Enum NormOp 

pub enum NormOp {
    Rms(Vec<f32>),
    LayerNormNoParams,
    None,
}

The normalisation applied at one norm site.

Variants

Rms(Vec<f32>)

RMSNorm with these learned weights. Every architecture on the generic path except the two families below.

LayerNormNoParams

Non-parametric LayerNorm: subtract the mean, divide by the standard deviation, no learned weight and no bias.

olmo (OLMo-1) and nothing else in llama.cpp. There is no LayerNorm(Vec<f32>) beside this because no architecture ferrox admits needs one: the weighted-LayerNorm rows (dbrx, nemotron, orion, stablelm, …) all refuse for other reasons too, and a variant with no caller is a variant that silently rots.

None

No norm at all: the branch reads the raw residual.

olmo2 and exaone4. NOT “an RMSNorm whose weights are all ones” – that would still divide by the RMS of the residual, and the whole point of this variant is that nothing is divided.

Implementations

impl NormOp

pub fn apply(&self, x: &[f32], eps: f32) -> Vec<f32>

The site’s output: rms_norm(x, w, eps), the non-parametric LayerNorm, or x itself.

Returns an owned vector in every arm, which is what every caller already had: rms_norm allocates too.

eps is ModelConfig::rms_norm_eps, which for a LayerNorm architecture is read from {arch}.attention.layer_norm_epsilon – llama.cpp’s f_norm_eps rather than f_norm_rms_eps. The two are one field here because no architecture reads both keys, and loader.rs already accepted either spelling into that field before this variant existed.

pub fn rms_weights(&self) -> Option<&[f32]>

The weights a fused GPU kernel needs, or None when this site has no RMSNorm weights and the kernel therefore must not run.

Every fused Metal launch that bakes an RMSNorm into its kernel goes through here and falls back to the host body on None. See the module docs for why that is a ? and not a comment.

Trait Implementations

impl Clone for NormOp

fn clone(&self) -> NormOp

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for NormOp

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

Formats the value using the given formatter.

impl PartialEq for NormOp

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for NormOp