Update GitHub URL to UVA-Computer-Vision-Lab/sab-3r

Keep upsampler on CPU on HF Spaces (fixes text-query _cuda_init crash) — verified locally

ZeroGPU: patch cached device descriptors (working_device) to CPU after roundtrip — verified locally

ZeroGPU: force return value to CPU via torch.save/load roundtrip

ZeroGPU: move tensors to CPU on return; drop @GPU from post-processing; harden SparseGAState.__del__

Move SAB3R model to module-level singleton for ZeroGPU (fixes CPU/CUDA weight mismatch on worker)

Make transpose_to_landscape picklable for ZeroGPU (closures -> __call__ classes)

Import spaces before torch (ZeroGPU init order requirement)

Fix CLIP BPE vocab download: derive path without importing the broken maskclip package

Download CLIP BPE vocab file missing from FeatUp package

Bump Gradio SDK to 5.33.0 (gradio 4.44.1 imports HfFolder which is gone from huggingface_hub>=1.0)

Set TORCH_CUDA_ARCH_LIST=7.5;8.0;8.6;8.9;9.0 for FeatUp runtime install (no GPU visible at app startup on ZeroGPU)

Move FeatUp install to runtime (app.py); drop git+ line from requirements.txt

Drop --no-build-isolation from requirements.txt (pip doesn't support it there); rely on PIP_NO_BUILD_ISOLATION=1 Space variable

Fix FeatUp install: use --no-build-isolation so setup.py sees pre-installed torch

Reorganize: demo lives in demo/ in the GitHub repo; at Space root we keep app.py+demo.py next to mast3r/ and dust3r/

Add ZeroGPU support (@spaces.GPU decorators)

Make demo device-adaptive (cuda if available, else cpu)

Remove RoPE CUDA build artifacts (rebuilt on startup)

Initial SAB3R demo release

initial commit