Enum ZeroKvLayer 

pub enum ZeroKvLayer {
    Linear,
    ShortConv,
    Mamba2,
    Mamba1,
    Plamo2,
    Mamba2UnlessFfn,
    Unserved(&'static str),
}

What head_count_kv == 0 with head_count > 0 MEANS for an architecture, because two graphs spell two different blocks with the same two counts and the counts alone cannot tell them apart.

Variants

Linear

deci.cpp:115-118: attn_norm then wo.

ShortConv

lfm2.cpp:197: the short convolution (crate::shortconv).

Mamba2

granite-hybrid.cpp:163: the Mamba-2 block (crate::mamba2).

Mamba1

jamba.cpp:128: the Mamba-1 block (crate::mamba1).

Plamo2

plamo2.cpp:218: PLaMo-2’s own block (crate::plamo2_ssm).

Mamba2UnlessFfn

nemotron-h.cpp:9-11: the Mamba-2 block when the layer’s FFN width is ALSO zero, and an FFN-only layer (no attention block at all, AttnShape::Absent) when it is not – every Nemotron-H layer is one block, and feed_forward_length is the second array that says which.

Unserved(&'static str)

A recurrent block frink has no body for; the reason names it.

Implementations

impl ZeroKvLayer

pub fn for_arch(arch: &str) -> Self

The table. Every architecture in PER_LAYER_SHAPE_ARCHS whose graph reads n_head_kv(il) == 0 as a layer kind has a row here; anything else that reaches a zero is Linear, the reading the generic path had before the table existed, which only deci’s converter writes.

Trait Implementations

impl Clone for ZeroKvLayer

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for ZeroKvLayer

impl Debug for ZeroKvLayer

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

Formats the value using the given formatter.

impl Eq for ZeroKvLayer

impl PartialEq for ZeroKvLayer

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for ZeroKvLayer