Enum RopeLayout 

pub enum RopeLayout {
    Norm,
    Neox,
}

Which RoPE pairing convention a model uses. Confirmed against llama.cpp’s llama_model_rope_type (src/llama-model.cpp): Norm is adjacent-pair / GPT-J (LLAMA_ROPE_TYPE_NORM); Neox is split-half / GPT-NeoX (LLAMA_ROPE_TYPE_NEOX). Getting this wrong silently produces fluent-but-wrong logits (the real Llama-3.1-8B early-stop bug: ferrox applied NeoX to a Norm architecture).

Variants

Norm

Adjacent pairs (2*i, 2*i+1) – llama.cpp LLAMA_ROPE_TYPE_NORM. Used by llama (including Llama 3/3.1/3.2), deepseek2, mistral3, and related families. (llama4 is DedicatedOnly — MoE graph — but its RoPE type in the inventory is still Norm.)

Neox

Split-half pairs (i, i+half) – llama.cpp LLAMA_ROPE_TYPE_NEOX. Used by olmoe, qwen2/qwen2moe/qwen3, phi3, gemma*, and related families. Ferrox’s historical default before architecture- aware dispatch existed.

Implementations

impl RopeLayout

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

Maps a GGUF general.architecture string onto the RoPE pairing llama.cpp selects for that family. Prefer crate::capability::resolve_architecture for load-time decisions — unknown architectures must fail closed there rather than guessing. This helper remains for tests and call sites that already know the arch is registered; unknowns still return Neox only as a last-resort historical default.

Trait Implementations

impl Clone for RopeLayout

fn clone(&self) -> RopeLayout

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for RopeLayout

impl Debug for RopeLayout

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

Formats the value using the given formatter.

impl Eq for RopeLayout

impl PartialEq for RopeLayout

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for RopeLayout