kwargs):
_id = id(module)
if (self.last_pre != _id) and (module.__class__.__name__ != self.last_cls) and self.offload_allowed(module): # offload every other module first time when new module starts pre-forward
do_offload = (self.last_pre != _id) or (module.__class__.__name__ != self.last_cls)
if do_offload and self.offload_allowed(module): # offload every other module first time when new module starts pre-forward
if shared.opts.diffusers_offload_pre:
t0 = time.time()
debug_move(f'Offload: type=balanced op=pre module={module.__class__.__name__}')
for pipe in get_pipe_variants():
for module_name in get_module_names(pipe):
@@ -220,15 +225,16 @@ class OffloadHook(accelerate.hooks.ModelHook):
if (_id != id(module_instance)) and (module_cls not in self.offload_never) and (not devices.same_device(module_instance.device, devices.cpu)):
apply_balanced_offload_to_module(module_instance, op='pre')
self.last_cls = module.__class__.__name__
self.last_pre = _id
process_timer.add('offload', time.time() - t0)
if not devices.same_device(module.device, devices.device): # move-to-device
t0 = time.time()
device_index = torch.device(devices.device).index
if device_index is None:
device_index = 0
max_memory = { device_index: self.gpu, "cpu": self.cpu }
device_map = getattr(module, "balanced_offload_device_map", None)
if device_map is None or max_memory != getattr(module, "balanced_offload_max_memory", None):
if (device_map is None) or (max_memory != getattr(module, "balanced_offload_max_memory", None)):
device_map = accelerate.infer_auto_device_map(module, max_memory=max_memory)
offload_dir = getattr(module, "offload_dir", os.path.join(shared.opts.accelerate_offload_path, module.__class__.__name__))
if devices.backend == "directml":
@@ -241,13 +247,15 @@ class OffloadHook(accelerate.hooks.ModelHook):
module._hf_hook.execution_device = torch.device(devices.device) # pylint: disable=protected-access
module.balanced_offload_device_map = device_map
module.balanced_offload_max_memory = max_memory
process_timer.add('onload', time.time() - t0)
if debug:
for pipe in get_pipe_variants():
for _i, pipe in enumerate(get_pipe_variants()):
for module_name in get_module_names(pipe):
module_instance = getattr(pipe, module_name, None)
shared.log.trace(f'Offload: type=balanced op=pre:status module={module_instance.__class__.__name__} device={module_instance.device} dtype={module_instance.dtype}')
shared.log.trace(f'Offload: type=balanced op=pre:status forward={module.__class__.__name__} module={module_name} class={module_instance.__class__.__name__} pipe={_i} device={module_instance.device} dtype={module_instance.dtype}')
self.last_pre = _id
return args, kwargs
def post_forward(self, module, output):
@@ -283,6 +291,7 @@ def get_module_names(pipe=None, exclude=[]):
modules_names = get_signature(pipe).keys()
modules_names = [m for m in modules_names if m not in exclude and not m.startswith('_')]
modules_names = [m for m in modules_names if isinstance(getattr(pipe, m, None), torch.nn.Module)]
modules_names = list(sorted(set(modules_names)))
return modules_names
@@ -308,6 +317,17 @@ def get_module_sizes(pipe=None, exclude=[]):
def move_module_to_cpu(module, op='unk', force:bool=False):
def do_move(module):
if shared.opts.diffusers_offload_streams:
global move_stream # pylint: disable=global-statement
if move_stream is None:
move_stream = torch.cuda.Stream(device=devices.device)
with torch.cuda.stream(move_stream):
module = module.to(devices.cpu)
else:
module = module.to(devices.cpu)
return module
try:
module_name = getattr(module, "module_name", module.__class__.__name__)
module_size = offload_hook_instance.offload_map.get(module_name, offload_hook_instance.model_size())
@@ -318,17 +338,17 @@ def move_module_to_cpu(module, op='unk', force:bool=False):
op = f'{op}:skip'
if force:
op = f'{op}:force'
module = module.to(devices.cpu)
module = do_move(module)
used_gpu -= module_size
elif module_cls in offload_hook_instance.offload_never:
op = f'{op}:never'
elif module_cls in offload_hook_instance.offload_always:
op = f'{op}:always'
module = module.to(devices.cpu)
module = do_move(module)
used_gpu -= module_size
elif perc_gpu > shared.opts.diffusers_offload_min_gpu_memory:
op = f'{op}:mem'
module = module.to(devices.cpu)
module = do_move(module)
used_gpu -= module_size
if debug:
quant = getattr(module, "quantization_method", None)
modules/shared.py
+1
View File
@@ -170,6 +170,7 @@ options_templates.update(options_section(('offload', "Model Offloading"), {
"diffusers_offload_nonblocking": OptionInfo(False, "Non-blocking move operations"),
"offload_balanced_sep": OptionInfo("<h2>Balanced Offload</h2>", "", gr.HTML),
"diffusers_offload_pre": OptionInfo(True, "Offload during pre-forward"),
"diffusers_offload_streams": OptionInfo(False, "Offload using streams"),
"diffusers_offload_min_gpu_memory": OptionInfo(startup_offload_min_gpu, "Offload low watermark", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01 }),
"diffusers_offload_max_gpu_memory": OptionInfo(startup_offload_max_gpu, "Offload GPU high watermark", gr.Slider, {"minimum": 0.1, "maximum": 1, "step": 0.01 }),
"diffusers_offload_max_cpu_memory": OptionInfo(0.90, "Offload CPU high watermark", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01, "visible": False }),
modules/ui_extra_networks_checkpoints.py
+2
View File
@@ -14,6 +14,8 @@ version_map = {
"SDXL Hyper": "SD XL",
"StableDiffusion3": "SD 3",
"StableDiffusionXL": "SD XL",
"WanToVideo": "Wan",
"WanVACE": "Wan",
}
class ExtraNetworksPageCheckpoints(ui_extra_networks.ExtraNetworksPage):
modules/video_models/models_def.py
+8 -2
View File
@@ -169,6 +169,12 @@ models = {
te_cls=transformers.T5EncoderModel,
dit_cls=diffusers.WanTransformer3DModel,
dit_folder=("transformer", "transformer_2")),
Model(name='WAN 2.2 14B VACE',
url='https://huggingface.co/Wan-AI/Wan2.2-14B-VACE-T2V-Diffusers',
repo='linoyts/Wan2.2-VACE-Fun-14B-diffusers',
repo_cls=diffusers.WanVACEPipeline,
te_cls=transformers.T5EncoderModel,
dit_cls=diffusers.WanVACETransformer3DModel),
Model(name='WAN 2.1 1.3B T2V',
url='https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers',
repo='Wan-AI/Wan2.1-T2V-1.3B-Diffusers',
@@ -204,13 +210,13 @@ models = {
repo='Wan-AI/Wan2.1-VACE-1.3B-diffusers',
repo_cls=diffusers.WanVACEPipeline,
te_cls=transformers.T5EncoderModel,
dit_cls=diffusers.WanTransformer3DModel),
dit_cls=diffusers.WanVACETransformer3DModel),
Model(name='WAN 2.1 VACE 14B',
url='https://huggingface.co/Wan-AI/Wan2.1-VACE-14B-diffusers',
repo='Wan-AI/Wan2.1-VACE-14B-diffusers',
repo_cls=diffusers.WanVACEPipeline,
te_cls=transformers.T5EncoderModel,
dit_cls=diffusers.WanTransformer3DModel),
dit_cls=diffusers.WanVACETransformer3DModel),
],
'SkyReels V2': [
Model(name='None'),
modules/video_models/video_overrides.py
+9 -6
View File
@@ -32,19 +32,22 @@ def set_overrides(p: processing.StableDiffusionProcessingVideo, selected: Model)
if selected.name == 'Latte 1 T2V':
p.task_args['enable_temporal_attentions'] = True
p.task_args['video_length'] = 16 * (max(p.frames // 16, 1))
# SkyReels
if 'SkyReelsV2DiffusionForcing' in cls:
p.task_args['overlap_history'] = 17
# LTX
if cls == 'LTXImageToVideoPipeline' or cls == 'LTXConditionPipeline':
p.task_args['generator'] = None
if cls == 'LTXConditionPipeline':
p.task_args['strength'] = p.denoising_strength
if 'LTX' in cls:
p.task_args['width'] = 32 * (p.width // 32)
p.task_args['height'] = 32 * (p.height // 32)
# WAN
if 'Wan' in cls:
p.task_args['width'] = 16 * (p.width // 16)
p.task_args['height'] = 16 * (p.height // 16)
p.frames = 4 * (max(p.frames // 4, 1)) + 1
# LTX
if 'LTX' in cls:
p.task_args['width'] = 32 * (p.width // 32)
p.task_args['height'] = 32 * (p.height // 32)
if 'SkyReelsV2DiffusionForcing' in cls:
p.task_args['overlap_history'] = 17
# WAN VACE
if 'WanVACEPipeline' in cls:
p.task_args['reference_images'] = 1
modules/video_models/video_run.py
+7 -3
View File
@@ -53,7 +53,10 @@ def generate(*args, **kwargs):
p.do_not_save_grid = True
p.do_not_save_samples = not save_frames
p.outpath_samples = shared.opts.outdir_samples or shared.opts.outdir_video
if 'I2V' in model:
if 'T2V' in model:
if init_image is not None:
shared.log.warning('Video: op=T2V init image not supported')
elif 'I2V' in model:
if init_image is None:
return video_utils.queue_err('init image not set')
p.task_args['image'] = images.resize_image(resize_mode=2, im=init_image, width=p.width, height=p.height, upscaler_name=None, output_type='pil')
@@ -66,9 +69,10 @@ def generate(*args, **kwargs):
p.task_args['image'] = images.resize_image(resize_mode=2, im=init_image, width=p.width, height=p.height, upscaler_name=None, output_type='pil')
p.task_args['last_image'] = images.resize_image(resize_mode=2, im=last_image, width=p.width, height=p.height, upscaler_name=None, output_type='pil')
shared.log.debug(f'Video: op=FLF2V init={init_image} last={last_image} resized={p.task_args["image"]}')
elif 'T2V' in model:
elif 'VACE' in model:
if init_image is not None:
shared.log.warning('Video: op=T2V init image not supported')
p.task_args['reference_images'] = [images.resize_image(resize_mode=2, im=init_image, width=p.width, height=p.height, upscaler_name=None, output_type='pil')]
shared.log.debug(f'Video: op=VACE reference={init_image} resized={p.task_args["reference_images"]}')
else:
shared.log.warning(f'Video: unknown model type "{model}"')