Fixing SageAttention in ComfyUI Portable: missing Python headers, 40% speedup on 4-step workflow
DeltaWaffleSyrup · reddit · 2026-09-14
A Reddit user shares a tutorial for installing SageAttention and Triton in ComfyUI Portable, yielding a 40% speed boost on a Minimax H3 4-step turbo workflow.
The key pitfall and fix (diagnosed by feeding error logs to Claude):
- Symptom: SageAttention errors on startup and silently falls back to PyTorch attention — the "running on startup" log line only means it's attempting, not that it works
- Root cause: Triton's portable build uses tcc to JIT-compile a CUDA helper (cudautils.c), but the embedded Python lacks Python.h and the include/, libs/ folders, so compilation fails immediately
- Fix: download the matching regular CPython installer (e.g. 3.13 per cp313), copy its include\ and libs\ folders into pythonembeded, restart, and confirm the sage attention error line is gone
The post also links the full workflow resources: a video walkthrough, the fused-turbo-int8 model on Hugging Face, and workflow files on GitHub.
More from Multimodal
- PNN Creator Challenge Offers $5,000 Prize Pool for AI Anime News Segments — Promptmethus · 2026-09-14
- MiniMax H3 generates an impressive Sailor Jupiter entrance animation — Certain_Potato_4509 · 2026-09-14
- MiniMax H3 video generation free for a week on Astorie, plus ad contest — aziz4ai · 2026-09-14
- 4 AI models colorize the same Berserk spread — GPT keeps linework best, Nano Banana adds Griffith — AKAMA199 · 2026-09-14
- Seedance 2.5 turns a school day into a stealth-game mission video — iamfakhrealam · 2026-09-14
- Which models natively generate true 2K/4K images? Reddit asks for recommendations — bananandsmoothie · 2026-09-14