Do not make any extra API call when endpoint URL is directly specified

[coyotte508]

Fix #1321 (cc @Kakulukian for viz)

The fix is not perfect, a greater refactoring would need to take place (eg to remove all concept of providerConfig when endpointUrl is provided, use a separate default provider config maybe not relying on model being defined)

[Copilot]

Pull Request Overview

This PR updates the inference API request logic to avoid making unnecessary API calls when a direct endpoint URL is provided. Key changes include:

• Making the "model" property optional in BodyParams.
• Adding a fast-path in makeRequestOptions to bypass model resolution when an endpoint URL is supplied.
• Adjusting the signature and usage of makeRequestOptionsFromResolvedModel to accommodate the optional model input.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
packages/inference/src/types.ts	Changed "model" from required to optional in BodyParams.
packages/inference/src/lib/makeRequestOptions.ts	Added branch logic for endpoint URL and updated request options construction accordingly.

Comments suppressed due to low confidence (2)

packages/inference/src/types.ts:130

• Making the 'model' field optional may lead to cases where a required model is not provided when an endpoint URL is absent. Ensure that downstream logic either supplies a valid model ID or handles the undefined case appropriately.

model?: string;

packages/inference/src/lib/makeRequestOptions.ts:210

• When using the endpointUrl branch, 'input.resolvedModel' might be undefined. Consider adding validation or a fallback to guarantee that a model value is provided when needed by providerConfig.makeBody.

model: input.resolvedModel,

[Kakulukian]

ty

[Wauplin]

Thanks! Just a nit regarding the potentially undefined model.

a greater refactoring would need to take place (eg to remove all concept of providerConfig when endpointUrl is provided, use a separate default provider config maybe not relying on model being defined)

What we do on the Python side is to use the "hf-inference" provider when an endpoint is passed. And in the hf-inference logic we take care of handling things correctly. If a user passes provider="<any-provider-except-hf-inference>", endpoint="https://..." we raise an exception.

[Wauplin]

I'm a bit afraid this PR and #1315 will highly conflict 😕 cc @hanouticelina