Spaces:
Sleeping
Sleeping
Commit History
29 days ago
-
Cut per-request time from ~25 s to ~4 s
Verified signatureProfiling showed the model itself takes <0.5 s; the rest was post-processing. - skip rendering and encoding the input-view path videos, which the app never shows (only moved/) - skip per-frame semantic decoding, whose video is never written - global PCA of the feature video on the GPU instead of scikit-learn on the CPU (same covariance-eigh math and sign convention; 8 s -> 0.1 s, <=1/255 difference) - write the PLY with a structured-array view instead of per-row tuples (byte-identical, 2.4 s -> 0.2 s) - GPU duration 120 s -> 60 s9b8f29a -
Port the demo to current ZeroGPU (torch 2.8, CUDA 12.8, Gradio 6)
Verified signatureZeroGPU now runs on RTX Pro 6000 Blackwell (sm_120) and only accepts torch>=2.8, so the torch 2.1.2 stack and its sm_60-86 extension wheels could no longer start. - requirements: torch 2.8.0, official flash-attn 2.8.3.post1 wheel (has sm_120), spconv-cu126 (compute_90 PTX JITs on sm_120), pytorch-lightning for unpickling the LSeg checkpoint - rebuild torch_scatter / diff_gaussian_rasterization (32 feature channels) / simple_knn / curope for sm_80-120; drop unused pointops; build notes in wheel/README.md - torch.load(weights_only=False) for the checkpoints, matplotlib 3.9 get_cmap, curope scalar_type - Gradio 6.28, lazy example caching, GPU duration 120 sfe8f1ba
-
Update Code
57746f1
