Struct VlmOptions 

pub struct VlmOptions {
    pub prompt: Option<String>,
    pub system_prompt: Option<String>,
    pub decoding: Decoding,
    pub max_new_tokens: Option<usize>,
    pub stop: Vec<String>,
    pub frames_per_window: Option<usize>,
    pub repetition_penalty: Option<f32>,
}

Options for a VLM call (Argus).

Gate 2 — the v1 surface, and what is deliberately NOT here

Settled once, before implementation, because every field is cheap now and a breaking change later (docs/plans/argus-launch-plan.md §2 Gate 2).

Excluded from v1, as decisions rather than oversights:

• Streaming. Tokens as produced. It changes the return type of every method, so it is a trait redesign and not a field; it waits until there is a consumer that needs it.
• Grounding. Region-in (“what is in this box”) and grounded-out (“the dog [x,y,w,h]”). Diana already returns boxes; a second, weaker box source in the toolkit needs a reason beyond “the model can”.
• Structured / JSON output. Needs constrained decoding to be worth anything — an unconstrained “please reply in JSON” is a request, not a guarantee. High value for a tooling product, so it is v2 rather than never; mistral.rs already provides grammar-constrained decoding, which is what makes it cheap when it lands.
• Token-level confidence (logprobs). Carmenta uses recognition confidence for its verifier and the analog is genuinely valuable, but it is only meaningful once an engine exists to produce it honestly.
• Conversation history / multi-turn. VlmPrompt can already express an interleaved turn sequence; a typed history is a chat-session concern, and Argus is a media op.

Not here because it belongs to the ENGINE, not the caller: the chat template, the special vision tokens (<image>, <|vision_start|>), and the M-RoPE time axis for video. Those are properties of a specific checkpoint. Exposing them as options would invite a caller to set them wrong — and the launch plan measured what that costs: 43 of 50 answers changed on identical weights purely from prompt formatting, with no error raised.

Fields

prompt: Option<String>

Instruction for the model; None = plain captioning.

Read by VlmEngine::describe_image only. VlmEngine::describe takes its text from the prompt’s own VlmPart::Text parts, because position matters there and a separate field could not say where the text goes.

system_prompt: Option<String>

Role/behaviour framing, when the checkpoint’s template has a slot for it. Engines without one must ignore it rather than concatenate it into the user turn, which changes the prompt the model was tuned on.

decoding: Decoding

Token-selection strategy. Defaults to Decoding::Greedy.

max_new_tokens: Option<usize>

Token budget. None = the engine’s own default.

stop: Vec<String>

Stop strings. Generation halts as soon as one is produced; the string itself is not included in the returned text.

EOS is not here — that is the tokenizer’s, and an engine that needed to be told its own EOS would be misconfigured.

frames_per_window: Option<usize>

Frames shown to the model per video caption. None = the engine’s own default.

This is the video knob that decides what a caption can be ABOUT, and the arithmetic is worth stating because it is not obvious. A still image is split into tiles so fine print survives — for SmolVLM that is 17 tiles at 64 tokens, 1088 tokens per image. A text tower with 8192 positions therefore holds seven such frames, which is not a window so much as a slideshow.

Video engines turn splitting off, making a frame one tile, and the same 8192 positions then hold a hundred. So this number trades fine detail within a frame against temporal context across frames, and the right value depends on the question being asked — which is why it is a caller’s option and not a constant.

1 degenerates to per-frame captioning: correct, and blind to motion.

repetition_penalty: Option<f32>

Repetition penalty, None = disabled.

Sits beside Self::decoding rather than inside Decoding::Sampled because it is a logit transform, not a sampling one: it applies to greedy decoding too, and small models loop under greedy more than under sampling.

Trait Implementations

impl Clone for VlmOptions

fn clone(&self) -> VlmOptions

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for VlmOptions

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

Formats the value using the given formatter.

impl Default for VlmOptions

fn default() -> VlmOptions

Returns the “default value” for a type.