Comfy-Org / Comfy-Org/ComfyUI

Regression: Qwen Image Edit FP8 + LoRA VRAM overflow after comfy-kitchen update

Open
#15,013 1 comment 1 reaction 0 assignees View on GitHub
Potential Bug
Dominant language
Python
Stars
133k
Forks
15.7k
Avg merge
1d 7h
Merged PRs (30d)
158

Description

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

The same workflow should run normally after updating ComfyUI dependencies.

With Qwen Image Edit FP8 model and LoRA:
- normal VRAM usage
- normal sampling speed
- VRAM should be released after generation

This workflow was working correctly with comfy-kitchen==0.2.7.

### Actual Behavior

After updating comfy-kitchen to newer versions (tested 0.2.16 and 0.2.20), Qwen Image Edit FP8 with LoRA causes abnormal VRAM usage.

Symptoms:
- GPU VRAM reaches 99%
- Sampling becomes extremely slow
- The whole system becomes laggy, similar to running out of VRAM
- ComfyUI may become disconnected

The issue does not happen with the same workflow when using comfy-kitchen==0.2.7.

Tested:
OK:
comfy-kitchen==0.2.7

Problem:
comfy-kitchen==0.2.16
comfy-kitchen==0.2.20

### Steps to Reproduce

Environment:
- Windows 11
- RTX 5090 32GB
- PyTorch 2.7.0 + CUDA 12.8
- Python 3.11

Steps:
1. Load Qwen Image Edit FP8 model.
2. Add any LoRA.
3. Run the same image edit workflow.
4. Compare behavior between comfy-kitchen==0.2.7 and newer versions.

The workflow and models are identical. Only comfy-kitchen version is changed.

### Debug Logs

```powershell
INFO] got prompt
0%| | 0/8 [00:00

Contributor guide

Open the contributing guide

Research direction

First reproduce the workflow with custom nodes disabled, then compare comfy-kitchen 0.2.7 with 0.2.16 or 0.2.20. Start at comfy/ldm/qwen_image/model.py and follow the sampling path through comfy/samplers.py; done means the Qwen Image Edit FP8 plus LoRA workflow completes with normal sampling speed, without VRAM overflow, and releases VRAM afterward.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.