Streaming model loader - #1642

Open

dxqb wants to merge 10 commits into

Nerogar:masterfrom

dxqb:streaming-squashed

Commits on Jul 23, 2026

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

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

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

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

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

Commits on Jul 24, 2026