Force model to real CUDA inside @spaces.GPU; add device/amp timing diag

Make the GPU call non-generator (fix ZeroGPU hang)

Fix FA3 kernel load: get_kernel requires version=

Use Flash-Attention 3 (prebuilt via HF kernels) as the attention backend

Perf: cap views to 16, duration 180s, phase timing, free memory, attn backend

Simplify app.py; bf16 inference; unified video->images input; drop onnx

Add inference logging + disable Gradio SSR

Consolidate into a single PEP 723 app.py; fix ZeroGPU GPU-fn detection

Commit the dvlt wheel (uv had dropped a .gitignore in wheels/)

Install dvlt from a vendored prebuilt wheel (ZeroGPU, no git/sys.path)

Vendor edited dvlt in packages/dvlt; install via git subdirectory (ZeroGPU)

Unofficial DVLT 3D reconstruction demo (Gradio + ZeroGPU)