Fix vision-tower numerical instability: mask-free segmented SDPA + pristine re-patch

diagnostic: fingerprint weights/kernels per GPU worker

Restore non-persistent indicator_token_indices buffer inside the GPU worker

diagnostic: A/B endpoint comparing old vs fixed scoring pipeline

diagnostic: log weight-loading info + temporary freeform debug endpoint

Drop flash_attn wheel (ABI-incompatible with torch 2.13)

Replace flash_attn with pure-torch SDPA stand-ins (no wheel for torch 2.13)

Match the authors' SA-IQA inference pipeline (system prompt, no think block, swift pixel budget, rating-token top-5 logprob scoring)

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload app.py with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

initial commit