Struct Half 

pub struct Half {
    pub shape: Vec<usize>,
    pub dtype: DType,
    /* private fields */
}

A half-precision tensor stored packed (2×16-bit per 32-bit word) in GPU memory.

Fields

shape: Vec<usize>

dtype: DType

Implementations

impl Half

pub fn numel(&self) -> usize

pub fn nbytes(&self) -> usize

Bytes actually stored on device (half of the f32 equivalent).

pub fn from_bits( ctx: &Arc<Context>, bits: &[u16], shape: &[usize], dtype: DType, ) -> Half

Build from raw 16-bit values (e.g. an fp16/bf16 slice straight out of a safetensors file).

pub fn dequant(&self) -> Tensor

Dequantize to an f32 compute tensor, on-device.