Understanding How Ollama Stores Models

---

---

AI Summary

Summary: Olama Model Storage and Management

• Common Questions:
  • Location of Olama model storage is confusing to users.
  • Models are stored in different directories based on the operating system.
• Storage Locations:
  • Mac: ~/olama/models
  • Linux: /usr/share/olama/olama/models
  • Two main directories: blobs and manifests
• Manifest Directory:
  • Organized by registry/organization/model.
  • Currently, a single registry at ol. is used.
  • Official models are under Library.
• Model Directories:
  • Example: llama 2 model directory contains files for each tag pulled.
• Manifest Files:
  • Schema version set to two, not related to Olama but to Docker.
  • Media type: application/vnd.docker.distribution-manifest.v1+json.
  • Olama team has roots in Docker, influencing the manifest format.
• Manifest Contents:
  • config: Contains media type, digest, and size (not detailed).
  • layers: Array of config objects, each with media type, digest, and size.
  • First layer usually contains the model with a unique SHA-256 digest.
  • Digests serve as fingerprints and filenames in the blobs directory.
• Blobs Directory:
  • Contains large files not suitable for opening in standard code editors.
• License and Templates:
  • License files and templates also have unique SHA-256 digests.
  • Templates can be shared across different models.
• Model Management Commands:
  • olama cp: Copies a model.
  • olama create: Creates a new model.
  • olama rm: Deletes a model but retains blobs if still referenced.
• Blob File Management:
  • Blob files are kept as long as a model or manifest references them.
  • Deleting models may not free up space if blobs are shared.
  • Adding models may not consume more space if based on the same source model.
• Conclusion:
  • Explains model storage, naming, and deletion in Olama.
  • Encourages questions, likes, and subscriptions for more content.

[Music]
Goodbye
[Music]