Struct StreamedTrunkExecution 

pub struct StreamedTrunkExecution<'a> { /* private fields */ }

Layer-at-a-time trunk execution over the exact per-layer math of execute, for checkpoint-scale runs where all weights cannot be resident at once. The driver materializes only the current layer’s tensors, calls StreamedTrunkExecution::step, and frees them before the next layer.

begin needs TokenEmbedding in globals; finish needs OutputNorm plus OutputProjection (falling back to the embedding for tied heads) and, for a gated-head collapse, the HyperHead* tensors.

Deliberate scope: trunk + final norm + LM head only. MTP blocks are not executed (mtp stays empty) and drafter plans are refused at begin — the streamed path has no per-layer tap capture. The glm5 checkpoint runner’s --self-test mode pins this path against execute bit-for-bit.

Implementations

impl<'a> StreamedTrunkExecution<'a>

pub fn begin( plan: &'a ModelPlan, globals: &ReferenceWeights, token_ids: &[u32], ) -> Result<Self, ReferenceError>

pub fn next_layer(&self) -> Option<&'a LayerPlan>

The plan layer the next Self::step call will execute, or None when the trunk is fully executed.

pub fn step( &mut self, weights: &ReferenceWeights, ) -> Result<u32, ReferenceError>

Execute the next trunk layer using only that layer’s tensors. Returns the executed layer’s plan index.

pub fn finish( self, globals: &ReferenceWeights, ) -> Result<ReferenceOutput, ReferenceError>

Collapse, final-norm, and project the trunk. MTP blocks are skipped by design.