AI BriefingSGLang Achieves Up to 6.24x Speedup in MiniMax-H3 Video Generation Using Cache-DiT and SubBlock
Key point
The SGLang Diffusion team accelerated the MiniMax-H3 video generation model by up to 6.24x on eight NVIDIA H200 GPUs using Cache-DiT and SubBlock sparse attention.
Details
The SGLang Diffusion team benchmarked the performance of the MiniMax-H3 video generation model using eight NVIDIA H200 GPUs. Compared to the Diffusers library as a baseline, the lossless SGLang runtime demonstrated a speedup of 1.85 to 1.95x.
Acceleration Strategies and Performance
For additional lossy acceleration, Cache-DiT (step reuse) and SubBlock Sparse Attention were combined. This combination achieved a maximum speedup of 6.24x, with SSIM (Structural Similarity Index) scores ranging from 0.76 to 0.91.
- Cache-DiT alone: Up to 2.99x speedup without quality degradation
- SubBlock 0.80 + Cache-DiT: 5.06 to 5.72x speedup for T2VA (Text-to-Video-Audio) and 5.86 to 6.24x speedup for FL2VA (First-Frame-to-Video-Audio)
- Quality trade-off: FL2VA tasks maintained high quality with SSIM scores of 0.85 to 0.91, while T2VA tasks saw a slight decline to SSIM scores of 0.76 to 0.78.
Technical Contributions
The acceleration stems from three key factors. First, Fused Kernels provide efficiency gains of 2.00 to 12.16x at individual operation sites. Second, Cache-DiT reduces computational load by skipping redundant denoising steps. Finally, SubBlock Sparse Attention lowers attention computation costs even during executed steps. This benchmark did not include other acceleration techniques such as quantization or progressive resolution adjustment.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.