trouble running newbie-image-exp0.1 and/or installing flash-attention

[manassm]

I've been trying to install flash attention in my comfy-desktop (.exe) environment to run newbie-image-exp0.1 but kept failing.
Below is an AI-summary of my current situation:

[Bug] ImportError: DLL load failed for flash_attn_2_cuda breaking core nodes (ComfyUI Desktop App)
System Information

ComfyUI Version: 0.4.0 (Desktop App)

OS: Windows 11

GPU: NVIDIA GeForce RTX 4080 (16GB VRAM)

Python Version: 3.12.10

PyTorch Version: 2.8.0+cu129

Python Executable: C:\Users\SEPHiA\Documents\ComfyUI\.venv\Scripts\python.exe

Describe the bug

After attempting to set up requirements for a new model, I am encountering import failures for core ComfyUI nodes (nodes_canny.py and nodes_morphology.py). The root cause appears to be an ImportError in the flash_attn dependency which is triggered when kornia is imported by these nodes.
Error Log
Plaintext

Traceback (most recent call last):
...
File "C:\Users\SEPHiA\Documents\ComfyUI.venv\Lib\site-packages\kornia\feature\lightglue.py", line 48, in
from flash_attn.modules.mha import FlashCrossAttention
File "C:\Users\SEPHiA\Documents\ComfyUI.venv\Lib\site-packages\flash_attn_init_.py", line 3, in
from flash_attn.flash_attn_interface import (
File "C:\Users\SEPHiA\Documents\ComfyUI.venv\Lib\site-packages\flash_attn\flash_attn_interface.py", line 15, in
import flash_attn_2_cuda as flash_attn_gpu
ImportError: DLL load failed while importing flash_attn_2_cuda: 지정된 프로시저를 찾을 수 없습니다. (The specified procedure could not be found.)

IMPORT FAILED: nodes_canny.py
IMPORT FAILED: nodes_morphology.py

Additional Context

I am using the ComfyUI Desktop App, not the Portable version.

The environment is the internal .venv created by the Desktop installer.

My system has CUDA 12.1 installed globally, but the ComfyUI Desktop log shows 2.8.0+cu129.

It seems kornia has a hard dependency chain leading to flash_attn, and because the flash_attn binary is incompatible with the current Desktop App environment, it prevents the standard comfy_extras nodes from loading.

┆Issue is synchronized with this Notion page by Unito

[dosubot]

Your ImportError with flash_attn_2_cuda is almost certainly due to a binary incompatibility between flash-attn, PyTorch, and CUDA in the ComfyUI Desktop App’s managed .venv. The Desktop App installs PyTorch 2.8.0+cu129 (CUDA 12.9) inside its own environment, which may not match your system’s global CUDA 12.1, and flash-attn binaries are extremely sensitive to these version mismatches. DLL load failures like this are a classic sign of such incompatibility, especially with packages that have compiled CUDA extensions reference.

Key points and steps to resolve:

• The Desktop App’s requirements do not list flash-attn directly; it’s pulled in transitively (often via kornia). The environment expects all CUDA/PyTorch-dependent packages to match the installed PyTorch version (2.8.0+cu129) reference.
• If you manually installed flash-attn, it may have been built for a different CUDA version (e.g., 12.1), which will not work with the Desktop App’s PyTorch (CUDA 12.9). You need a flash-attn wheel built for PyTorch 2.8.0+cu129, Python 3.12, and Windows—these are often not available upstream.
• The Desktop App’s .venv is isolated; global CUDA installations do not affect it. All dependencies must be installed inside this .venv, and all CUDA/PyTorch versions must match reference.

What you can do:

1. Delete the .venv folder in your ComfyUI install directory to force a clean environment rebuild.
2. Restart the Desktop App and let it recreate the environment and install dependencies.
3. If you need to install additional dependencies (like flash-attn for a custom node), ensure you are using a wheel built for PyTorch 2.8.0+cu129, CUDA 12.9, Python 3.12, and Windows. If such a wheel does not exist, flash-attn will not work in this environment.
4. If you cannot get a compatible flash-attn, consider disabling the custom node or workflow that requires it, or try running ComfyUI in a manual/portable install where you have more control over the environment and can match all versions exactly.

If you continue to see import errors, try disabling all custom nodes to isolate the issue, and only re-enable them one by one after confirming the core app works reference.

If you need to manually install PyTorch or other dependencies, always use the exact CUDA version index used by the Desktop App (for CUDA 12.9: pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129) reference.

Let me know if you need step-by-step commands for any of these actions.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}