feat(inference): multi-GPU auto-layout and cross-GPU routing (PR2) - #1263

Open

greenstephen wants to merge 14 commits into

ace-step:mainfrom

greenstephen:feat/multi-gpu-pr2-auto-layout

Commits on Jul 6, 2026

feat(inference): add ComponentDeviceMap for multi-GPU placement (PR1)

Steve
and
cursoragent
committed
4cc120b
feat(inference): add multi-GPU auto-layout and cross-GPU routing (PR2)

Steve
and
cursoragent
committed
a38527f
refactor(device_map): split package to satisfy 200-LOC module cap.

Steve
and
cursoragent
committed
83c3c0e
fix(inference): complete cross-GPU diffusion routing and APG device index.

Steve
and
cursoragent
committed
7d2b302
fix(device_map): reserve DiT VRAM for co-located LM auto-layout fallback.

Steve
and
cursoragent
committed
18c8d4a
fix: address CodeRabbit parsing dead code and APG MPS regression.

Steve
and
cursoragent
committed
c1d9d98
Fix silence_latent device check for multi-GPU CUDA indices.

Steve
and
cursoragent
committed
b6d4061
Hoist normalize_component_device import and narrow device errors.

Steve
and
cursoragent
committed
0c21299
Catch specific exceptions in device string parsing helpers.

Steve
and
cursoragent
committed
f7ba683
Fix VAE VRAM preflight to use mapped component device index.

Steve
and
cursoragent
committed
8f1463f
Raise DeviceMapError for malformed CUDA device indices.

Steve
and
cursoragent
committed
1d19b9b
Split generate_music_decode tests to satisfy 200 LOC cap.

Steve
and
cursoragent
committed
c3ef2df
Use normal imports in generate_music_decode test support.

Steve
and
cursoragent
committed
7aee60f

Commits on Jul 11, 2026

fix(inference): honor mapped cuda:N for LM init and Gradio wiring

Steve
and
cursoragent
committed
ec02094