Keep native ZeroGPU module-scope pack; build generator on CPU to avoid poisoning fork

Fix module-scope CUDA crash: defer generator/timesteps/empty_cache, finalize on GPU in worker

Preload model + AoTI kernels at module scope (avoid per-session reinit)

Populate /api/settings schema; hide shared-GPU queue note

fix: explicit / route so StaticFiles no longer shadows /gradio_api/startup-events

Upload folder using huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

add fork-queue bridge probe (parent -> held GPU worker)

add cross-call GPU state probe (gpu_fuse / gpu_step)

Upload app.py with huggingface_hub

add AoTI constant-FQN diagnostic + alt-portrait check

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Add AoTI loader + verification endpoint

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Unpin transformers (4.36.2 caps hub<1.0, conflicts with gradio 6); CLIP usage is version-stable

Drop hub pin (conflicts with gradio 6); shim removed hub symbols for diffusers 0.27 / transformers 4.36

ZeroGPU backend self-test: PersonaLive pipeline on Blackwell

initial commit