Enum GgmlDType 

pub enum GgmlDType {
    F32,
    F16,
    BF16,
    Q4_0,
    Q4_1,
    Q5_0,
    Q5_1,
    Q8_0,
    Q8_1,
    Q2K,
    Q3K,
    Q4K,
    Q5K,
    Q6K,
    Q8K,
    Iq4Nl,
    Iq4Xs,
    Iq2Xxs,
    Iq2Xs,
    Iq3Xxs,
    Iq1S,
    Iq3S,
    Iq2S,
    Iq1M,
    Tq1_0,
    Tq2_0,
    Mxfp4,
    Nvfp4,
}

The ggml block-format dtype tag. Mirrors llama.cpp’s ggml_type @9d57ce456c for the subset fuel supports (the classic 15 — F32/F16/BF16 plus Q4_0..Q8_1 and Q2K..Q8K, plus every IQ family: IQ4_NL/IQ4_XS (GAP-211 B1) and the decode-only IQ1_S/IQ1_M/IQ2_XXS/IQ2_XS/IQ2_S/IQ3_XXS/IQ3_S (GAP-211 B2); block geometry and the dequant formula verified against ggml-common.h and ggml-quants.c at that commit), plus the decode-only TQ1_0/TQ2_0/MXFP4/NVFP4 (GAP-211 C); NOT the other ggml types, which an unsupported-tensor GGUF load reports by name instead).

Geometry checked against that commit (GAP-248 instance 5), and the referee differs by column:

• type_size (bytes/block) is ADJUDICATED — upstream’s own static_assert on each block struct is the referee; 12/12 for the quant types here.
• block_size (elems/block) and the numeric codes are CORROBORATED — both are transcribed upstream (type_traits[].blck_size, enum ggml_type) and agree 12/12 with MLMF’s INDEPENDENT transcription; no referee exists.
• per-block alignment is UNANCHORED by both projects; fuel tracks none, so an upstream alignment change is undetectable here.

NO DETECTOR: this pin makes drift ANSWERABLE, not detected — it ages silently. Re-run the comparison on any llama.cpp bump that touches ggml_type or type_traits[].

Lives here (rather than in quantized/mod.rs) because per-backend kernel crates need to name it without depending on fuel-core.

Variants

F32

F16

BF16

Q4_0

Q4_1

Q5_0

Q5_1

Q8_0

Q8_1

Q2K

Q3K

Q4K

Q5K

Q6K

Q8K

Iq4Nl

llama.cpp GGML_TYPE_IQ4_NL (20): 32 elems/block, non-linear 4-bit codebook.

Iq4Xs

llama.cpp GGML_TYPE_IQ4_XS (23): 256 elems/block, 4-bit codebook, 6-bit sub-block scales.

Iq2Xxs

llama.cpp GGML_TYPE_IQ2_XXS (16): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Iq2Xs

llama.cpp GGML_TYPE_IQ2_XS (17): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Iq3Xxs

llama.cpp GGML_TYPE_IQ3_XXS (18): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Iq1S

llama.cpp GGML_TYPE_IQ1_S (19): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Iq3S

llama.cpp GGML_TYPE_IQ3_S (21): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Iq2S

llama.cpp GGML_TYPE_IQ2_S (22): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Iq1M

llama.cpp GGML_TYPE_IQ1_M (29): 256 elems/block, grid-table quantization; DECODE-ONLY (GAP-211 B2).

Tq1_0

llama.cpp GGML_TYPE_TQ1_0 (34): 256 elems/block, ternary (5 trits/byte); DECODE-ONLY (GAP-211 C).

Tq2_0

llama.cpp GGML_TYPE_TQ2_0 (35): 256 elems/block, ternary (2 bits/elem); DECODE-ONLY (GAP-211 C).

Mxfp4

llama.cpp GGML_TYPE_MXFP4 (39): 32 elems/block, E2M1 with an E8M0 block scale; DECODE-ONLY (GAP-211 C).

Nvfp4

llama.cpp GGML_TYPE_NVFP4 (40): 64 elems/block, E2M1 with four UE4M3 sub-block scales; DECODE-ONLY (GAP-211 C).

Implementations

impl GgmlDType

pub fn from_u32(u: u32) -> Result<GgmlDType, Error>

pub fn to_u32(self) -> u32

pub fn type_size(&self) -> usize

pub fn block_size(&self) -> usize

pub fn supports_quantize(self) -> bool

false for the grid-table IQ formats (GAP-211 B2) and the TQ/MXFP4/NVFP4 formats (GAP-211 C): Fuel decodes them but has no quantizer for them.

pub fn check_quantize(self) -> Result<(), Error>

Declines quantizing to a decode-only format with a typed error that names it, instead of reaching GgmlType::from_float, which has nothing to run.

pub fn check_dequant_count( self, op_label: &str, numel: usize, byte_len: usize, ) -> Result<(), Error>

GAP-333: numel dequantized elements are read from numel / block_size whole blocks of type_size bytes. Declines a count that is not a whole number of blocks, or whose blocks do not fit in byte_len.

The one copy every backend’s dequantize calls before it reads, so the predicate cannot drift between them.

Trait Implementations

impl Clone for GgmlDType

fn clone(&self) -> GgmlDType

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for GgmlDType

impl Debug for GgmlDType

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

Formats the value using the given formatter.

impl Eq for GgmlDType

impl Hash for GgmlDType

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 GgmlDType

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for GgmlDType