Struct Calibrator 

pub struct Calibrator<'g> { /* private fields */ }

Compiled calibration harness. The graph is owned by the caller — we hold a reference and the compiled artifacts (arena + schedule). The caller writes inputs and parameters into arena_mut() between batches.

Implementations

impl<'g> Calibrator<'g>

pub fn new(graph: &'g Graph, taps: Vec<NodeId>) -> Self

Build a calibrator over graph that records max-abs at each tap after every step(). The graph must already have those taps in its outputs list (so the memory planner keeps their arena slots alive to end-of-execution); this constructor asserts the precondition.

pub fn arena_mut(&mut self) -> &mut Arena

Mutable arena access — for writing inputs/params before each step() and (typically once at startup) for filling Op::Constant data via rlx_runtime’s loader.

pub fn arena(&self) -> &Arena

Read-only arena view — for reading the tap values manually if the caller wants something fancier than max-abs.

pub fn step(&mut self)

Run one forward batch, then update each tap’s running max-abs.

pub fn max_abs(&self) -> &[f32]

Per-tap max-abs accumulated so far (in input order).

pub fn scales(&self) -> Vec<f32>

Per-tap scale = max_abs / 127.0, clamped up to 1e-6. Use directly as the scale for Op::Quantize / Op::Dequantize or rlx_opt::CalibrationEntry::per_tensor.

pub fn graph(&self) -> &Graph

Borrow the inner graph (for the caller to re-look-up NodeIds after compilation).