Struct TabFMModel 

pub struct TabFMModel { /* private fields */ }

Implementations

impl TabFMModel

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

Start building a TabFMModel — see TabFMModelBuilder.

pub fn load(gguf_path: &Path, config: InferConfig) -> Result<Self>

pub fn is_classifier(&self) -> bool

pub fn predict( &self, x: &[Vec<f32>], y: &[f32], train_size: usize, cat_mask: Option<&[bool]>, d: Option<usize>, ) -> Result<Vec<Vec<f32>>>

Run one table (train rows followed by test rows) through TabFM. A thin B=1 wrapper around predict_batch — see that method to run many tables (e.g. ensemble members) in one forward pass.

x: [T][H] padded feature matrix (numeric; categorical columns are pre-encoded to floats by the caller). y: [T] labels (any finite placeholder at test-row positions is fine — it’s masked internally and never influences the output). train_size: number of leading rows that are training rows. cat_mask: [H], which columns are categorical (defaults to all-false). d: actual (unpadded) feature count (defaults to H).

Returns [T][out_dim] raw logits (classification, out_dim = max_classes) or a [T][1] scalar (regression) — only rows >= train_size are meaningful predictions.

pub fn predict_batch( &self, x_batch: &[Vec<Vec<f32>>], y_batch: &[Vec<f32>], train_size: usize, cat_mask_batch: &[Vec<bool>], d: Option<usize>, ) -> Result<Vec<Vec<Vec<f32>>>>

Runs B independent tables through one forward pass, sharing the fixed cost of the 24-block ICL stage (and every other stage) across all of them instead of paying it once per table. Every table must share the same row count T and feature count H, and train_size/d are shared scalars across the whole batch — this holds for TabFM’s ensemble members, which only differ in cell values (per-member feature permutation/scaling) and cat_mask (which position is categorical shifts with the permutation), never in table shape. Returns [B][T][out_dim].