Warm VoxCPM2 at startup (under ZeroGPU emulation) + load_denoiser=False — fixes lazy-import CUDA-init error

Part 2: @app .api game endpoints + @gradio/client frontend (ZeroGPU auth-forwarded GPU calls)

Part 1: mamba-ssm wheel + module-level Nemotron load (real dispatcher on GPU)

Register a module-level @spaces.GPU function (HF ZeroGPU requires one at startup)

Bundle assets (drop snapshot_download — its async teardown corrupted the server fd)

Non-daemon keepalive thread so the process stays resident under reload

Test python 3.10.13 (her-trace ver) + lazy model — isolate the serving keep-alive

Force module-level model load (her-trace keep-alive) + plain blocking launch

Keep main thread alive after launch (SSR on) so the Node proxy survives

Plain blocking launch with SSR on (match her-trace exactly)

Block main thread after launch (reload makes gr.Server launch non-blocking)

Disable GRADIO_WATCH_DIRS hot-reload (it breaks gr.Server launch blocking)

Self-contained root app.py = gr.Server + module-level launch (her-trace clone)

Remove importlib root app.py (app_file now frontend/server/app.py)

Launch gr.Server at MODULE level (HF gradio reload mode skips __main__)

Switch to gr.Server (Gradio Server mode) — fixes the 7860 double-bind

Fix Space entry: gradio host (demo.launch) instead of uvicorn (was double-binding 7860)

Deploy Mumbai Local Panic — Nemotron(ZeroGPU) dispatcher + VoxCPM2 voice