Ben Browning

Testing the arm64 built starter distribution image on a mac with podman desktop, I'm able to pull the image and run the image by giving it a custom entrypoint. However,...

Thanks for the pointer! We may need to rethink the default vector store provider in the starter distribution so that it also works for Mac users...

I had the same reservations as Ashwin, but discussed this with some of the contributors here offline and think it's a reasonable compromise to give a place for new APIs...

I wonder if we should set the `UV_PYTHON` environment variable ourselves to the currently running Python installation before calling `uv venv` at https://github.com/meta-llama/llama-stack/blob/37b6da37ba3f62a9267f6242997ee17a4f070b1a/llama_stack/distribution/build_venv.sh#L86 if that env variable is not already...

From my local testing of this PR, to get this working with more models we'll probably need to explicitly tell the models to only return one tool call per turn,...

The following change gets this test consistently passing for me with Llama 4 Scout using the fireworks provider, although it fails when using the together provider due to some tool...

(comment made on wrong GitHub issue, so removed)

We have both an inline and a remote vLLM provider now. Can we close this one as done?

What provider was in use here? For example, the remote vLLM provider doesn't actually pass `tool_choice` or `tool_config.tool_choice` through to the backend vLLM. Other providers may suffer from the same...

I believe this is fixed by #1529 however that fix has not made it into a released Llama Stack yet. If you're willing and able to run with a distribution...