Simplex offloading (2x speed) - #1644

Open

dxqb wants to merge 15 commits into

Nerogar:masterfrom

dxqb:streaming-simplex-squashed

Commits on Jul 23, 2026

Replace per-model device-placement methods with a generic materialize()/evict() API

dxqb
and
claude
committed
1150660
Use materialize_only("vae") consistently in debug-save and VAE sampler
dxqb
committed
43824a5
Extract shared transformer-loading helper into HFModelLoaderMixin

dxqb
and
claude
committed
35a5dfa
Arena MemPool + single-buffer offload cache

dxqb
and
claude
committed
90f1616
Pool each component's LoRA in a per-stem MemPool
dxqb
committed
e27dc4c
[centralize-setup-optimizations] Extract shared setup_optimizations helper into BaseModelSetup
dxqb
committed
ed56354
Add ModelType.trainable_parts() and TrainConfig.part_trained_in_place()

dxqb
and
claude
committed
44c80f7
Stream layers through the offload conductor from checkpoint (disk-offload)
dxqb
committed
3d9dc78

Commits on Jul 24, 2026

Skip dequantize for unquantized W8A8 weights

dxqb
and
claude
committed
bb4fbb6
Full-model-buffer offload for frozen-base simplex mode
dxqb
committed
96ffea2
Join stream-reader threads on abort, not just on success

dxqb
and
claude
committed
417cb5d

Commits on Aug 13, 2026