Make livekit-local-inference an optional dependency

[ytausch]

Feature Type

Would make my life easier

Feature Description

It would be nice if livekit-agents declared livekit-local-inference as an optional dependency, possibly behind a feature like livekit-agents[local-inference].

While VAD and turn detection are likely used by most LiveKit voice-agent projects, some consumers only use lower-level functionality, such as LLM abstractions, provider plugins, or telemetry, and do not need the bundled models.

Rationale: We use automated license checking in our organization, and projects that use livekit-agents must set manual exclusion rules even if they don't use local inference. Also, we don't want to whitelist the model license org-wide (like we do with OSS licenses), because individual projects should acknowledge compliance with it.

Another nice aspect would be smaller environments if local inference isn't actually used.

In addition to adjusting pyproject.toml, the implementation would need to use lazy imports and raise a clear runtime error.

Somewhat related: #6192

Workarounds / Alternatives

No response

Additional Context

No response

[KXHXK]

I reviewed current main and would be interested in implementing this once the compatibility shape is confirmed. This is broader than moving one dependency: livekit.agents eagerly imports inference.VAD, inference.eot.transports imports the native EOT, and the Linux forkserver unconditionally preloads inference._warmup.

A backward-compatible option would be to keep livekit-local-inference in the default install via a local-inference extra included by a standard/default installation path, while making a documented minimal core install possible. The code side would defer native imports until local VAD/EOT construction and skip forkserver warmup when the package is unavailable, with a clear ImportError pointing to livekit-agents[local-inference]. Tests would verify that importing core LLM/telemetry APIs works with the native module blocked, while local VAD/EOT fail only at use time with the install hint.

Would you prefer that compatibility-preserving shape, or should plain pip install livekit-agents become the minimal install immediately? I will hold the PR until a maintainer confirms the intended default, since that changes the package installation contract.