MODAL_GUIDE.md

Modal Guide

Cloud Parade Cabinet can use Modal in two ways.

1. Cheap Smoke Test

This confirms your Modal account/profile works without starting a GPU:

$env:PYTHONUTF8="1"
$env:PYTHONIOENCODING="utf-8"
modal run cloud_parade_cabinet/modal_smoke.py

Expected output includes:

status: ok
project: Cloud Parade Cabinet

2. OpenBMB GPU Endpoint

The GPU scaffold serves openbmb/MiniCPM4-8B with a vLLM OpenAI-compatible server on Modal.

Create a Modal secret once:

modal secret create huggingface-secret HF_TOKEN=$env:HF_TOKEN

Deploy:

modal deploy cloud_parade_cabinet/modal_openbmb.py

Then point the app at the Modal web endpoint once Modal prints it:

$env:PARADE_MODEL="openbmb/MiniCPM4-8B"

Cost Notes

• The smoke test should be negligible.
• The OpenBMB endpoint uses A10G by default to keep spend controlled.
• Upgrade to A100 or H100 only if latency matters for a demo recording.
• Stop deployed GPU apps when not testing:

modal app stop cloud-parade-openbmb