feat(models) - Smart Partial Loading - Attempt To Measure Actual VRAM Requirements#9266

Open

kappacommit wants to merge 21 commits into

invoke-ai:mainfrom

kappacommit:feat/denoise-working-memory-estimator

Commits on Jun 2, 2026

feat(mm): estimate denoise working memory so more of the model loads into VRAM

Your Name
and
claude
committed
feat(mm): wire denoise working-memory estimator into the remaining 5 architectures

Your Name
and
claude
committed

Commits on Jun 3, 2026

Commits on Jun 4, 2026

fix(mm): partial-load headroom to prevent high-res thrash under smart_partial_loading

Your Name
and
claude
committed
style: ruff format (wrap long model_on_device call in denoise_latents)

Your Name
and
claude
committed
chore(api): regenerate openapi schema + types for smart_partial_loading config field

Your Name
and
claude
committed
Merge branch 'main' into feat/denoise-working-memory-estimator
kappacommit
authored

Commits on Jun 8, 2026

Merge branch 'main' into feat/denoise-working-memory-estimator
Pfannkuchensack
authored

Commits on Jun 9, 2026

fix(mm): tune partial-load headroom to 1.5 and flux2 multiplier to 2.2

Your Name
and
claude
committed
docs(mm): drop internal memory-audit/ reference from denoise estimator docstring

Your Name
and
claude
committed
fix(mm): keep partial-load headroom reserve in effect through the self-unload path

Your Name
and
claude
committed
test(mm): cover cache-side working-memory reserve semantics

Your Name
and
claude
committed
refactor(mm): derive the raised-working-mem flag at the app boundary

Your Name
and
claude
committed
refactor(mm): consolidate the denoise working-memory estimator API

Your Name
and
claude
committed
docs: regenerate docs data for smart_partial_loading + LoggerInterface.isEnabledFor

Your Name
and
claude
committed
docs: regenerate settings data with posix path separators

Your Name
and
claude
committed
test(mm): stub the RAM-cache sizing heuristic for CPU-only CI

Your Name
and
claude
committed