lucasbrianpiveta/Hetu-DiT
🚀 Optimize your Diffusion Transformers with Hetu-DiT, a dynamic parallel serving system that reduces latency and enhances GPU utilization.
What's novel
🚀 Optimize your Diffusion Transformers with Hetu-DiT, a dynamic parallel serving system that reduces latency and enhances GPU utilization.
Code Analysis
9 files read · 3 roundsA distributed GPU inference serving system for Diffusion Transformers that implements multi-dimensional parallelism (data, CFG, sequence, tensor, pipeline) with ILP-based task scheduling and custom ring/ulysses attention kernels for models like Flux, CogVideoX, HunyuanVideo, and SD3.
Strengths
Genuine distributed systems engineering with real ILP optimization for GPU scheduling, proper multi-dimensional parallelism decomposition, and custom distributed attention implementations with KV cache support.
Weaknesses
README is completely misleading (describes a simple GUI app), no visible test suite, significant code duplication, and extremely large monolithic files (76KB scheduler, 114KB engine) that hurt maintainability.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
8
Contributors
1
Files
149
Active weeks
4
Repository
Language
Python
Stars
1
Forks
1
License
Apache-2.0