mirror of
https://github.com/vladmandic/automatic
synced 2026-09-18 16:54:33 +02:00
fix ipadapters and improve offloading
Co-authored-by: Copilot <copilot@github.com> Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
@@ -66,6 +66,7 @@ For full details, see [ChangeLog](https://github.com/vladmandic/automatic/blob/m
|
||||
- remove "processed preview" from ui
|
||||
preprocessor output can still be generated by clicking preview button in in control unit and it will render into normal output area
|
||||
- **Internal**
|
||||
- `offload` auto-reapply hook on error
|
||||
- refactor `pip` installer, thanks @awsr
|
||||
- remove obsolete `lora` stepwise and functional code, thanks @awsr
|
||||
- interrupt model loading between components
|
||||
@@ -90,6 +91,8 @@ For full details, see [ChangeLog](https://github.com/vladmandic/automatic/blob/m
|
||||
- `detailer` handle `lora` internally
|
||||
- vae preview flashes previous image
|
||||
- `torch.compile` improvements
|
||||
- `gradio` preprocess exception handling
|
||||
- `ipadapters` with offloading
|
||||
|
||||
## Update for 2026-04-28
|
||||
|
||||
|
||||
@@ -1,16 +1,14 @@
|
||||
# TODO
|
||||
|
||||
- Outpaint
|
||||
- IPAdapter
|
||||
- https://github.com/PrunaAI/pruna
|
||||
|
||||
## Features
|
||||
|
||||
### Assigned
|
||||
|
||||
- Check Outpaint, @vladmandic
|
||||
- Chat-based interface, @vladmandic
|
||||
- Control tab verify overrides handling, @vladmandic
|
||||
- Reimplement `llama` remover for Kanvas, @vladmandic
|
||||
- Implement [pruna](https://github.com/PrunaAI/pruna), @vladmandic
|
||||
|
||||
- Cloud providers, @CalamitousFelicitousness
|
||||
- Video processing add full API support, @CalamitousFelicitousness
|
||||
|
||||
@@ -153,6 +153,7 @@ def patch_gradio():
|
||||
orig_cancel_tasks = gradio.utils.cancel_tasks
|
||||
orig_restore_session_state = gradio.route_utils.restore_session_state
|
||||
orig_call_prediction = gradio.queueing.Queue.call_prediction
|
||||
orig_blocks_preprocess_data = gradio.blocks.Blocks.preprocess_data
|
||||
|
||||
async def wrap_cancel_tasks(task_ids: set[str]):
|
||||
log.error(f'Gradio cancel: task={task_ids}')
|
||||
@@ -200,9 +201,17 @@ def patch_gradio():
|
||||
log.error(f"Gradio queue: events={len(events)} batch={batch} error: {e}")
|
||||
raise
|
||||
|
||||
def wrap_blocks_preprocess_data(self, fn_index: int, inputs: list, state: dict):
|
||||
try:
|
||||
return orig_blocks_preprocess_data(self, fn_index, inputs, state)
|
||||
except Exception as e:
|
||||
log.error(f"Gradio preprocess: {e}")
|
||||
raise
|
||||
|
||||
gradio.queueing.Queue.call_prediction = wrap_call_prediction
|
||||
gradio.route_utils.restore_session_state = wrap_restore_session_state
|
||||
gradio.utils.cancel_tasks = wrap_cancel_tasks
|
||||
gradio.blocks.Blocks.preprocess_data = wrap_blocks_preprocess_data
|
||||
|
||||
|
||||
def patch_gradio_future():
|
||||
|
||||
@@ -201,10 +201,10 @@ def load_image_encoder(pipe: DiffusionPipeline, adapter_names: list[str]):
|
||||
else:
|
||||
if clip_subfolder is None:
|
||||
image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True, **offline_config)
|
||||
log.debug(f'IP adapter load: encoder="{clip_repo}" cls={pipe.image_encoder.__class__.__name__}')
|
||||
log.debug(f'IP adapter load: encoder="{clip_repo}" cls={image_encoder.__class__.__name__}')
|
||||
else:
|
||||
image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, subfolder=clip_subfolder, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True, **offline_config)
|
||||
log.debug(f'IP adapter load: encoder="{clip_repo}/{clip_subfolder}" cls={pipe.image_encoder.__class__.__name__}')
|
||||
log.debug(f'IP adapter load: encoder="{clip_repo}/{clip_subfolder}" cls={image_encoder.__class__.__name__}')
|
||||
sd_models.clear_caches()
|
||||
image_encoder = model_quant.do_post_load_quant(image_encoder, allow=True)
|
||||
if hasattr(pipe, 'register_modules'):
|
||||
@@ -212,12 +212,16 @@ def load_image_encoder(pipe: DiffusionPipeline, adapter_names: list[str]):
|
||||
else:
|
||||
pipe.image_encoder = image_encoder
|
||||
clip_loaded = f'{clip_repo}/{clip_subfolder}'
|
||||
pipe = sd_models.apply_balanced_offload(pipe, force=True)
|
||||
except Exception as e:
|
||||
log.error(f'IP adapter load: encoder="{clip_repo}/{clip_subfolder}" {e}')
|
||||
errors.display(e, 'IP adapter: type=encoder')
|
||||
return False
|
||||
shared.state.end(jobid)
|
||||
sd_models.move_model(pipe.image_encoder, devices.device)
|
||||
if hasattr(pipe.unet, 'balanced_offload_device_map') and pipe.unet.balanced_offload_device_map.get('encoder_hid_proj', None) is None:
|
||||
# image encoder patches unet, but diffusers creates full device map only during model load so if module is loaded later it will be missing
|
||||
pipe.unet.balanced_offload_device_map['encoder_hid_proj'] = 0
|
||||
return True
|
||||
|
||||
|
||||
@@ -235,13 +239,14 @@ def load_feature_extractor(pipe):
|
||||
pipe.register_modules(feature_extractor=feature_extractor)
|
||||
else:
|
||||
pipe.feature_extractor = feature_extractor
|
||||
sd_models.apply_balanced_offload(pipe.feature_extractor)
|
||||
pipe = sd_models.apply_balanced_offload(pipe, force=True, silent=True)
|
||||
log.debug(f'IP adapter load: extractor={pipe.feature_extractor.__class__.__name__}')
|
||||
except Exception as e:
|
||||
log.error(f'IP adapter load: extractor {e}')
|
||||
errors.display(e, 'IP adapter: type=extractor')
|
||||
return False
|
||||
shared.state.end(jobid)
|
||||
sd_models.move_model(pipe.feature_extractor, devices.device)
|
||||
return True
|
||||
|
||||
|
||||
|
||||
+15
-8
@@ -276,13 +276,20 @@ class OffloadHook(accelerate.hooks.ModelHook):
|
||||
log.trace(f'Offload: type=balanced op=dispatch map={device_map}')
|
||||
if device_map is not None:
|
||||
skip_keys = getattr(module, "_skip_keys", None)
|
||||
module = accelerate.dispatch_model(module,
|
||||
main_device=torch.device(devices.device),
|
||||
device_map=device_map,
|
||||
offload_dir=offload_dir,
|
||||
skip_keys=skip_keys,
|
||||
force_hooks=True,
|
||||
)
|
||||
try:
|
||||
module = accelerate.dispatch_model(module,
|
||||
main_device=torch.device(devices.device),
|
||||
device_map=device_map,
|
||||
offload_dir=offload_dir,
|
||||
skip_keys=skip_keys,
|
||||
force_hooks=True,
|
||||
)
|
||||
except Exception as e: # reapply hook
|
||||
log.warning(f'Offload: type=balanced op=dispatch module={module.__class__.__name__} {e}')
|
||||
module = accelerate.hooks.remove_hook_from_module(module, recurse=True)
|
||||
module.balanced_offload_device_map = None
|
||||
sd_models.move_model(module, devices.device, force=True)
|
||||
module = accelerate.hooks.add_hook_to_module(module, self, append=True)
|
||||
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
|
||||
@@ -498,7 +505,7 @@ def apply_balanced_offload(sd_model=None, exclude: list[str] | None = None, forc
|
||||
continue
|
||||
module.module_name = module_name
|
||||
module.offload_dir = os.path.join(shared.opts.accelerate_offload_path, checkpoint_name, module_name)
|
||||
apply_balanced_offload_to_module(module, op='apply')
|
||||
apply_balanced_offload_to_module(module, op='apply', force=force)
|
||||
if not silent:
|
||||
report_model_stats(module_name, module)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user