Struct DenseFp8Plan 

pub struct DenseFp8Plan {
    pub ffn_nvfp4: bool,
    pub attn_nvfp4: bool,
    pub attn_fp8_twins: Fp8TwinSet,
}

2026-09-25: Which derived copies one dense layer builds beyond the checkpoint bytes. Each field means “build this copy”.

Fields

ffn_nvfp4: bool

2026-09-25: The NVFP4 gate/up/down weights and their transposed twins (fp8_lut::load_dense_ffn).

attn_nvfp4: bool

2026-09-25: The NVFP4 q/k/v/o weights, their transposed twins, and the fused [q|k|v] transposed twin.

attn_fp8_twins: Fp8TwinSet

2026-09-25: Which FP8 prefill twins transpose_fp8_for_prefill_selected builds.

Implementations

impl DenseFp8Plan

pub fn resolve(i: DenseFp8Inputs) -> Self

2026-09-25: The decision table.

• FFN NVFP4: built only when the layer has no FFN overlay, or under keep_nvfp4 (the module docs say why the FP8 FFN never reads it).
• Attention NVFP4: set_fp8_weights replaces q_weight, k_weight, v_weight and o_weight with QuantWeight::Fp8. The NVFP4 copies are kept when a METRALE_CUTLASS_NVFP4_* attention arm (cutlass_nvfp4_* in the dispatch) or the W4A4 o_proj arm (attn_w4a4) can read them.
• FP8 twins K and V: built on every FP8 attention layer unless the umbrella flag is set (last item). A single stream’s first prefill chunk (seq_len_start == 0) runs prefill/cache_skip_qkv.rs, whose only W8A8 arm is the cuBLASLt one (cache_skip_qkv_cublas_selected, which needs METRALE_CUBLAS_GEMM to cover attn). When it declines, K and V read k_fp8w_t and v_fp8w_t if present.
• FP8 twins Q and O: Q is read on that chain only with attn_prefill_q_t, and on the paged chain after its W8A8 arm declines; both chains run O through prefill/paged_oproj.rs, which reads the O twin after its W8A8 arm declines. The W8A8 arms decline when block-scaled prefill is off or a W8A8 kernel is missing.
• Under the umbrella cutlass_nvfp4_gemm no FP8 twin is planned, because transpose_fp8_for_prefill_selected builds none under that flag.

Trait Implementations

impl Clone for DenseFp8Plan

fn clone(&self) -> DenseFp8Plan

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DenseFp8Plan

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

Formats the value using the given formatter.

impl PartialEq for DenseFp8Plan

fn eq(&self, other: &DenseFp8Plan) -> 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 DenseFp8Plan

impl Eq for DenseFp8Plan

impl StructuralPartialEq for DenseFp8Plan