README.md

metadata

title: OneAlign Scorer (CPU fork)
emoji: ⚖️
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 4.44.1
python_version: '3.10'
app_file: app.py
pinned: false
license: mit

OneAlign Scorer — CPU Fork (personal experiment)

This is a fork of teowu/OneScorer — the official Space for Q-Align (ICML 2024).

What this fork changes

1. Remove broken examples=... — the original Space's example assets are git-LFS pointers that fail to resolve on cold start, causing gradio.Interface(... examples=[...]) to throw PIL.UnidentifiedImageError and the whole Space to crash.
2. Auto-pick device (CUDA → CPU) so it can attempt to boot on free CPU hardware. Original was hard-coded to cuda:0.
3. cache_examples=False, show_api=False for stability.

⚠️ Performance warning

The underlying model is q-future/one-align, a ~8B mPLUG-Owl2 LMM.

• On GPU (1620GB VRAM): ~13 s per image. Works great.
• On free CPU Space (this fork's hardware): the model may load successfully but a single inference can take many minutes, or hit RAM/timeout limits. This Space is uploaded purely for boot-time troubleshooting and as a stepping stone toward a paid GPU upgrade.

If you need real usage, please either:

• Run locally on a GPU machine, or
• Duplicate this Space and upgrade to a GPU runtime (T4-small ≈ $0.40/hour) in Space settings.

Credits

Paper: Wu et al. Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels (ICML 2024). arXiv:2312.17090

Original code: https://github.com/Q-Future/Q-Align Original Space: https://huggingface.co/spaces/teowu/OneScorer