Skip to content

stdlib/inferView source ↗

infer

9 items · generated by scribe fromstdlib/infer.fd· synced 2026-11-15

stdlib/infer9 items

Contents

infer_matvec

y = W x for W in row-major [rows][cols] f64.

pub fn infer_matvec(w: Vec<f64>, rows: i64, cols: i64, x: Vec<f64>) -> Vec<f64> {

Parameters:

  • w: Vec<f64>
  • rows: i64
  • cols: i64
  • x: Vec<f64>

Returns: Vec<f64>

Defined at line 7


infer_rmsnorm

RMSNorm: x / sqrt(mean(x^2) + eps) * weight.

pub fn infer_rmsnorm(x: Vec<f64>, w: Vec<f64>, eps: f64) -> Vec<f64> {

Parameters:

  • x: Vec<f64>
  • w: Vec<f64>
  • eps: f64

Returns: Vec<f64>

Defined at line 25


infer_rope

Apply rotary position embedding to a q/k head: pairs (x[2i], x[2i+1]) rotate by pos * theta^(-2i/head_dim). Returns a new Vec so it works on every backend.

pub fn infer_rope(v: Vec<f64>, head_dim: i64, pos: i64, theta: f64) -> Vec<f64> {

Parameters:

  • v: Vec<f64>
  • head_dim: i64
  • pos: i64
  • theta: f64

Returns: Vec<f64>

Defined at line 45


infer_softmax

Numerically stable softmax (max subtracted).

Example (native only): use vec; var x: Vec= Vec::new(); x.push(1.0); x.push(1.0); var p = infer_softmax(x); println(float_to_str(p.get(0))); // 0.5

pub fn infer_softmax(x: Vec<f64>) -> Vec<f64> {

Parameters:

  • x: Vec<f64>

Returns: Vec<f64>

Defined at line 73


infer_argmax

Index of the largest element (ties -> lowest index), -1 when empty.

pub fn infer_argmax(x: Vec<f64>) -> i64 {

Parameters:

  • x: Vec<f64>

Returns: i64

Defined at line 104


infer_attn

Single-head attention: q (dim), keys/values flat [seq][dim]. Returns sum_i softmax(q.k_i / sqrt(dim)) * v_i — a Vec of dim f64.

pub fn infer_attn(q: Vec<f64>, k: Vec<f64>, v: Vec<f64>, seq: i64, dim: i64) -> Vec<f64> {

Parameters:

  • q: Vec<f64>
  • k: Vec<f64>
  • v: Vec<f64>
  • seq: i64
  • dim: i64

Returns: Vec<f64>

Defined at line 121


infer_top_k

Top-k indices of logits, largest-first (selection sort over k).

pub fn infer_top_k(x: Vec<f64>, k: i64) -> Vec<i64> {

Parameters:

  • x: Vec<f64>
  • k: i64

Returns: Vec<i64>

Defined at line 154


infer_sample

Sample from a probability Vec using u in [0,1). Returns an index. Caller normalizes/softmaxes first (or passes probs from top-k).

pub fn infer_sample(probs: Vec<f64>, u: f64) -> i64 {

Parameters:

  • probs: Vec<f64>
  • u: f64

Returns: i64

Defined at line 188


infer_sample_top_p

Top-p (nucleus): keep the smallest prefix of sorted probs whose mass reaches p, then sample among them. order from infer_top_k over the softmax’d logits, probs the same softmax values. Returns the chosen LOGIT index.

pub fn infer_sample_top_p(probs: Vec<f64>, order: Vec<i64>, p: f64, u: f64) -> i64 {

Parameters:

  • probs: Vec<f64>
  • order: Vec<i64>
  • p: f64
  • u: f64

Returns: i64

Defined at line 208


Generated by scribe, the Fides documentation generator