Key Info

A developer asked Claude to fully implement Trellis.2 in CUDA C++, resulting in a 5.3x speedup over the original PyTorch implementation. The code is open-sourced and can be used directly or as a source for agent-based model optimization.

Highlights

  • 5.3x faster than the original PyTorch implementation, with zero Python runtime (all C++).
  • The GLB export function was the bottleneck, not the neural net; moving export to GPU gave the biggest gains.
  • Still delivered 2.2x better performance while being bitwise identical.
  • Cost: ~$7 on Claude and ~$60 on compute for testing on H200s via Hugging Face Jobs.