Quantize Models to NVFP4 with NVIDIA Model Optimizer
60 MIN
Cut memory ~3.5× vs FP16 while keeping accuracy close to FP8, then validate with an OpenAI-compatible endpoint
Common issues
The Hardware platform column shows where an issue is most relevant. "All hardware platforms" applies to every supported platform.
| Symptom | Hardware platform | Cause | Fix |
|---|---|---|---|
| "permission denied" when running docker | All hardware platforms | User not in docker group | Run sudo usermod -aG docker $USER && newgrp docker |
| "Permission denied" when accessing Hugging Face | All hardware platforms | Missing or invalid HF token | Export a valid HF_TOKEN, or run huggingface-cli login |
| Cannot access gated repo for URL | All hardware platforms | Restricted Hugging Face model | Regenerate your Hugging Face token and request access to the gated model |
| Container exits with CUDA out of memory | All hardware platforms | Insufficient GPU memory | Quantize a smaller model, or free GPU memory and retry |
| Model files not found in output directory | All hardware platforms | Volume mount failed or wrong path | Verify ./output_models resolves from your working directory and remount |
| Git clone fails inside container | All hardware platforms | Network connectivity issues | Check internet access and retry |
| Quantization process hangs | All hardware platforms | Container resource limits | Increase Docker memory limits or keep --ulimit flags from Instructions |
Log ends with MPI or ModuleNotFoundError: No module named 'mpi4py' | DGX Station | Optional MPI runner step after quant | Confirm NVFP4 artifacts exist under ./output_models; quantization may have succeeded even if the final runner step fails |
| Model runs on wrong GPU | DGX Station | Dual-GPU default selection | Set GPU_ID to the GB300 device from nvidia-smi and use --gpus "device=$GPU_ID" |
| Memory pressure within capacity | DGX Spark | UMA buffer cache not released | See UMA note below |
Permission denied removing ./output_models | All hardware platforms | Root-owned files from container | Use sudo rm -rf ./output_models |
NOTE
Unified memory (UMA). On hardware platforms with unified memory, GPU and CPU share memory dynamically. Some applications have not yet been updated for UMA, so you may hit memory issues even within capacity. If that happens, manually flush the buffer cache:
sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'