Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix(lora): silence offload re-init logging on network changes

Browse Source 
Rebuilding balanced offload before touching weights reconstructs the
OffloadHook, whose constructor prints the op=init banner and module
inventory meant for model load, so every network switch replayed the
full load-time announcement. The hook constructor and the model summary
now honor the silent flag and the network activate, deactivate, and
selection paths pass it; real model loads keep the full output.
This commit is contained in:
CalamitousFelicitousness
2026-07-19 19:10:20 +01:00
parent b90afe1253
commit c75e9410fe
3 changed files with 9 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/lora/lora_stack.py
+1 -1
View File
@@ -252,7 +252,7 @@ def materialize_model():
"""Weight-kind selection rewrites module weights outside the activation walk; rebuild balanced-offload modules real first (mirrors network_activate)."""
from modules import sd_models
if getattr(shared.opts, 'diffusers_offload_mode', None) == 'balanced' and getattr(shared, 'sd_model', None) is not None:
sd_models.apply_balanced_offload(shared.sd_model, force=True)
sd_models.apply_balanced_offload(shared.sd_model, force=True, silent=True)
def finalize(total_steps):
modules/lora/networks.py
+2 -2
View File
@@ -63,7 +63,7 @@ def network_activate(include=None, exclude=None):
sd_models.disable_offload(sd_model)
sd_models.move_model(sd_model, device=devices.cpu)
elif shared.opts.diffusers_offload_mode == "balanced":
sd_model = sd_models.apply_balanced_offload(sd_model, force=True) # dispatched modules hold meta tensors backed by the offload map; rebuild them real on cpu with hooks intact before touching weights
sd_model = sd_models.apply_balanced_offload(sd_model, force=True, silent=True) # dispatched modules hold meta tensors backed by the offload map; rebuild them real on cpu with hooks intact before touching weights
group_offload = shared.opts.diffusers_offload_mode == "group"
group_stripped = {}
device = None
@@ -259,7 +259,7 @@ def network_deactivate(include=None, exclude=None):
sd_models.disable_offload(sd_model)
sd_models.move_model(sd_model, device=devices.cpu)
elif shared.opts.diffusers_offload_mode == "balanced":
sd_model = sd_models.apply_balanced_offload(sd_model, force=True) # dispatched modules hold meta tensors backed by the offload map; rebuild them real on cpu with hooks intact before touching weights
sd_model = sd_models.apply_balanced_offload(sd_model, force=True, silent=True) # dispatched modules hold meta tensors backed by the offload map; rebuild them real on cpu with hooks intact before touching weights
group_offload = shared.opts.diffusers_offload_mode == "group"
group_stripped = {}
modules = {}
modules/sd_offload_balanced.py
+6 -5
View File
@@ -14,7 +14,7 @@ import modules.sd_offload_state as s
class OffloadHook(accelerate.hooks.ModelHook):
def __init__(self, checkpoint_name):
def __init__(self, checkpoint_name, silent=False):
if shared.opts.diffusers_offload_max_gpu_memory > 1:
shared.opts.diffusers_offload_max_gpu_memory = 0.75
if shared.opts.diffusers_offload_max_cpu_memory > 1:
@@ -32,8 +32,9 @@ class OffloadHook(accelerate.hooks.ModelHook):
self.last_pre = None
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}'
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}')
if not silent:
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}'
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__()
@@ -255,7 +256,7 @@ def apply_balanced_offload(sd_model=None, exclude: list[str] | None = None, forc
checkpoint_name = sd_model.sd_checkpoint_info.name if getattr(sd_model, "sd_checkpoint_info", None) is not None else sd_model.__class__.__name__
if force or (s.offload_hook_instance is None) or (s.offload_hook_instance.min_watermark != shared.opts.diffusers_offload_min_gpu_memory) or (s.offload_hook_instance.max_watermark != shared.opts.diffusers_offload_max_gpu_memory) or (checkpoint_name != s.offload_hook_instance.checkpoint_name):
cached = False
s.offload_hook_instance = OffloadHook(checkpoint_name)
s.offload_hook_instance = OffloadHook(checkpoint_name, silent=silent)
if cached and shared.opts.diffusers_offload_pre:
s.debug_move('Offload: type=balanced op=apply skip')
@@ -277,6 +278,6 @@ def apply_balanced_offload(sd_model=None, exclude: list[str] | None = None, forc
process_timer.add('offload', t1 - t0)
fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access
s.debug_move(f'Apply offload: time={t1 - t0:.2f} type=balanced fn={fn}')
if not cached:
if not cached and not silent:
log.info(f'Model class={sd_model.__class__.__name__} modules={len(s.offload_hook_instance.offload_map)} size={s.offload_hook_instance.model_size():.3f}')
return sd_model