Skip to main content

​
Overview

When you run cassian down, your workspace is saved to the cloud. Next cassian up restores it: fast, incremental, only downloading what’s new or changed.

​
What persists

PathPersists?Mechanism
/workspace (your code, scripts)YesSynced to cloud on down, restored on up
/workspace/storageYesAlways available, persists across sessions
no_sync foldersYesCloud only, doesn’t copy locally
Installed pip packagesNoLost on down
~/.cache/huggingfaceNoLost on down, use /workspace/storage
System packages (apt)NoPart of base image

​
Keeping model weights persistent

HuggingFace models and large datasets should live in /workspace/storage. Downloads happen once and stay available across all your sessions. 
# In your container, point HF to persistent storage
export HF_HOME=/workspace/storage/huggingface
Or in cassian.yaml: 
workspace:
  setup: "export HF_HOME=/workspace/storage/huggingface && pip install -r requirements.txt"

​
Keeping generated files (checkpoints, logs)

Use no_sync for large generated outputs you want to keep but don’t need locally: 
workspace:
  no_sync:
    - "checkpoints/"
    - "outputs/"
    - "wandb/"
These persist in the cloud across sessions but never download locally, keeping things fast.

​
Pip packages

Packages are re-installed on every cassian up. Keep a requirements.txt in your project root. Cassian detects it automatically and runs pip install as part of startup. Or specify explicitly: 
workspace:
  setup: "pip install -r requirements.txt"
Pre-installed in the default image: torch, transformers, accelerate, peft, datasets, bitsandbytes, trl, wandb, numpy, pandas, scipy, jupyter.
Package caching is coming soon. Your installed packages will be cached between sessions so you only pay the install cost once.

​
Switching GPU types

Your volumes are cloud-backed and node-agnostic. cassian switch lets you pick a different GPU type. Your entire workspace moves with you, no re-download.