TRELLIS.2 CUDA kernels on HF Kernels cut BVH build from 7.1s to 5ms
ariG23498 · x · 2026-11-19
Hugging Face's kernels project shipped two standalone CUDA C++ kernels from Microsoft's TRELLIS.2 image-to-3D pipeline (suvadityamuk/trellis2-kernels), moving serial CPU mesh post-processing steps to GPU:
- meshudf: point-to-mesh distance with nearest face and barycentric coordinates, building a GPU linear BVH per call—BVH build drops from 7.1s to 5ms on H200
- teleainpaint: GPU TELEA inpainting replacing cv2.inpaint's TELEA mode—a 4096² texture atlas goes from 8.2s to 0.15s
Usage is pip install -U kernels then getkernel("suvadityamuk/trellis2-kernels"); Metal and ROCm kernels are coming next.
More from Infra
- FT: Anthropic needs ~$602B yearly revenue to earn 20% on $1T compute commitments — rohanpaul_ai · 2026-11-19
- FT: Anthropic needs ~$602B yearly revenue to earn 20% on ~$1T compute commitments — rohanpaul_ai · 2026-11-19
- Linux Mint website goes dark, showing hosting placeholder pages — kingslyj · 2026-11-19
- Tencent Hunyuan's Hy4-Flash engine PR reveals tensor parallel, MTP drafts, CPU expert lane — TencentHunyuan · 2026-11-19
- Commentator calls Jensen Huang's China compute claim bullshit: won't outpace rest of world — teortaxesTex · 2026-11-19
- OpenCode fix cuts GPT-6.2 uncached input per step from ~14,000 tokens to 3 — victor_explore · 2026-11-19