DEV Community

Cover image for πŸ› οΈ I Built a One-Click ComfyUI Setup for RTX 5090 on Windows β€” No WSL2, No Docker
GeneLab_999
GeneLab_999

Posted on

πŸ› οΈ I Built a One-Click ComfyUI Setup for RTX 5090 on Windows β€” No WSL2, No Docker

#comfyui #ai #showdev #python

I bought an RTX 5090. 32GB VRAM. The most powerful consumer GPU on the planet.

Then I tried to run ComfyUI on Windows. It broke immediately.

RuntimeError: sm_120 is not compatible
Enter fullscreen mode Exit fullscreen mode

Three days later, I had a fully working solution. I packaged it and open-sourced it:

ComfyUI-Win-Blackwell

Here's the whole story.

Why RTX 50-series Breaks Everything

NVIDIA's Blackwell architecture (RTX 5090/5080/5070) uses a new compute capability code called sm_120. The problem? PyTorch's stable release doesn't include kernels for it.

This means:

  • pip install torch β†’ doesn't work on Blackwell
  • You need PyTorch nightly with CUDA 13.0 (cu130)
  • But then xformers (the standard ComfyUI speed boost) forces PyTorch back to stable
  • And custom nodes silently pull stable PyTorch through their dependencies

It's a dependency trap. Every fix creates a new problem.

The 5 Rules I Discovered

After 3 days of trial and error, I distilled everything into 5 rules. Break any one of them and your environment dies.

Rule 1: Use PyTorch nightly cu130 β€” stable doesn't have sm_120 kernels.

Rule 2: Never install xformers β€” it force-downgrades PyTorch to stable. This is the trap that got me twice.

Rule 3: Strip torch from every requirements.txt β€” custom nodes list torch as a dependency, and pip will happily replace your nightly build with stable.

Rule 4: Verify PyTorch after every custom node install β€” run python -c "import torch; print(torch.__version__)" and check that it still says cu130.

Rule 5: Clear proxy environment variables β€” system proxies block pip and git silently on Windows.

πŸ’‘ Pro Tip: Rule 2 was the hardest to figure out. xformers installs successfully, ComfyUI starts fine, and then crashes mid-inference with sm_120 is not compatible. You don't even realize PyTorch was downgraded until you check the version.

What I Built

I automated all 5 rules into a one-click setup:

git clone https://github.com/hiroki-abe-58/ComfyUI-Win-Blackwell.git
cd ComfyUI-Win-Blackwell
# Double-click setup.bat β€” done in ~20 minutes
Enter fullscreen mode Exit fullscreen mode

What setup.bat handles:

  • Python 3.13 environment
  • PyTorch nightly cu130 (not stable, not cu128)
  • triton-windows + torch.compile (replaces xformers)
  • ComfyUI core + custom dependencies (with torch stripped out)
  • 28 verified custom nodes
  • Post-install verification that cu130 is still intact

I also built companion tools:

  • verify_env.py β€” Blackwell-specific environment checker (sm_120, cu130, Triton, torch.compile)
  • fix_windows_compat.py β€” Converts Linux workflow JSON paths to Windows format, replaces SageAttention with SDPA
  • update.bat β€” Updates everything while preserving Blackwell compatibility

What I Verified

I tested 28 custom nodes one by one. Install β†’ check PyTorch version β†’ run test β†’ record result. That was the most tedious part.

I also tested 5 Image-to-Video pipelines on 32GB VRAM:

  • HunyuanVideo 1.5 I2V (8.3B params, ~16GB) β€” Smooth. My top recommendation.
  • Kandinsky 5.0 Lite I2V (2B, ~4GB) β€” Very smooth. Great for quick tests.
  • LTX-2 I2V (19B, ~25GB) β€” Works in FP8. Tight but fine.
  • LongCat-Video TI2V (13.6B, ~14.5GB) β€” Works with adjustments.
  • Kandinsky 5.0 Pro I2V (19B, ~40GB) β€” Needs CPU offload. Slow.

Why Not Just Use WSL2 or Docker?

The short answer: performance.

Loading safetensors through WSL2's NTFS translation layer is noticeably slower. Docker has the same issue plus additional setup complexity. For a tool like ComfyUI where you're iterating on workflows and loading large models frequently, native Windows file I/O makes a real difference.

Also, most AI artists using ComfyUI on Windows aren't Docker experts. A .bat file they can double-click is the right UX.

Wrapping Up

If you have an RTX 5090/5080/5070 and want to run ComfyUI on Windows without WSL2 or Docker, give it a try:

github.com/hiroki-abe-58/ComfyUI-Win-Blackwell

GitHub logo hiroki-abe-58 / ComfyUI-Win-Blackwell

ComfyUI for GeForce RTX 50-Series (Blackwell)

The first fully documented, Windows-native ComfyUI setup for NVIDIA GeForce RTX 5090/5080/5070 (Blackwell architecture, sm_120) with CUDA 13.0.

Other languages: ζ—₯本θͺž | δΈ­ζ–‡ | ν•œκ΅­μ–΄

What Makes This Special

RTX 50-series GPUs (Blackwell, Compute Capability sm_120) are not supported by PyTorch stable releases as of early 2026. Running ComfyUI on these GPUs requires specific versions and workarounds that are not documented anywhere else in a single, reproducible package.

Technical Highlights










































Feature Details
GPU Architecture NVIDIA Blackwell (sm_120) -- RTX 5090 / 5080 / 5070
CUDA Version 13.0 (cu130) -- the latest CUDA runtime
PyTorch Nightly cu130 build (not stable, not cu128)
Python 3.13 (latest)
Triton triton-windows fork (official Triton is Linux-only)
xformers Deliberately excluded (causes PyTorch downgrade)
Custom Nodes 28 verified nodes including video & music generation
Platform Windows Native (no WSL2, no Docker required)

Why This Is Unique

  1. Blackwell + Windows Native +…



View on GitHub


MIT licensed. Stars and PRs welcome β€” especially if you verify additional custom nodes on Blackwell hardware.

Have you tried running AI tools on RTX 50-series? What was your experience? Let me know in the comments! πŸ‘‡

If you found this helpful, consider following me for more AI + GPU content!

πŸ“ Japanese version: Qiita / Zenn
🐦 Follow me on X: @geneLab_999
πŸ’» GitHub: hiroki-abe-58

Top comments (0)