Struct TimesFMConfig 

pub struct TimesFMConfig {

    pub input_patch_len: usize,
    pub output_patch_len: usize,
    pub output_quantile_len: usize,
    pub num_layers: usize,
    pub d_model: usize,
    pub d_ff: usize,
    pub num_heads: usize,
    pub head_dim: usize,
    pub quantiles: Vec<f32>,
    pub n_outputs: usize,
    pub decode_index: usize,
    pub context_limit: usize,
    pub rms_norm_eps: f64,
    pub rope_theta: f64,
}

Hardcoded architecture constants for TimesFM 2.5 200M.

These match TimesFM_2p5_200M_Definition in the Python source exactly. No config.json is needed — the architecture is fixed for this model.

Fields

input_patch_len: usize

Input patch length (p = 32).

output_patch_len: usize

Output patch length (o = 128).

output_quantile_len: usize

Output quantile length for the continuous quantile head (os = 1024).

num_layers: usize

Number of stacked transformer layers.

d_model: usize

Transformer model dimension.

d_ff: usize

Feed-forward hidden dimension.

num_heads: usize

Number of attention heads.

head_dim: usize

Head dimension = d_model / num_heads.

quantiles: Vec<f32>

Quantile levels predicted (excludes the implicit point forecast at index 0).

n_outputs: usize

Total number of per-timestep outputs = len(quantiles) + 1 (point).

decode_index: usize

Index used to extract point/AR forecast from the 10-dim output.

context_limit: usize

Maximum context tokens the model accepts (in individual timesteps).

rms_norm_eps: f64

RMS norm epsilon.

rope_theta: f64

RoPE base frequency (theta = 10000).

Implementations

impl TimesFMConfig

pub fn new() -> Self

Trait Implementations

impl Clone for TimesFMConfig

fn clone(&self) -> TimesFMConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TimesFMConfig

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

Formats the value using the given formatter.

impl Default for TimesFMConfig

fn default() -> Self

Returns the “default value” for a type.