Spaces:
Running on Zero
Running on Zero
Commit History
-
feat: Update context. Running
29acae6 -
fix: Make the demo usable from the page, not just the API.
Three things blocked the browser while the endpoint itself was fine: - gr.Examples cannot hold a File input. Gradio drops the File column from the examples table, so the row carried five values into six inputs and every argument after the scene name shifted by one, putting the frame count where the archive path goes. Dropped the table; the dropdown already selects the only example scene and the sliders already default to its settings. - Spaces enable Gradio's server-side rendering through GRADIO_SSR_MODE, and its node server was answering the page's POSTs with `405 POST method not allowed`. Rendering client side keeps every request on the Python backend. - The upload rejected archives client side on extension alone. unpack_scene checks the magic bytes, which is the stronger test. Also prune unpacked scenes, not just recordings, so uploads cannot fill the disk. Verified over HTTP with a real multipart upload: example scene and uploaded zip both run. Co-Authored-By: Clau…eee12ae -
fix: Compile the vis4d kernels here instead of installing them.
pip cannot install vis4d_cuda_ops on a Space. It publishes no wheels, and pip compiles it in an isolated build environment that resolves its own newest torch (2.13.0) while the Space caps the runtime torch at 2.11.0, so the extension imports with `undefined symbol: c10::NotImplementedError`. Setting PIP_NO_BUILD_ISOLATION as a Space variable does not help, the build does not see it. A local vis4d_cuda_ops package now shadows the real one and compiles only the upstream CPU translation units against the torch the Space runs, which cannot drift. iou_box3d runs on CPU for the few boxes the track graph compares, and deformable attention runs the grid-sample path already in mapdet3d, on the GPU. Verified against the CUDA kernels on a 4090: same 19 tracks over the example scene, same 5.5 s for 21 frames, and ~10 s to compile on first import. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>7266be4 -
fix: Build vis4d_cuda_ops against the torch the Space runs.
With pip build isolation the extension compiled against torch 2.13.0 while the Space caps the runtime at 2.11.0, so importing it died on `undefined symbol: c10::NotImplementedError`. The Space now sets PIP_NO_BUILD_ISOLATION=1. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>791f52a -
feat: Add the Gradio demo for Map-Det3D.
Ports scripts/demo.py to a ZeroGPU Space. Posed frames are streamed through the model one at a time and the tracked world-frame boxes are shown in an embedded Rerun viewer. - Vendors mapdet3d so the Space does not need a published package. - Builds the MapAnything backbone from its config alone. The Map-Det3D checkpoint overwrites every backbone weight, so fetching the MapAnything and torch hub DINOv2 weights only cost ~10 GB per cold start. - Logs the scene mesh once instead of once per frame, and logs frames at a 640 px long side with matching intrinsics, which keeps the recording at ~27 MB instead of ~250 MB. - Starts frame IDs at 0 so the streaming window and the track graph reset between runs. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>0122a25
-
initial commit
Verified signature861c9ce