[Windows] Crash “Windows fatal exception: access violation” when loading FLUX checkpoints (fp8 / nf4) — RTX 4060 Laptop 8 GB #9505
Description
Custom Node Testing
- I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)
Your question
Environment
- OS: Windows 11
- Machine: Laptop
- CPU: AMD Ryzen 7 (mobile)
- GPU: NVIDIA GeForce RTX 4060 Laptop GPU, 8 GB VRAM
- System RAM: 16 GB
- ComfyUI install: ComfyUI_windows_portable (official ZIP)
- Python: 3.13.6 (embedded)
- PyTorch: 2.8.0+cu129 (as printed at startup)
- ComfyUI: 0.3.51 — frontend 1.25.9
- Launch command: run_nvidia_gpu.bat (sometimes with --lowram, --force-fp16)
- Debug/diagnostic env vars tried:
1.Model paths: via extra_model_paths.yaml (also pointing to A1111 folders: checkpoints, VAE, LoRA, etc.)
2. ComfyUI-Manager: installed (clone in ComfyUI\custom_nodes\ComfyUI-Manager). Issue reproduced both with Manager present and with custom_nodes temporarily renamed (i.e., disabled).
FLUX models tested
- Comfy-Org/flux1-dev fp8 (flux1-dev-fp8.safetensors)
- Comfy-Org/flux1-schnell fp8 (flux1-schnell-fp8.safetensors)
- ../flux1-dev-bnb-nf4-v2.safetensors
FYI: SD 1.5 checkpoints (e.g. v1-5-pruned-emaonly.safetensors) run fine on this machine in ComfyUI.
Minimal Repro Steps
- Start ComfyUI (portable) on GPU (startup log shows Device: cuda:0 NVIDIA GeForce RTX 4060 Laptop GPU : cudaMallocAsync). Following the run_nvidia_gpu.bat
-
Load the FLUX dev checkpoint example workflow (flux_schnell)
-
In Load Checkpoint, pick a FLUX checkpoint (fp8 dev/schnell or the nf4 one).
- Click Execute.
Observed result
After few seconds server crash with “Windows fatal exception: access violation”. Process exit code: -1073741819 (0xC0000005).
----------------
What I already tried
- fp8 (dev & schnell) → crash
- nf4 (flux1-dev-bnb-nf4-v2) → crash
- Flags --lowram, --force-fp16 → still crash
Expected behavior
FLUX checkpoints should load on an RTX 4060 (8 GB) (even if we must use --lowram)
Thanks a lot for any guidance