Struct PromptLayout 

pub struct PromptLayout {
    pub image_token: String,
    pub fake_token: String,
    pub global_token: String,
    pub row_col_fmt: fn(usize, usize) -> String,
    pub tokens_per_tile: usize,
}

The pieces of the layout that vary by checkpoint, read from config.json and the processor rather than hard-coded.

Fields

image_token: String

Repeated tokens_per_tile times per image block.

fake_token: String

Wraps every image block and closes the run.

global_token: String

Marks the global thumbnail’s block.

row_col_fmt: fn(usize, usize) -> String

<row_{r}_col_{c}>, 1-based.

tokens_per_tile: usize

(image_size / patch_size)^2 / scale_factor^2 — 64 for SmolVLM-256M.

Implementations

impl PromptLayout

pub const fn with_geometry( self, image_size: usize, patch_size: usize, scale_factor: usize, ) -> Self

Derive tokens_per_tile from the vision geometry.

Computed rather than constant: it is (image_size / patch_size)^2 / scale_factor^2, so a different SmolVLM size changes it and a hard-coded 64 would be silently wrong there — the same class of defect this whole module is guarding against.

pub fn image_block(&self, rows: usize, cols: usize) -> String

The image-block run: 16 tiles row-major, then \n\n, then the global thumbnail, then a closing fake token.

rows/cols describe the tile grid the preprocessor chose. rows == 0 means the image was small enough that only the thumbnail exists — the reference emits just the global block then, with no grid and no \n\n.

pub fn user_turn(&self, question: &str, rows: usize, cols: usize) -> String

The full user turn, chat template included.

The template is SmolVLM’s own — <|im_start|>User: … Assistant: — and it is written here only because the tokenizer’s Jinja template is not available to this crate. It is checked against the reference’s own output, which is the only thing that makes writing it acceptable at all.

Trait Implementations

impl Clone for PromptLayout

fn clone(&self) -> PromptLayout

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PromptLayout

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for PromptLayout

fn default() -> Self

Returns the “default value” for a type.