n8n Workflow: Check Hugging Face Model VRAM Fit (Ollama)

Instant VRAM fit check for Hugging Face models—before you pull

This n8n workflow automatically checks whether a Hugging Face model is likely to fit on your GPU VRAM budget by combining Hugging Face model-card data with a local Ollama verdict for recommended quantizations—so you can avoid failed downloads and wasted time on a 12GB-class setup.

What this workflow does

• You POST a Hugging Face model id (and optional vram_gb budget) to the webhook endpoint /webhook/vram-fit-check.
• n8n fetches the model’s public Hugging Face model card (no API key required).
• A local Ollama model writes a short verdict that includes:
  • likely quantization fit (e.g., FP16 / Q8 / Q4)
  • license/gate considerations
  • the next action to take
• The workflow returns a structured JSON response containing the briefing.
• If a model id is provided, it also uses LangChain via a local Ollama call to generate a <150-word plain-text verdict about whether the model fits your VRAM and at which quantizations.
• If something goes wrong, it formats a helpful `error/fallback` message into the JSON payload.

Use cases

• Before running ollama pull, confirm your model can fit on a 12GB GPU (or your chosen budget).
• Screen Hugging Face Hub repos privately and quickly, without exposing tokens.
• Automate preflight checks for model downloads in a self-hosted ML/SaaS workflow.

Technical details

• Outbound HTTPS to huggingface.co (no token required).
• Self-hosted n8n + Ollama reachable by n8n (e.g., http://host.docker.internal:11434 for Docker).
• Webhook endpoint: /webhook/vram-fit-check.
• Nodes/logic: webhook, http request, code, if, error trigger, and sticky note.

Default local Ollama tag: gemma4:e4b (pull it locally or update the model tag to match what you have).