| # gradio itself is provided by the Space SDK (sdk_version in README.md). | |
| # `spaces` provides the @spaces.GPU decorator used on ZeroGPU. app.py shims it to | |
| # a no-op when it's absent, so it's only actually required on the Space. | |
| spaces | |
| torch # ZeroGPU requires torch >=2.8; the Space runtime supplies the CUDA build | |
| transformers # MusicGen + (on ZeroGPU) the MiniCPM5-1B enricher; needs >=5.6 for MiniCPM5 | |
| # Enrichment LLM: on a ZeroGPU Space it's openbmb/MiniCPM5-1B, pulled from the Hub | |
| # at startup (no extra deps — standard Llama arch + fast tokenizer). For LOCAL dev | |
| # instead, run Ollama with the model below pulled: | |
| # ollama pull llama3.2:3b | |
| # The app runtime needs nothing more (ambience.py reads beds via stdlib `wave`). | |
| # To (re)populate the sampled ambience beds in assets/ambience/, pick one: | |
| # download from Wikimedia Commons: uv pip install soundfile av | |
| # python scripts/fetch_ambience.py | |
| # or generate them: uv pip install diffusers | |
| # python scripts/make_ambience.py | |