Spaces:

lemonaddie

/

geowizard

like 71

Running on Zero

Hi @lemonaddie , we've assigned ZeroGPU to this Space. Please check the compatibility and usage sections of this page so your Space can run on ZeroGPU.

Hi @lemonaddie , we've assigned ZeroGPU to this Space. Please check the compatibility and usage sections of this page so your Space can run on ZeroGPU.

Dear hyst, thank you for your nice ZeroGPU! Could you please share some information about the exact version (i.e. 11.8, 12.2) of the CUDA running on ZeroGPU? We are now planning to release a version supporting 3D reconstruction, which requires a package [cupy] corresponding to some specific CUDA version.

Hi @lemonaddie , we've assigned ZeroGPU to this Space. Please check the compatibility and usage sections of this page so your Space can run on ZeroGPU.

Dear hyst, thank you for your nice ZeroGPU! Could you please share some information about the exact version (i.e. 11.8, 12.2) of the CUDA running on ZeroGPU? We are now planning to release a version supporting 3D reconstruction, which requires a package [cupy] corresponding to some specific CUDA version.

We tried a lot of times but still can not install [cupy] on the zerogpu.....

@lemonaddie Thanks for testing ZeroGPU and sorry for the inconvenience. I think the CUDA version in ZeroGPU environment is 11.7. As CUDA is not available outside of the function decorated with @spaces.GPU, some Python packages are not compatible with ZeroGPU, and cupy might be one of them. cc @cbensimon
If this Space doesn't work on ZeroGPU, we can switch the hardware to a normal A10G.

Oh, looks like you solved the cupy issue yourself. I just duplicated this Space and tried app_recon.py, but the reconstruction seemed to work.
But I wonder why this Space is currently using app.py. Are there any other issues?

Oh, looks like you solved the cupy issue yourself. I just duplicated this Space and tried app_recon.py, but the reconstruction seemed to work.
But I wonder why this Space is currently using app.py. Are there any other issues?

Thanks for your testing hysts! We tried different versions of [cupy] but non of them can support the reconstruction procedure (After depth and normal prediction) for the online ZeroGPU environment. We can install the [cupy] packages on the device, but:
(1) For cupy_cuda11x, we can import cupy but there will be some error. After we get depth and normal, the succeeding reconstruction was interrupted by some unexpected issues.
(2) For all other versions, we cannot import cupy successfully.

For users' convenience, we might not be able to test the environments again and again.

We wonder how did you make it work? We would gratefully appreciate it if you could help us out!

@lemonaddie

(1) For cupy_cuda11x, we can import cupy but there will be some error. After we get depth and normal, the succeeding reconstruction was interrupted by some unexpected issues.

Hmm, weird. I simply duplicated your Space and changed app_file. (I also changed the sdk_version to 4.23.0 too, but I don't think it's related.)

What was the error? If it was GPU task aborted error, it's raised when the function takes longer than the time limit for the function. The time limit is 60 seconds by default and you can change it by specifying duration parameter, like @spaces.GPU(duration=120). This parameter is supposed to be set as close as the maximum expected inference time of the function. Each user has only a few minutes worth of GPU quota per a few hours, and users cannot run the function when their remaining quota is smaller than the duration.
Also, we had a bug on our infra side a few weeks ago that raises the error unexpectedly from time to time. The bug should be fixed now, but there's a possibility that the bug persists.

For users' convenience, we might not be able to test the environments again and again.

Ah, it totally makes sense. Sorry, I hadn't thought of it. I'll add you to the ZeroGPU explorers org so that you can duplicate this Space privately and assign ZeroGPU yourself to test if the reconstruction feature can work with ZeroGPU. You can update this main Space once you found the fix and delete the private duplicate.

I just sent you an invitation to join the org.

I just sent you an invitation to join the org.

Thank you, let me shift to [app_recon.py]