If you’re running ComfyUI on an NVIDIA RTX 5090 with PyTorch 2.9.1 and CUDA 13.0, you may encounter startup crashes with cryptic error messages. This guide explains the causes and provides working solutions.
The Problem
After updating ComfyUI to version 0.8.2 on Windows with an RTX 5090, the application crashes on startup. You might see one of these errors:
- Python process exited with code 3221226505 (0xC0000409)
- “CUDA error: no kernel image is available for execution on the device”
- bitsandbytes library load error: Configured CUDA binary not found at libbitsandbytes_cuda130.dll
The RTX 5090 uses NVIDIA’s new Blackwell architecture (sm_120), and several popular Python packages haven’t released compatible CUDA binaries yet.
Root Causes
Two packages are the main culprits:
1. bitsandbytes – This library is used for quantization and memory optimization. As of early 2025, it only includes precompiled DLLs up to CUDA 12.9. When PyTorch runs with CUDA 13.0, bitsandbytes fails to find a matching binary and crashes.
2. triton-windows – This package provides optimized CUDA kernels but hasn’t been compiled for the Blackwell architecture yet. When ComfyUI tries to load nodes that use Triton, it triggers a “no kernel image available” error.
The Solution
Step 1: Disable bitsandbytes
Navigate to your ComfyUI virtual environment’s site-packages folder and rename the bitsandbytes directory:
ren "C:\Users\[YourUsername]\Documents\ComfyUI\.venv\Lib\site-packages\bitsandbytes" bitsandbytes.disabled
This prevents the library from loading while preserving it for future use when CUDA 13.0 support is added.
Step 2: Uninstall triton-windows
Remove triton-windows using pip:
C:\Users\[YourUsername]\Documents\ComfyUI\.venv\Scripts\pip.exe uninstall triton-windows
Step 3: Install the Frontend Package
If you see an error about comfyui-frontend-package, install it manually:
C:\Users\[YourUsername]\Documents\ComfyUI\.venv\Scripts\pip.exe install comfyui-frontend-package
ComfyUI Desktop App Workaround
The ComfyUI Desktop application may still crash even after these fixes, as it might have its own startup routines that conflict with CUDA 13.0. The solution is to launch ComfyUI manually using a batch file.
Create a file named ComfyUI.bat on your desktop with the following content:
@echo off
cd /d C:\Users\[YourUsername]\Documents\ComfyUI
call .venv\Scripts\activate
python F:\ComfyUI\resources\ComfyUI\main.py --base-directory "C:\Users\[YourUsername]\Documents\ComfyUI" --auto-launch
Adjust the paths to match your installation. The --base-directory parameter ensures ComfyUI finds your custom nodes, models, and other resources. The --auto-launch flag automatically opens your browser to the ComfyUI interface.
What You Lose
Disabling these packages means some features won’t work:
- NF4 quantization nodes (require bitsandbytes)
- Some memory optimization features
- Certain custom nodes that depend on Triton kernels
However, core ComfyUI functionality including Stable Diffusion, SDXL, Flux, and most workflows will work normally.
Looking Ahead
Both bitsandbytes and triton-windows will eventually release updates with RTX 5090 support. Check their respective GitHub repositories periodically for new releases. Once updated versions are available, you can:
- Rename bitsandbytes.disabled back to bitsandbytes
- Reinstall triton-windows with pip
System Specifications
This solution was tested on:
- Windows 11
- NVIDIA GeForce RTX 5090 (32GB VRAM)
- PyTorch 2.9.1+cu130
- CUDA 13.0
- ComfyUI 0.8.2
- Python 3.12.8
Conclusion
The RTX 5090 is a powerful GPU for AI image generation, but being on the bleeding edge means waiting for software to catch up. By disabling incompatible packages and using a manual launch script, you can run ComfyUI successfully while the ecosystem adapts to the new Blackwell architecture.
