Model Repository

Use a Model Repository to keep the weight and configuration files of a custom or fine-tuned model in one place. You can store, version, and share the model there. You can also create a Model Endpoint that serves it through a REST API or gRPC. A repository stores model files, and a Model Endpoint is the running service that answers requests using files from a repository. A repository is useful on its own as well.

Each repository is backed by a bucket in E2E Object Storage (EOS). You can work with it as you would with a cloud file system. A model in TIR is a directory in the bucket. A repository takes any file, in any structure, file format, or framework.

You move files in and out through the portal, the MinIO CLI (mc), s3cmd, or the Python SDK, rather than mounting the repository like a disk. A Model Endpoint downloads from the repository when it starts. A repository grows with its files, so its size always matches what you store.

When you create a repository, you choose where its bucket comes from. It can be a new bucket that TIR creates, one of your own EOS buckets, or an EOS bucket that another account owns.

Where to find it

In TIR, open your project. In the side menu, under Inference, select Model Repository. You see the repositories of the current project and region. The dashboard also has a Create Model Repo shortcut.

What you can do

Create and connect

• Create a repository with a name, a model type, and a bucket.
• Use an existing or external bucket for a repository.
• Find a repository’s access keys and endpoint to connect an S3 tool or the SDK.
• Connect to a repository with the MinIO CLI or s3cmd using the ready-made commands on the Using CLI tab.

Work with model files

• Upload model files with the portal, the MinIO CLI, s3cmd, or the Python SDK.
• Browse, download, and delete files on the Model Files tab.
• Use a repository from an instance or notebook to download a model or save a trained one.

Serve and train

• Use a repository from a Model Endpoint so the endpoint downloads the model when it starts.
• Store fine-tuned model output and continue training from a checkpoint.

Delete

• Delete a repository when you are finished with it.

Use the API

The TIR API reference lists the repository operations in Model Repository.

Where to start

Start with Get started with Model Repository. It creates a repository and uploads a first model. For help with a message, see Troubleshooting Model Repository. The Model Repository FAQ has short answers about file formats, versions, access, and billing.