Struct DenseFp8Plan 

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

What a native-FP8 dense layer must materialise beyond the checkpoint bytes.

Every field is “build this derived copy”: false means the selected kernels provably never read it.

Fields

ffn_nvfp4: bool

NVFP4 gate/up/down and their w4a16_gemm_t_m128 transposed twins.

attn_nvfp4: bool

NVFP4 q/k/v/o, their transposed twins, and the fused [q|k|v] twin.

attn_fp8_twins: Fp8TwinSet

Which Fp8WeightTransposed twins transpose_fp8_for_prefill builds.

Implementations

impl DenseFp8Plan

pub fn resolve(i: DenseFp8Inputs) -> Self

The decision table. Pure — no environment reads, no allocation.

• FFN NVFP4 — unreachable the moment set_fp8_weights runs (see the module docs: every entry point returns from inside the FP8 arm and the w8_gemm! transposed operands are a literal None). Built only under the escape hatch.
• Attention NVFP4 — set_fp8_weights overwrites q_weight/k_weight/v_weight/o_weight with QuantWeight::Fp8, so the base NVFP4 weights are orphaned at load. The transposed and fused twins survive only behind the ATLAS_CUTLASS_NVFP4_* levers, which default off.
• Attention FP8 twins, K and V — KEPT UNCONDITIONALLY. The first prefill chunk (seq_len_start == 0, the default for every request) does not go through paged_qkv.rs at all: it goes through prefill/cache_skip_qkv.rs, whose dispatch chain has no W8A8 arm, so k_fp8w_t/v_fp8w_t are dereferenced at cache_skip_qkv.rs:218 / :235 on every request regardless of fp8_blockscaled_prefill. Freeing them is a NULL-pointer kernel launch on the first token.
• Attention FP8 twins, Q and O — Q on that same chain is behind ATLAS_ATTN_PREFILL_Q_T=1 (cache_skip_qkv.rs:142) and O is routed to paged_oproj.rs from both chains, so both are reachable only after the W8A8 arm declines — block-scaled prefill off, or a target missing one of the two kernels.

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