Connect Training and Inference Services
Definition
Open Admin → Integrations, select the provider and configure the connection for the intended account/resource. Use the resulting connection in Weights training or the consuming inference workflow. Connection validation is separate from permission, capacity and compatibility for a particular job.
Provider prerequisites
| Provider | Connection inputs | Separate job requirements |
|---|---|---|
| Azure OpenAI | Resource name, region, resource-origin endpoint and matching key | Eligible model, quota, fine-tuning and deployment rights |
| Microsoft Foundry | Resource name, region, resource-origin endpoint and matching key | Eligible model and training/deployment permission |
| Together | Key for the intended project | Model eligibility and job capacity |
| Fireworks | Account ID and matching key | Selected service/model eligibility and capacity |
| Baseten | Key permitted to list training projects | Training project, trainer image/command and data access |
| Amazon Bedrock | AWS account, region and integration role ARN | Customization model, service role and job storage access |
| Amazon SageMaker AI | AWS account, region and integration role ARN | Training execution role, compute and input/output storage |
| Vertex AI | Project ID, region and connection service-account email | Scoped impersonation and job execution identity/resource access |
For Azure OpenAI use the resource origin https://<resource>.openai.azure.com; for Foundry use https://<resource>.services.ai.azure.com. A project endpoint, deployment URL or API path is a different boundary. Match endpoint and key to the same resource.
Configure and verify
- Obtain the intended account/resource identifiers and credential or role details from its administrator.
- For key-based providers, enter values only in the owning connection flow and use Verify and save. Rotate using the existing connection's credential-replacement control.
- For AWS role connections, obtain the exact trusted backend AWS principal from your Teammately operator and require the displayed workspace external ID in role trust. An external ID is not a principal. Do not substitute a long-lived access key.
- For Vertex, obtain the exact backend impersonating principal from the operator, and authorize it only for the selected connection service account and resources. Keep connection identity and training execution identity distinct.
- Select the saved connection, compatible model/method and job-specific resources in Weights training.
A Baseten connection check lists accessible training projects. AWS and Vertex checks do not establish every training/storage permission. A successful connection never certifies quota, job creation or training quality.
Other inference endpoints
Choose Other inference endpoint for a public HTTPS API base, including an API prefix such as /v1. The connection supports Bearer token or Baseten Api-Key authentication. Local/private/IP endpoints and URLs with embedded credentials, ports, queries or fragments are rejected by this connection flow.
Verification checks model-list access. It does not test invocation or chat-completion compatibility and does not create a training connection. Check invocation in the consuming workflow against the actual deployment.
Fields, states, or lifecycle rules
Keep connection state, successful probe, job admission, job completion and callable inference deployment separate. Provider training does not automatically allocate serving resources. Inspect the actual job result and separately authorized evaluation before interpreting a candidate's quality.
When a connection fails, check matching account/region/endpoint and credential or role trust first. When a job fails after connection validation, inspect its execution role, selected model, storage and quota. Correct the owning provider connection; retain its intended identity/resource boundary.