Struct DenseFp8Inputs 

pub struct DenseFp8Inputs {
    pub ffn_fp8: bool,
    pub attn_fp8: bool,
    pub keep_nvfp4: bool,
    pub dispatch: GemmDispatch,
    pub w8a8_kernels: bool,
    pub attn_w4a4: bool,
    pub attn_prefill_q_t: bool,
}

The inputs DenseFp8Plan::resolve decides from. Taken as a struct so the CPU decision-table test can pin every clause without touching the process environment (the OnceLock resolvers cannot be toggled per test).

Fields

ffn_fp8: bool

The native FP8 dense-FFN overlay will be installed on this layer (ATLAS_DENSE_FP8=1, tp_size 1, Fp8Dequanted, gate_proj native FP8).

attn_fp8: bool

The native FP8 attention overlay will be installed on this layer.

keep_nvfp4: bool

ATLAS_DENSE_FP8_KEEP_NVFP4 — restore the pre-#915 behaviour.

dispatch: GemmDispatch

Resolved exactly as model/impl_a1.rs:836 resolves it.

w8a8_kernels: bool

Whether the target’s per_token_group_quant_fp8 + fp8_gemm_t_blockscaled kernels are both loaded. Both are required by the W8A8 prefill arm in qwen3_attention/prefill/paged_qkv.rs:220 and prefill/paged_oproj.rs:94; without them those two chains fall through to the transposed W8A16 kernels, which read the FP8 twins.

attn_w4a4: bool

ATLAS_ATTN_W4A4 is set. prefill/paged_oproj.rs:38-42 builds its W4A4 arm with NO weight-type predicate and then feeds it &self.attn.o_proj — the NVFP4 o_proj — so this one lever keeps the NVFP4 attention weights alive even under a full FP8 overlay. (The QKV side at paged_qkv.rs:51 does check as_nvfp4(), so it is already closed; the o_proj asymmetry is not.)

attn_prefill_q_t: bool

ATLAS_ATTN_PREFILL_Q_T=1. prefill/cache_skip_qkv.rs:142 reads it per projection per prefill and, when set, dispatches Q through q_fp8w_t.

Trait Implementations

impl Clone for DenseFp8Inputs

fn clone(&self) -> DenseFp8Inputs

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DenseFp8Inputs

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

Formats the value using the given formatter.

impl PartialEq for DenseFp8Inputs

fn eq(&self, other: &DenseFp8Inputs) -> 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 DenseFp8Inputs

impl Eq for DenseFp8Inputs

impl StructuralPartialEq for DenseFp8Inputs