Vae-class components never take group hooks, so group mode kept them
resident on the gpu; a MiniMax-class video vae holds about 10GB that
way while running only seconds per generation. Components above 1GB
now rest in system memory: the apply_forward_hook bridge on encode and
decode fires an on-demand hook that moves the whole module to the
device, so tiled calls find every weight already loaded, and the
processing seams return it to cpu once outputs are materialized. Small
vaes stay resident since the transfer would cost more than it frees.
- placement is decided per component by measured size and requires the
entry bridge; components without it stay resident
- move_model no longer forces on-demand vaes to the gpu for
non-txt2img tasks, and full_vae_encode onloads before binding the
input, which otherwise lands on the resting device
- mode switches clear the stamp and hook in both directions
Group offload is applied per component through one engine shared by
regular and modular pipelines. Each component carries a config
signature: re-application with unchanged settings is a no-op instead
of raising before the first forward or silently keeping a stale
config, and changed settings remove the hooks and reapply. Switching
offload modes cleans up the previous mode's hooks in both directions.
- text encoders always offload at leaf level without streams, so
their weights are never held in pinned host memory
- the vae never takes group hooks and stays resident: the hooks are
forward-scoped, while pipelines enter through encode/decode and
tiled calls re-enter per tile
- new pin offload memory option: disabled pins one group at a time
instead of holding the whole module in non-pageable memory, and
modules larger than half of system memory degrade automatically
- record stream is clamped to stream mode; the standalone
combination is rejected upstream
Entries in the always and never lists are matched against the pipeline
component name (text_encoder, vae) as well as the model class name, so
one entry covers every architecture instead of needing a new class name
per model. Class entries keep working unchanged.
Never is still tested first, so a class entry there exempts a single
model from a component entry in the always list.
This line resulted in the cleanup check never running if the current gallery has fewer than 1000 files, no matter how many excess entries there are in the cache.
- fold the enable rows into the type rows: disabled, enabled or an explicit dtype
- pv keep-unquantized value is disabled; enabled means int8 pv
- add a text encoder row judged at te geometry
- update locale hints, drop the orphaned checkbox hint
The refine toggle doubles as the slot selector on the UNet/DiT page:
with it active a card click sets sd_unet_secondary instead of sd_unet,
mirroring base/refiner selection on the model page.
One UNET override cannot serve dual-transformer arches: ideogram4
conditional/unconditional and wan combined-stage experts need separate
files, and previously a single override landed on both experts.
- sd_unet_secondary option with per-slot tracking, consumed-state sync,
arch-change reset, and incompatible-override fallback
- dropdown renders beside the primary, follows it into quicksettings,
and is visible only for dual-transformer model types
- ideogram4 native single-file spec with a quant-aware fused-qkv
converter; such converters run before comfy_quant detection via
TransformerSpec.converter_handles_quant
- quicksettings render in configured order (sort keyed on the option
object and always fell back to alphabetical)
- post-load dtype warning skips quantized transformers
Wan 2.2 A14B ships a per-model boundary_ratio (0.9 I2V, 0.875 T2V) that selects the high- or low-noise expert per step. The video and base-model image loaders both load the shipped value; the slider override is applied at generation time in set_pipeline_args, the one point both paths pass through before invoking the pipeline.
The denoising loop reads config.boundary_ratio each call, so tuning takes effect with no reload for video and base-model images alike. The slider defaults to -1, meaning use the model's value; 0 to 1 set the boundary explicitly. Single-expert stages stay load-time because they drop a transformer to free VRAM.
A single handler bound to the AR dropdown, width and height (in the
shared create_resolution_inputs and the resize section) wrote both
sliders on every change. With a ratio selected it locked one axis and
snapped it back on each edit, the two sliders looped, and the math ran
server-side one round-trip per keystroke, echoing a value back into the
field being typed in and yanking it.
Move aspect-ratio linking to the browser (ui/resolutionLock.ts): a
debounced edit writes only the partner axis, never the field being
edited, and commits immediately on blur, enter, or slider release. Keep
the kanvas notify on notifyKanvasResize wired to the resize sliders'
gradio .change, so it still fires on programmatic size updates (detect,
paste, swap) that client-side listeners miss. Drop the per-change AR
wiring, res_apply, and the resolutionChange* helpers.
CalamitousFelicitousness
Vladimir Mandic
awsr