Struct SundialModel 

pub struct SundialModel { /* private fields */ }

Sundial reads its own architecture (layer count, flow depth, output length, sampling steps) straight out of the GGUF metadata written at conversion time — there is no separate InferConfig/config.json needed for inference (unlike every other model in the workspace). SundialModelBuilder only exposes the two things a caller can actually tune at load time: the compute device and an ODE step-count override.

Implementations

impl SundialModel

pub fn builder(gguf_path: impl Into<PathBuf>) -> SundialModelBuilder

Start building a SundialModel — see SundialModelBuilder.

pub fn load( path: &Path, device: &Device, steps_override: Option<usize>, ) -> Result<Self>

pub fn forecast(&self, context: &[f32], device: &Device) -> Result<Vec<f32>>

Trait Implementations

impl Forecaster for SundialModel

type Config = ()

No config.json is consulted for inference — see the struct docs.

fn forecast( &self, context: &[Vec<f32>], _mask: &[Vec<bool>], horizon: usize, ) -> Result<QuantileMatrix>

Sundial is univariate-only and point-forecast-only (flow-matching samples a single path, not a quantile distribution) — mask is unused and the returned matrix has exactly one row (the point forecast) and one variate.

fn load(gguf_path: &Path, _config: ()) -> Result<Self>