Struct MtmdContext 

pub struct MtmdContext { /* private fields */ }

Context for assembling multimodal turns. Holds the chat template and (eventually) the tokenizer; per-family runners hand the resulting AssembledTurn into their prefill path.

Implementations

impl MtmdContext

pub fn from_template_source(src: impl Into<String>) -> Self

Build a context from a Jinja chat template (typically loaded from a GGUF via crate::ChatTemplate::from_gguf).

pub fn with_tokens(self, bos: Option<String>, eos: Option<String>) -> Self

pub fn template_source(&self) -> &str

pub fn bos_token(&self) -> Option<&str>

pub fn eos_token(&self) -> Option<&str>

pub fn build_turn( &self, turns: &[MtmdTurn], tokenizer_fn: Option<&(dyn Fn(&str) -> Result<Vec<u32>> + '_)>, ) -> Result<AssembledTurn>

Assemble one turn list into AssembledTurn.

Renders the text using the registered chat template, replaces each <|image|> / <|audio|> marker with the per-family placeholder token id (resolved from the optional tokenizer vocabulary), and records the order of media in image_refs / audio_refs so the runner can insert the corresponding embeddings at decode time.

tokenizer_fn lets the caller plug in a per-family text→ids encoder (typically auto_tokenize). Passing None populates text_tokens with an empty vec — useful for callers that own tokenization separately.