Yukari

Yukari is a pull-through cache for Ollama registries. The Ollama registry is somewhat a Docker registry, but also somewhat not. It's just compatible enough with the Docker registry that you can use one as storage for Ollama models, but incompatible with pull-through caching. This project offers a simple pull-through cache that you can deploy to your networks to speed up pulling models.

As a side effect, this also makes your models resistant to "left-pad" style attacks where the models you rely on are no longer available. This stores models in Tigris, but theoretically can be extended to support any S3 compatible object storage system (S3, Ceph, etc).

Deploying

First, follow the Kubernetes quickstart and put the Tigris credentials into a secret named yukari-tigris-creds:

# yukari-tigris-creds.yaml
apiVersion: v1
kind: Secret
metadata:
  name: yukari-tigris-creds
type: Opaque
stringData:
  AWS_ACCESS_KEY_ID: tid_*
  AWS_SECRET_ACCESS_KEY: tsec_*
  AWS_ENDPOINT_URL_S3: https://fly.storage.tigris.dev
  AWS_ENDPOINT_URL_IAM: https://fly.iam.storage.tigris.dev
  AWS_REGION: auto
  TIGRIS_BUCKET: mybucket

Deploy this to your Kubernetes cluster by using the manifests in manifest, be sure to edit the following fields:

• The DNS hostnames in manifest/ingress.yaml
• Any configuration in manifest/deployment.yaml's env section

A Helm manifest is in the works.

Use the cache

This is quite easy, just prepend your.yukari.instance/library/ to the image you want to run/pull

This ollama pull <image>:<tag> becomes

ollama pull your.yukari.instance/library/<image>:<tag>

Architecture

This proxy will forward all uncached requests to the upstream Ollama registry. When it sees you fetching a manifest, it'll scrape that manifest for the component layers and start caching them in Tigris. All subsequent fetches will be from Tigris instead of the Ollama registry.

Every half an hour, Yukari will check if any manifests it has cached are more than 240 hours (10 days) old. If it finds any, it schedules reprocessing of those manifests. Any new model versions will automatically be put into Tigris, making things faster.

Configuration options (via environment variables)

Environment Variable	Description	Default
BIND	The TCP host:port to bind on when serving HTTP.	:9200 (port 9200 on all addresses)
INVALIDATOR_PERIOD	How often the cache invalidator logic runs.	30m (30 minutes)
MANIFEST_LIFETIME	How long a manifest can live before it is considered invalid.	240h (240 hours, or 10 days)
SLOG_LEVEL	The log level for slog.	ERROR
TIGRIS_BUCKET	The Tigris bucket to cache model information in.	yukari (you will need to change this)
UPSTREAM_REGISTRY	The upstream Ollama registry you are mirroring.	https://registry.ollama.ai/

Contributing

Feel free to create issues and PRs. The project is tiny as of now, so no dedicated guidelines.

Disclaimer: This is a side project. Don't expect any fast responses on anything.

Related Information

Yukari is a fork of simonfrey/ollama-registry-pull-through-proxy, but there has been an almost complete rewrite during the process of making it use Tigris as a storage backend.

• It is a fix to ollama/ollama#914 (comment)
• To make its behavior work better, we would need this PR merged: ollama/ollama#5241

License

MIT