Enum TernaryFormat 

#[non_exhaustive]
pub enum TernaryFormat {
    Tq1_0,
    Tq2_0,
    I2sInt8,
}

Canonical packing schemes for ternary weights — the shared vocabulary across crates. The byte-exact layout, pack, and unpack live in tritium-format; this enum only fixes the names, block size, and effective width so every crate agrees on them.

Both follow ggml’s 256-element super-block with one fp16 block scale.

Variants (Non-exhaustive)

Non-exhaustive enums could have additional variants added in future. Therefore, when matching against variants of non-exhaustive enums, an extra wildcard arm must be added to account for any future variants.

Tq1_0

llama.cpp TQ1_0 — 5 trits per byte (3^5 = 243 < 256). 1.6875 bpw incl. block scale. Most compact; unpack needs base-3 division — CPU/edge oriented.

Tq2_0

llama.cpp TQ2_0 — 2 bits per trit, 4 per byte. 2.0625 bpw incl. block scale. Cheap shift/mask unpack; matches BitNet’s int8 packing — GPU oriented.

I2sInt8

I2sInt8 — the IMMA (int8 tensor-core) GPU layout, derived from an I2_S checkpoint at load (tritium_format::convert_i2s_to_int8). Weights stay 2-bit packed in VRAM (≈2.0 bpw, no per-block scale — the I2_S source carries a single per-tensor f32 scale), interleaved so the IMMA kernel can unpack them to int8 operands for mma.m16n8k32 in shared memory. Added in v0.30 (ADR 0005) for the compute-bound prefill path; the CPU/reference backends do not consume it and return an unsupported-format error — it is a GPU-only packing.

Implementations

impl TernaryFormat

pub const fn bits_per_weight(self) -> f32

Effective bits per weight, block scale amortized over the block.

pub const fn block_size(self) -> usize

Weights per quantization block (ggml QK_K). One fp16 scale per block.

pub const fn is_bit_aligned(self) -> bool

Whether unpack is a cheap shift/mask (true, GPU-friendly) or base-3 division (false, CPU/edge). Both Tq2_0 and I2sInt8 are 2-bit aligned.

Trait Implementations

impl Clone for TernaryFormat

fn clone(&self) -> TernaryFormat

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for TernaryFormat

impl Debug for TernaryFormat

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

Formats the value using the given formatter.

impl Display for TernaryFormat

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

Short canonical token per variant (e.g. tq1_0, tq2_0).

impl Eq for TernaryFormat

impl Hash for TernaryFormat

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for TernaryFormat

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for TernaryFormat