Fixes the ZeroGPU build so the Space matches the working local version.
Functional fixes:
- requirements.txt: add voxcpm (was missing -> narration always fell back to a silent clip) + numpy.
- app.py: load FLUX / MiniCPM5 / VoxCPM2 on cuda at MODULE scope (the documented ZeroGPU pattern) instead of lazy-loading inside @spaces.GPU with enable_model_cpu_offload() -- that was why FLUX failed and images dropped to the sketch fallback. Flow is now sequential (one GPU per request) and real errors surface in the trace instead of silently degrading.
Cleanup:
- Remove Tiny Mode (sd-turbo/sdxl-turbo) -- redundant on a real GPU.
- Trim config.py to the 3 sponsor models and add an explicit models: list in the README so the Space links exactly FLUX.2-klein-4B, MiniCPM5-1B, VoxCPM2.
- Delete local-only / Modal / doc files that do not belong on a ZeroGPU Space.
Restore Codex zerogpu-space-copy: model caching, theme templates, parallel TTS, heartbeat
Codex
0babffc
Use Codex zerogpu-space-copy app.py with model caching, theme templates, parallel TTS
Codex
18d8424
DoodleBook: ZeroGPU app with parallel TTS, heartbeat, queue
