Struct Boot 

pub struct Boot<'a> {

    pub gpu: &'a dyn GpuBackend,
    pub config: &'a ModelConfig,
    pub config_json: &'a str,
    pub levers: &'a ModelLevers,
    pub instance: &'a Instance,
    pub policy: Policy,
    pub layers: Vec<Option<CircuitLayer>>,
    pub head: HeadBinding,
    pub fixed: Fixed,
    pub fusions: Fusions,
    pub modules: &'a TargetModules,
    pub multi_seq_rows: Vec<u64>,
    pub verify_rows: Vec<u64>,
    pub draft: Option<CircuitLayer>,
    pub verify_batch_rows: Option<u64>,
    pub draft_rows: Option<u64>,
}

2026-09-28: What a build reads.

Fields

gpu: &'a dyn GpuBackend

config: &'a ModelConfig

config_json: &'a str

2026-09-30: The checkpoint’s config.json, which the circuit’s config map turns into the served shape (sources::checkpoint_shape).

levers: &'a ModelLevers

2026-09-28: The model’s levers, as its legacy layers read them (ForwardContext::levers).

instance: &'a Instance

policy: Policy

layers: Vec<Option<CircuitLayer>>

head: HeadBinding

fixed: Fixed

fusions: Fusions

modules: &'a TargetModules

multi_seq_rows: Vec<u64>

2026-09-28: The padded multi-sequence batch widths to compile (the decode graph ladder up to the serve’s widest batch).

verify_rows: Vec<u64>

2026-09-29: The MTP verify widths K to compile (none without speculative decode).

draft: Option<CircuitLayer>

2026-09-29: The MTP draft head, when one is loaded; its buffers are fixed.draft.

verify_batch_rows: Option<u64>

2026-09-30: The widest batched MTP verify, in rows (Σ k), when the serve verifies several sequences in one forward; None otherwise.

draft_rows: Option<u64>

2026-09-30: The widest batched propose (propose_batch_max), whose widths from 2 get n-row draft programs; None when the head proposes one sequence at a time.