Feature Request: Support Anthropic on Vertex models.

[simonbutt]

Currently Anthropic models on AWS Bedrock is supported but not Google Vertex.

"As a user of Inspect eval framework and also GCP, I'd like to use Anthropic models through Vertex AI in the Google Cloud project and region of my choice, rather than having to create a new billing process to use through the Anthropic API"

Details on all anthropic models:
https://docs.anthropic.com/en/docs/about-claude/models

[jjallaire-aisi]

I don't have access to Vertex to test, however I just made a change that should enable you to put together a PR for this. The Anthropic client already supports Bedrock, and I just made it possible to use an anthropic/bedrock/ model name prefix to opt into using this client: ca04ef5

Anthropic also has a built-in Vertex client, so you could add a vertex flag (and service name check) that binds to Vertex rather than Bedrock or the default Anthropic endpoint.

[lizaitang]

Currently Anthropic models on AWS Bedrock is supported but not Google Vertex.

"As a user of Inspect eval framework and also GCP, I'd like to use Anthropic models through Vertex AI in the Google Cloud project and region of my choice, rather than having to create a new billing process to use through the Anthropic API"

Details on all anthropic models: https://docs.anthropic.com/en/docs/about-claude/models

Hi, Simon, have you solved this issue?

[jjallaire]

I have attempted to implement this here: #1122

There are docs on how to make this work in the PR notes. Unfortunately when I try to make this work on my own account I get 429 errors (hard to know how to overcome these, billing and other config looks fine).

If others could install this branch and try it out I'd be appreciative.

[jjallaire]

I have merged this as it appears to all work correctly (save for the fact that all of my requests are rate limited and attempting to request an increase in rate limits for Anthropic models on Vertex results in an immediate "can't increase your quota above 0 at this time"). It may just be difficult to get Anthropic API resources provisioned through Vertex right now (I'm sure it depends on the customer though).

If people want to try this out on their own accounts here are the basic docs: #1122

You might also want to add --log-level=http so you can see the HTTP rate limits in the console.