← sources

Tutorial For Installing Sage and Triton to ComfyUI Portable. Gave Me A 40% Speed Boost On Minimax H3 4-Step Workflow

source: reddit (reddit:StableDiffusion)published 2026-09-14 09:01ingested 2026-09-15 02:38

retained text

I know Sage and Triton are nothing new but if you're a beginner or doing a fresh re-install like me, these resources helped me get faster gens on the 4-step turbo workflow I use (will post a link to the workflow lower down in the post here.

The below YouTube video helped me get the main files installed:

[https://youtu.be/7ZDi2CwkS5U](https://youtu.be/7ZDi2CwkS5U)

  
However I still did not fully get Sage to work until I followed these additional steps from Claude AI after getting some error logs in my SLA workflow (link to workflow at bottom of this post):

  
***\~\~\~FROM CLAUDE, AFTER FEEDING IT MY ERROR LOGS:\~\~\~***

Diagnosing a compiler error blocking kernel compilation.

This one's a real (but non-fatal) problem — it's why you're silently falling back to slower attention.

**Root cause:** SageAttention needs Triton to JIT-compile a small CUDA helper (`cuda_utils.c`) at runtime. Triton's portable build uses `tcc` (Tiny C Compiler) to do this instead of MSVC. But your embedded Python distro (`python_embeded`) doesn't include the `Python.h` header or matching `libs/`, `include/` folders that a normal Python installation has — so the compile step fails immediately at `#include Python.h`.

That failure is caught, ComfyUI/Triton logs the error, and **silently falls back to PyTorch attention** — which is why your first message ("sage attention running on startup") was misleading: it's only *attempting* to use it, not confirming it works. Generations still complete because of the fallback, just without SageAttention's speed benefit.

**How to actually fix it — add the missing Python dev headers to your embedded Python:**

1. Check your embedded Python version — your paths show `cp313`, so Python 3.13.
2. Download the matching **regular (non-embeddable) CPython 3.13 Windows installer** from [python.org](http://python.org) (does not need to be installed system-wide — you just need the files).
3. From that install (or by installing it to a throwaway location), copy:
   * `include\` folder → into `C:\Ai\ComfyUI_Portable\3\ComfyUI_windows_portable\python_embeded\include\`
   * `libs\` folder → into `C:\Ai\ComfyUI_Portable\3\ComfyUI_windows_portable\python_embeded\libs\`
4. Restart ComfyUI and try a generation again — watch for the `[ERROR] Error running sage attention` line; it should disappear.



***\~\~\~END OF CLAUDE OUTPUT\~\~\~***



I had to ask Claude to break some of the steps down, but I basically just had to do the first 4.

  
**MINIMAX H3 4-STEP WORKFLOW RESOURCES:**

**A walkthrough of the workflow is here (MAKE SURE YOU WATCH IN FULL):** [**https://www.youtube.com/watch?v=4AqLGMkJwUQ**](https://www.youtube.com/watch?v=4AqLGMkJwUQ)

**The main model used is here:** [**https://huggingface.co/MATLOWAI/minimax-h3-fused-turbo-int8-convrot**](https://huggingface.co/MATLOWAI/minimax-h3-fused-turbo-int8-convrot)

  
**A link to the workflow is here (the same model is used in all workflows):** [**https://github.com/amao2001/ganloss-latent-space/tree/main/workflow/2026-09-07%20minnimax%20fused%20turbo**](https://github.com/amao2001/ganloss-latent-space/tree/main/workflow/2026-09-07%20minnimax%20fused%20turbo)





  

document fetched 2026-09-15 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1wfy61m/tutorial_for_installing_sage_and_triton_to