0.00 MB usable + hangs at Model Initializing with Flux GGUF on 8GB NVIDIA GPU (lowvram + comfy-aimdo)
- 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
### Expected Behavior
When running ComfyUI v0.25 with `--lowvram` on a NVIDIA RTX 4070 Laptop GPU (8GB VRAM) using Flux GGUF models (quantized Flux + T5), the model should receive a reasonable VRAM allocation (>0 MB usable) and proceed to generate without hanging.
### Actual Behavior
The model loads as "loaded partially" with **0.00 MB usable**, **0.00 MB loaded**, and **~23 GB offloaded**. Generation then hangs indefinitely at "Model Initializing..." with no error message or crash. The progress bar stays at 0% for the entire sampling loop.
Log output:
```
[INFO] Requested to load Flux2
[INFO] 0 models unloaded.
[INFO] loaded partially; 0.00 MB usable, 0.00 MB loaded, 23374.75 MB offloaded, 2425.78 MB buffer reserved, lowvram patches: 0
Attempting to release mmap (299)
0%| | 0/8 [00:00
Contributor guide
Research direction
Start in model_management.py at MIN_WEIGHT_MEMORY_RATIO and the lowvram_model_memory calculation around line 935, then review how comfy-aimdo manages T5 GGUF weights through its virtual VRAM allocator. Reproduce the Flux workflow with --lowvram and --enable-dynamic-vram, and compare the allocation log. Done means the model receives nonzero usable VRAM and generation completes instead of hanging at Model Initializing.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Clearly specified
- Newbie friendliness
- 55/100