Spaces:
Running on Zero
Running on Zero
Commit History
-
refactor: minimal hacks — own inference wrapper, trimmed stubs and deps, gradio 6.20
Claude Sonnet 4.5 +1: Claude Fable 5Unverified signature- sam3d_inference.py replaces notebook/inference.py: imports only the pipeline, dropping the kaolin.visualize/SceneVisualizer/plotly chain - attention backend pinned to sdpa via import order (upstream env mechanism) instead of source patching; flash_attn stub removed - kaolin stub reduced to utils.testing.check_tensor (flexicubes); pytorch3d stub reduced to transforms/structures/renderer - requirements: drop ~25 training-only packages (verified unreferenced in the inference import graph); move spconv/open3d/iopath from runtime pip into requirements; pin hydra-core 1.3.2 - drop hydra patch script call: it was invoked with bash against a python script and never ran — the successful e2e run proves it is not needed - drop utils3d monkey-patches: the pinned commit ships those functions - gradio 6.20.0 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>7444d60 -
fix: stub files contained HF redirect pages instead of Python
Claude Sonnet 4.5 +1: Claude Fable 5Unverified signaturecurl without -L saved the 307 redirect body. Re-downloaded with redirect-following and validated all 19 files with ast.parse. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>94b4398 -
fix: rebuild for ZeroGPU torch whitelist (2.8-2.11)
Claude Sonnet 4.5 +1: Claude Fable 5Unverified signatureZeroGPU rejects any torch pin outside 2.8.0/2.9.1/2.10.0/2.11.0, and there are no kaolin/pytorch3d/gsplat wheels for those versions. Instead of pinning old torch: - drop torch from requirements — use the ZeroGPU-provided build - stub kaolin/pytorch3d/flash_attn (adopted from jboth/sam3d-objects-fixed); only their pure-python surface is used since texture baking, mesh postprocess and layout postprocess are disabled in Inference.__call__ - install spconv-cu124 (torch-independent), gsplat (PyPI JIT mode, kernels never compiled), sam2/MoGe/utils3d at runtime against the live torch - force sdpa attention (patch inference_pipeline.py H200 flash_attn autodetect) - checkpoints via symlink to snapshot_download, config = pipeline.yaml - add Diagnose tab for module/GPU health checks Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>ee8c843