Fix: disable track_tqdm (ProgressUnit crash) for API use

VAE decode untiled (80GB) - tiled_decode frame-dim mismatch 141!=138

Fix scheduler IndexError: use enumerated timestep, not mutable timesteps[progress_id]

Persist model cache on /data (symlink) for fast cold-wakes

Native GPU load (revert offload) for full speed+quality on 48GB

Load model to CPU then enable_vram_management streams to GPU (was OOM placing on cuda)

L4/A10G fit: enable_vram_management(num_persistent_param_in_dit=0) CPU-stream DiT

Pin python 3.10 (scipy/numpy/torch wheels exist; avoids gfortran source build)

Update app.py

Update app.py

Update app.py

Update requirements.txt

Update OmniAvatar/models/wan_video_dit.py

Update requirements.txt

error fixed

Update OmniAvatar/models/wan_video_dit.py

Update requirements.txt

Update requirements.txt

Update OmniAvatar/models/wan_video_dit.py

Upload fa3.py

Update requirements.txt

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

gpu based

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

clean up

do faster CPU based tts

Update app.py

Update app.py

Update requirements.txt

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py