Enum GgufBlockFormat 

#[non_exhaustive]
#[repr(u16)]
pub enum GgufBlockFormat {
    Q4_0 = 2,
    Q4_1 = 3,
    Q5_0 = 6,
    Q5_1 = 7,
    Q8_0 = 8,
    Q2K = 10,
    Q3K = 11,
    Q4K = 12,
    Q5K = 13,
    Q6K = 14,
    Q8K = 15,
}

GGUF block-format selector for QuantizeKind::GgufDequantize / QuantizeKind::GgufMmvq plans. Mirrors the discriminants used by llama.cpp / ggml so a descriptor can be round-tripped to a GGUF file header without translation.

Block sizes:

• Type-0/1 variants (Q4_0, Q4_1, Q5_0, Q5_1, Q8_0) pack 32 quantized values per block plus a shared FP scale (+ min for the _1 variants).
• k-quants variants (Q2_K … Q8_K) pack 256 values per super-block with a multi-level scale hierarchy (quantized sub-block scales + FP super-block scale).

Discriminant values match the GGML_TYPE_* enum in upstream ggml.h, ensuring binary compatibility with GGUF file headers.

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.

Q4_0 = 2

4-bit, 32-element block, single FP scale. block_q4_0.

Q4_1 = 3

4-bit, 32-element block, FP scale + FP min. block_q4_1.

Q5_0 = 6

5-bit, 32-element block, single FP scale. block_q5_0.

Q5_1 = 7

5-bit, 32-element block, FP scale + FP min. block_q5_1.

Q8_0 = 8

8-bit, 32-element block, single FP scale. block_q8_0.

Q2K = 10

2.5-bit (effective), 256-element super-block. block_q2_K.

Q3K = 11

3.4-bit (effective), 256-element super-block. block_q3_K.

Q4K = 12

4.5-bit (effective), 256-element super-block. block_q4_K.

Q5K = 13

5.5-bit (effective), 256-element super-block. block_q5_K.

Q6K = 14

6.6-bit (effective), 256-element super-block. block_q6_K.

Q8K = 15

8-bit, 256-element super-block (CPU-side intermediate). block_q8_K. Dequant supported; MMVQ NOT supported (matches llama.cpp — no upstream MMVQ specialization).

Implementations

impl GgufBlockFormat

pub const fn block_size(self) -> usize

Number of FP elements per packed block.

pub const fn type_size(self) -> usize

Number of bytes per packed block. Matches sizeof(block_q*) from ggml.h. Used by the Rust plan layer to size the input weight buffer.

pub const fn is_type_01(self) -> bool

true for the type-0/1 family (32-element blocks); false for the k-quants family (256-element super-blocks).

pub const fn has_mmvq(self) -> bool

true if MMVQ (fused dequant + matvec) is supported for this block format. As of Phase 11.4, all 11 GGUF block formats ship a MMVQ kernel. Q8_K MMVQ is a bespoke baracuda addition (upstream llama.cpp / Fuel reserve Q8_K as a CPU-side intermediate and ship dequant only); we close that gap to avoid 2× memory traffic on the inference decode step.

Trait Implementations

impl Clone for GgufBlockFormat

fn clone(&self) -> GgufBlockFormat

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for GgufBlockFormat

impl Debug for GgufBlockFormat

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

Formats the value using the given formatter.

impl Eq for GgufBlockFormat

impl Hash for GgufBlockFormat

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

where __H: Hasher,

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 GgufBlockFormat

fn eq(&self, other: &GgufBlockFormat) -> 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 StructuralPartialEq for GgufBlockFormat