Struct CublasScope 

pub struct CublasScope {
    pub ffn: bool,
    pub attn: bool,
    pub ssm: bool,
    pub head: bool,
}

Which projection FAMILIES take a cuBLASLt GEMM arm.

WHY a set and not a bool — H100, 2026-09-11, Qwen/Qwen3.8-27B-FP8 native FP8, tip 5f78270dc, “config B” of the round-3 receipt. ATLAS_CUBLAS_GEMM=1 resolved to ONE global boolean, so the variable that arms #917’s dense-FFN W8A8 fast path ALSO armed the SSM in_proj_qkvz arm, which materialised a cached BF16 dequant of the fused QKVZ weight ([10240,5120] + [6144,5120] x 2 B = 167772160 bytes per layer, ~10.3 GiB over 48 SSM layers) outside the buffer ledger. One 28-token prefill ate 6120 MiB and died at layer 36 with cuMemAlloc_v2 ... status 2. The FFN arm under test could not be exercised end-to-end on an 80 GB card, and no knob separated the two. The CUTLASS family next door already spells its projections out (ATLAS_CUTLASS_NVFP4_QKVZ, ..._ATTN_Q, ..._ATTN_KV, ..._ATTN_O, ..._SSM_OUT); this gives cuBLASLt the same property in one variable instead of five.

GRAMMAR — ATLAS_CUBLAS_GEMM=<token>[,<token>]*, ASCII-case-insensitive, whitespace around a token ignored:

token	meaning
ffn	dense-FFN + MoE shared-expert projections
attn	attention Q/K/V, O and the output gate
ssm	SSM/GDN in_proj_qkvz
head	LM / MTP head (see CublasScope::head)
all, 1, true	every family — the pre-2026-09-11 spelling
off, 0, false, empty	the empty set

The result is the UNION of the tokens, so off adds nothing rather than clearing what another token armed (ffn,off is ffn). Unknown tokens are dropped with a warning and never widen the set: a typo must not silently arm an arm, which is the exact failure above.

Fields

ffn: bool

Dense-FFN gate/up/down and the MoE shared expert — #917’s W8A8 arm.

attn: bool

Attention Q/K/V, O projection and the output gate.

ssm: bool

SSM/GDN fused in_proj_qkvz prefill projection.

head: bool

LM / MTP head. Parsed, carried and covered by all, but NO dispatch site reads it yet — setting head is inert today. Named anyway so the grammar is the complete family list and all has a fixed meaning; a lever that silently drops a spelling is worse than one that documents an unclaimed slot.

Implementations

impl CublasScope

pub const OFF: Self

No family armed — the shape an absent or off ATLAS_CUBLAS_GEMM resolves to, and the default for every build.

pub const ALL: Self

Every family — what all / 1 / true resolve to.

pub fn any(&self) -> bool

Whether any family is armed (for the resolved-set log line).

Trait Implementations

impl Clone for CublasScope

fn clone(&self) -> CublasScope

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CublasScope

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

Formats the value using the given formatter.

impl Default for CublasScope

fn default() -> CublasScope

Returns the “default value” for a type.

impl PartialEq for CublasScope

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

impl Eq for CublasScope

impl StructuralPartialEq for CublasScope