Struct LoadedCheckpoint 

pub struct LoadedCheckpoint {
    pub plan: ModelPlan,
    pub weights: ReferenceWeights,
    /* private fields */
}

A checkpoint materialized through the pack contract: reference-layout weights for the trunk + globals (effective norms — the (1+w) fold applied per the module-header rule), plus the table carrier that stays out of ReferenceWeights and the LAZY bank plans.

The open safetensors mmap is part of the value: expert banks are read from it per layer at consume time (see BankPlan). It stays mapped until the checkpoint is dropped, so a consumer must not outlive it — every consumer here is a constructor that finishes uploading before returning.

Fields

plan: ModelPlan

weights: ReferenceWeights

Implementations

impl LoadedCheckpoint

pub fn bank_fingerprints(&self) -> Result<Vec<BankFingerprint>, Box<dyn Error>>

Per-projection byte fingerprints for every bank, read one layer at a time (so this costs one layer of host memory, not the artifact). Gate instrument only — see BankFingerprint; the tiny-fixture gate compares these against banked goldens.

pub fn into_reference_weights(self) -> Result<ReferenceWeights, Box<dyn Error>>

Expand banks and n-gram tables into plain ReferenceWeights entries so memra-reference can execute the checkpoint. TINY/SIBLING SCALE ONLY — the real artifact’s banks/table do not fit host f32; the GPU path never takes this.

impl LoadedCheckpoint

pub fn mtp_reference_weights(&self) -> Result<ReferenceWeights, Box<dyn Error>>

CLONE the float weights and expand ONLY the MTP bank(s) into ReferenceWeights entries — the real-checkpoint draft-parity instrument (mtp-spec lane): the host reference twin needs the mtp.* rows + embed/head, and must NOT expand the trunk banks (48 layers of f32 experts do not fit anywhere). Borrowing form so ONE checkpoint read serves both the engine model and the host twin.