Spaces:
Configuration error
Space down on a torch pin ZeroGPU no longer accepts โ diagnosis and the rebuild it needs
This Space is currently down before any user code runs, and the cause is in its own configuration.
Checked today against the Hub API: the runtime stage is CONFIG_ERROR, no hardware is assigned, and the platform's message is that the torch version in requirements.txt is not compatible with ZeroGPU. The supported versions it lists are 2.11.0, 2.10.0, 2.9.1, and 2.8.0. The file here still pins torch==2.4.0 with torchvision==0.19.0, so the configuration fails before assignment. The requested hardware is zero-a10g, the custom domain is ready, and the application code itself is not the blocker โ it uses a plain PyTorch surface.
One honest caution: bumping the pin alone will not revive the demo. Four pinned artifacts are compiled against the old stack โ the flash-attention wheel (cu12 / torch 2.4 / cp310), spconv-cu120==2.3.6, and the two cp310 wheels for diff_gaussian_rasterization and nvdiffrast. On a current ZeroGPU image those will not install or import as they stand.
What must hold for a revival:
torchandtorchvisionmove to a supported, matched pair โ 2.8.0 with 0.23.0 is the conservative step, and the index line must point at an index that serves that release.- Flash-attention is replaced with a wheel built for the chosen torch, CUDA, and Python, or the code falls back to PyTorch SDPA.
spconvmoves to a build matching the new torch.- The two rasterizer extensions are built from source in the Space's build instead of loading the cp310 wheels.
xformersmoves to a release matching the new torch, or is removed if the inference path does not use it.
A simple check would prove this wrong: change only the torch pin and restart โ if the Space builds and serves, the rebuild list above is more than it needs.
The small next step is step 1 plus a build log: apply the pin change, let the build fail or pass on the compiled wheels in the open, and work through the list from there. I am glad to share a worked requirements and build diff for the full five steps if a maintainer would like it.
What this is: the worked requirements and build diff for the full
five steps, as offered in the diagnosis discussion โ complete, not a
step-1 fragment. It contains the full replacement requirements.txt
and README.md, a unified diff for each against the files currently
in the repo, the two rebuilt rasterizer wheels (cp312 / torch 2.8),
the script that built them, a runbook the owners can execute end to
end, and the verification record.
The diagnosis in one line: the Space sits at CONFIG_ERROR becauserequirements.txt pins torch 2.4.0, which ZeroGPU no longer accepts โ
and four more artifacts in the same file (the flash-attention wheel,
spconv, the two rasterizer wheels, xformers) are pinned to that same
old stack, so the pin alone cannot revive it.
What was verified: every version pairing against its project's own
index or release listing; a full pip resolution of the new
requirements.txt on Python 3.12.3 (136 packages, no conflicts โ
including torch 2.8.0+cu128 with xformers 0.0.32.post2, which pins
that exact torch); and both rasterizer wheels genuinely built from
the sources the code calls and run on a CUDA GPU.
The named remainder: the Space's own build on ZeroGPU cannot be
run from outside โ that last check belongs to a maintainer's restart,
and the runbook says what each possible failure line would mean.
The kit: https://huggingface.co/datasets/nootxlm/trellis-rebuild-kit