Spaces:
Running on Zero
Official Space requests GPU duration beyond anonymous allowance
The official Space's demo.py currently estimates GPU duration as min(900, max(240, 240 + 60 * file_count)), independent of the beam count. For one file this estimates 300 seconds.
An anonymous single-model test on 2026-11-06 was rejected before inference with a requested duration of 450 seconds exceeding the allowed maximum. That observation and the current estimator are separate: the cause of the difference in reported durations was not established. The inference has not been rerun against the current source.
Could the Space's duration estimate and startup path be reviewed against actual cold-start and inference measurements? Reducing the requested duration without measuring the cold path could cause tasks to time out, so an input-dependent estimate should preserve sufficient time for the chosen workload.
Source: https://huggingface.co/spaces/VAST-AI/SkinTokens/blob/main/demo.py