Spaces:

---

Duplicated from  FINAL-Bench/Darwin-35B-A3B-Opus

build-small-hackathon

/

Darwin-35B-A3B-Opus

like 2

Paused

The current app.py loads the 67GB model at module level (outside @spaces.GPU), which crashes during Gradio startup on CPU. This PR moves model loading into a lazy cache inside @spaces.GPU, so it only loads when GPU is active. Also increases GPU duration from 180s to 300s to accommodate cold model load.

Changes:

• Replace module-level AutoTokenizer.from_pretrained + AutoModelForCausalLM.from_pretrained with lazy _load() inside @spaces.GPU
• Increase duration from 180s to 300s
• Add _model_cache to avoid reloading per call