Enum WeightDtype 

pub enum WeightDtype {
    BF16,
    FP32,
    FP8E4M3,
    FP8E8M0,
    UInt8,
    Int64,
    PackedQ2_0 {
        group: u16,
    },
}

Data type of a weight tensor.

Variants

BF16

FP32

FP8E4M3

FP8E8M0

UInt8

Int64

PackedQ2_0

Keep-packed PrismML ternary Q2_0 (ggml id 42): raw on-disk blocks stay 2-bit in VRAM (fp16 scale + 2-bit codes per group of group elements), dequantized in-kernel by the native q2_0_gemv decode path. Only produced by the GGUF loader under ATLAS_GGUF_NATIVE_Q2=1. Its byte footprint is NOT a per-element size (2-bit codes + an inline scale per group), so WeightDtype::byte_size returns 0 for this variant and the real size is computed in WeightTensor::byte_size (shape + group).

Fields

group: u16

Implementations

impl WeightDtype

pub fn byte_size(self) -> usize

Bytes per element for the fixed-width dtypes. Returns 0 for the block-based WeightDtype::PackedQ2_0 — WeightTensor::byte_size handles that variant directly, and no caller multiplies its numel by this.

pub fn from_safetensors_str(s: &str) -> Result<Self>

Map a raw safetensors header dtype STRING (as it appears in the JSON header, e.g. "BF16", "F8_E4M3") to a WeightDtype, factored out so the RDMA weight loader (which receives dtype as a wire string in the peer manifest, not a safetensors::Dtype) resolves it identically to the disk loaders — byte-identity depends on the two ends agreeing.

Trait Implementations

impl Clone for WeightDtype

fn clone(&self) -> WeightDtype

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for WeightDtype

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

Formats the value using the given formatter.

impl PartialEq for WeightDtype

fn eq(&self, other: &WeightDtype) -> 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 Copy for WeightDtype

impl Eq for WeightDtype

impl StructuralPartialEq for WeightDtype