Struct DeferredTensor 

pub struct DeferredTensor {
    pub path: PathBuf,
    pub offset: u64,
    pub shape: Vec<usize>,
    pub dtype: WeightDtype,
}

Where a skipped tensor lives, so a consumer can read it in place.

Fields

path: PathBuf

Shard file containing the tensor.

offset: u64

ABSOLUTE byte offset of the tensor’s first element in that file (safetensors header length + the tensor’s data_offsets[0]).

shape: Vec<usize>

dtype: WeightDtype

Implementations

impl DeferredTensor

pub fn byte_size(&self) -> usize

The tensor’s on-disk footprint.

Only the fixed-width dtypes a safetensors header can carry reach here, so numel * byte_size is exact — the block-based GGUF dtypes (WeightDtype::PackedQ2_0 and the K-quants) report 0 bytes/element and are never deferred by either rule.

pub fn read_host_bytes(&self) -> Result<Vec<u8>>

Read the whole tensor’s raw bytes into host memory.

The consumer for a deferred tensor is usually a row cache that faults single rows off NVMe; this is the other shape of consumer — a loader that needs the tensor ONCE, at bind time, to derive the buffer it will actually upload.

🪤 read_exact rather than read_to_end: a short read means the shard is truncated or the offset is wrong, and both must be an error here rather than a shorter tensor whose element count is rejected later with a message pointing at the wrong thing. Seek-then-read rather than pread, because WeightStore builds on Windows too.

Trait Implementations

impl Clone for DeferredTensor

fn clone(&self) -> DeferredTensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DeferredTensor

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

Formats the value using the given formatter.