Bump sympy to 1.13.3 (torch 2.8.0 requires >=1.13.3)

Pin torch==2.8.0 / torchvision==0.23.0 (prevent transitive downgrade to 2.6 without sm_120)

Pin starlette<0.38 / fastapi<0.112 (0.38+ passes dict cache key, breaks jinja2 LRU)

Drop fastapi/starlette pins (starlette<0.39 triggers Jinja unhashable-dict in template cache)

Update requirements.txt

Unpin bitsandbytes (0.43.3 imports triton.ops, removed in triton 3.x)

Adapt to ZeroGPU: load models at module scope; @spaces.GPU on guess/segment; drop ProcessPool + triton pin

Update app.py

Update app.py

Update README.md

Update app.py

Update app.py

Update app.py

Update requirements.txt

Initial Commit