Keep model-specific conversation formatting and special-token choices in the host.
Prepared prompt input
LocalEndpoint accepts a prepared prompt. The host chooses system/user/assistant formatting, separators, roles, and turn history.
Template source
Prefer template information from a trusted model source or reviewed GGUF metadata, then validate it against observed model behavior.
Special-token flags
Choose AddSpecialTokens, ParseSpecialTokens, RemoveSpecialTokens, UnparseSpecialTokens, and CleanSpaces deliberately; do not copy settings across models without evidence.
UI presentation
Show the active template policy, truncation decision, token budget, and any normalization so users can inspect what the host sent.
Available now
- The public UAIX.LmRuntime package family and canonical package documentation.
- The host-facing workflow and evidence distinctions described on this page.
Host responsibilities
- Select and trust inputs, define policy and limits, integrate the UI or service, retain application evidence, and test the exact deployment.
Not claimed
- Metadata does not guarantee template correctness.
- The runtime does not silently repair prompts.
Operational boundary
What this page covers: Keep model-specific conversation formatting and special-token choices in the host.
Available now: The documented package surfaces and this host-facing guidance are public.
Host responsibility: Own acquisition, trust, configuration, presentation, persistence, transport, policy, tests, and deployment.
Not claimed: Universal compatibility, production certification, safety certification, hidden fallbacks, or behavior beyond retained evidence.
Canonical sources: Canonical package documentation · Current release metadata
Last reviewed UTC:
Source and review record
Prerequisites
- Read Get Started
- Review host/runtime boundaries