Struct PredictedDerived 

pub struct PredictedDerived {
    pub attn_fp8_twins: u64,
    pub ssm_fp8_concat: u64,
    pub ffn_gateup_fused: u64,
    pub ffn_gateup_pruned: u64,
    pub twins: TwinsBuilt,
    pub attn_twin_set: Fp8TwinSet,
}

The derived bytes the native-FP8 dense loader will allocate and KEEP, broken out so the preflight log can name each term.

Mirrors DerivedResidency::kept term for term — see the module docs’ receipt table — so predicted.total() and the serve log’s native FP8 dense residency: ... derived X GB are the same arithmetic evaluated at two different times.

Fields

attn_fp8_twins: u64

Fp8Weight::transpose_for_gemm twins for the projections fp8_residency::DenseFp8Plan selects, summed over full-attention layers.

ssm_fp8_concat: u64

The fused [QKV|Z] FP8 weight, its block-scale grid, the out_proj block-scale grid and the interleaved in_proj_ba, summed over linear-attention layers.

ffn_gateup_fused: u64

The fused [2*inter, hidden] dense-FFN gate+up weight and its block-scale grid (#927), summed over dense-FFN layers — 0 when the target does not arm the arm.

ffn_gateup_pruned: u64

The checkpoint bytes prune_after_load gives BACK because the fusion consumed them: mlp.gate_proj.weight + mlp.up_proj.weight over the same layers. EQUAL to Self::ffn_gateup_fused minus the scale grids, by construction — the fused weight IS those two tensors copied side by side — which is why the fusion nets out of Self::total below.

twins: TwinsBuilt

Which twin families the prediction expects, for the log line.

attn_twin_set: Fp8TwinSet

The twin set the attention term was priced at.

Implementations

impl PredictedDerived

pub fn total(&self) -> u64

Derived bytes ABOVE the on-disk checkpoint count, which is what headroom.rs adds to weights to build the post-load yardstick.

The gate+up fusion appears as a DIFFERENCE and not as a term: the caller’s weights is the checkpoint’s on-disk size, which still counts gate_proj.weight and up_proj.weight — and the loader releases both once the fused copy exists. Adding the fused weight without subtracting what it replaces would over-state pre-KV by 11.4 GB on Qwen3.8-27B and silently shrink — or refuse — the decode-rollback ring this yardstick exists to fit. The two terms are equal by construction, so the difference is exactly zero and every prediction taken before #927 is unchanged; both are carried so that is legible rather than asserted (the_gateup_fusion_is_residency_neutral).

Trait Implementations

impl Clone for PredictedDerived

fn clone(&self) -> PredictedDerived

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PredictedDerived

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for PredictedDerived

fn default() -> PredictedDerived

Returns the “default value” for a type.

impl PartialEq for PredictedDerived

fn eq(&self, other: &PredictedDerived) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for PredictedDerived

impl Eq for PredictedDerived

impl StructuralPartialEq for PredictedDerived