pub struct LoadOptions {
pub host_bf16_banks: bool,
pub indexer_norm_raw: bool,
pub load_mtp: bool,
}Expand description
Real-checkpoint loader knobs (defaults = the tiny-gate behavior).
Fields§
§host_bf16_banks: boolKeep BF16 expert banks HOST-resident (raw bf16) and upload+upcast per ROUTED expert at forward time. Gate-mode residency for artifacts whose f32 banks exceed device memory; value chain identical to the f32 device arm (bf16→f32 is exact). Never a serving configuration.
indexer_norm_raw: boolBind the indexer q/k layernorms RAW (skip the (1+w) fold) — the two-arm probe for the SEMANTICS.md VERIFY marker. Default keeps the family fold.
load_mtp: boolMaterialize the mtp.* namespace (the NextN draft block) — the mtp-spec lane.
The MTP expert bank keeps its raw BF16 bytes at read time and goes DEVICE
bf16-resident at build (BankHalf::DeviceBf16, ~5 GB beside the NVFP4 trunk).
Default OFF: the plain eager arm executes no draft.
Trait Implementations§
Source§impl Clone for LoadOptions
impl Clone for LoadOptions
Source§fn clone(&self) -> LoadOptions
fn clone(&self) -> LoadOptions
1.0.0 (const: unstable) · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read more