Spaces:

---

fffiloni

/

PSHuman

like 192

Running on Zero

Hi @fffiloni ,

first of all, thank you for maintaining the PSHuman Space on Hugging Face.

I have been working on a Blackwell / ZeroGPU compatibility fix after the recent migration to NVIDIA RTX PRO 6000 Blackwell and the torch 2.11.0+cu130 stack.

I managed to get my PSHuman Space running again on Hugging Face ZeroGPU with the full pipeline working:

• Multiview generation works again
• SMPL body optimization runs through
• Reconstruction runs through
• MP4 video export via MoviePy completes successfully

The working setup uses:

• torch 2.11.0+cu130
• CUDA 13.0
• NVIDIA RTX PRO 6000 Blackwell / sm_120
• backend:cudaMallocAsync for the subprocess CUDA allocator
• custom cu130 wheels for pytorch3d, torch_scatter, nvdiffrast, kaolin, and xformers
• sys.executable instead of "python" for subprocess calls
• SDPA as a safe fallback for multiview attention
• removal of the deprecated verbose argument from ReduceLROnPlateau in reconstruct.py

The most important fixes were:

1. Move the whole stack from cu128 to cu130.
2. Use sys.executable for inference.py subprocess calls.
3. Set PYTORCH_CUDA_ALLOC_CONF=backend:cudaMallocAsync in the subprocess environment.
4. Replace the torch.bmm attention fallback with torch.nn.functional.scaled_dot_product_attention to avoid huge quadratic memory usage.
5. Patch the diffusers 0.31.0 Attention API incompatibility by using keyword arguments instead of positional arguments.
6. Remove verbose=0 from ReduceLROnPlateau in reconstruct.py because torch 2.11 no longer accepts it.
7. Use warp-lang >= 1.5.0 for kaolin 0.18.0.
8. Use a real CUDA-built nvdiffrast wheel, not a py3-none-any wheel without the compiled plugin.

I documented the full fix notes here:

https://huggingface.co/spaces/painter3000/PSHuman/blob/main/PSHUMAN_BLACKWELL_FIX_NOTES_ENG_US.md

My working Space is here:

https://huggingface.co/spaces/painter3000/PSHuman

If this is useful, feel free to use my notes or compare the changes with your Space to restore PSHuman on the current ZeroGPU Blackwell/cu130 environment.

Best regards,
Michael / Painter3000