FLUX.1-dev FP8 extremely slow loading and heavy system lag on RTX 3060 (12GB)

[Aivoice96]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Hi,

I’m experiencing extremely slow loading and severe system lag when using FLUX.1-dev FP8 in ComfyUI.

System:

• GPU: RTX 3060 (12GB VRAM)
• RAM: [16 GB]
• OS: Windows
• ComfyUI: latest version (portable)
• Launch: run_nvidia_gpu.bat
• Browser: Firefox / Chrome (same behavior)

Description:

• FLUX.1-dev FP8 does work and completes successfully, but:
  • It takes 10–20 minutes before the KSampler actually starts
  • VRAM fills up (~9.5GB)
  • ComfyUI switches to partial loading / offloading / lowvram patches
  • Windows becomes very laggy (mouse stutter, UI freeze)
• Generation itself eventually finishes, but the loading phase is extremely slow.

Observations:

• This only happens with FLUX
• SDXL works normally on the same system (even with complex workflows)
• Task Manager shows very high memory usage and swapping behavior
• Logs show partial loading and offloading to CPU

Relevant log lines:

• model_type FLUX
• Loaded partially
• offloaded device: cpu
• lowvram patches

Question:
Is this expected behavior for FLUX on 12GB GPUs, or is there any recommended configuration to reduce the extreme loading time and system lag?
Is this mainly caused by RAM / pagefile swapping?

Thanks for your help.

Logs

Can provide full logs if needed.

Other

No response

[godlatero]

I ran the standard Flux.1 Dev workflow on version CUI 0.12.3 and the 3060 12GB without any changes. It works without any issues, and run about 1 min after maybe 5 min load all models.

But I have 112 GB RAM.
Windows uses more than 15 GB RAM, and during peak models load, memory usage is over 61 GB.
Most likely, you're using a paging file on disk, which is slowing down the entire process. Check this using Task Manager.
If page file on the main C:\ drive - it can slowdown all sysytem more.
All you need is more memory.

[Aivoice96]

up

[Arvamer]

It might be #11845, or at least reverting to commit before the PR was merged fixed the issue for me. 64 GB RAM, 32 GB VRAM. Moreover, even though RAM usage is only 50%, the swap usage is over 10 GB. Perhaps --disable-mmap flag no longer works?

Edit: I’m talking about flux2dev, didn’t notice original report was about flux1.

[github-actions]

This issue is being marked stale because it has not had any activity for 30 days. Reply below within 7 days if your issue still isn't solved, and it will be left open. Otherwise, the issue will be closed automatically.