Struct Tensor 

pub struct Tensor {
    pub name: String,
    /* private fields */
}

A tensor in a checkpoint, with its bytes not yet read.

Opening a checkpoint parses its metadata only. read is what produces the tensor’s bytes, so a checkpoint can be inspected without loading it, or loaded one tensor at a time. ZIP and legacy containers read the file at that point; a TAR container’s storages are held in memory from open.

The element type and shape are fixed at parse time, when the reader checks that the tensor fits its storage, so they are read-only: byte_len is what read returns because nothing can move them apart.

Fields

name: String

Dotted path of the tensor through the checkpoint’s dictionaries ("encoder.0.weight"). Free to rename; only error messages look at it.

Implementations

impl Tensor

pub fn dtype(&self) -> DType

Element type.

pub fn shape(&self) -> &[usize]

Extents, outermost first. Empty for a scalar.

pub fn num_elements(&self) -> usize

Number of elements.

pub fn byte_len(&self) -> usize

Number of bytes read returns.

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

Read the tensor’s bytes from the checkpoint.

The bytes are contiguous in row-major order and native-endian, whatever layout the file stored the tensor in: a strided or offset view is gathered, and a bool tensor holds only 0 and 1. Nothing is cached, so each call reads the storage again.

Reads from different threads run at once: a ZIP read holds the archive lock only to locate its entry. A compressed entry, which torch.save does not write, streams under the lock instead, as does every entry on a target other than Unix.

Errors

A storage that is missing, corrupt or unreadable by its container fails with io::ErrorKind::InvalidData (a ZIP storage’s checksum is verified when the tensor spans all of it, unless the file was saved without checksums, as compute_crc32=False does), and one shorter than the pickle declared with io::ErrorKind::UnexpectedEof: the file disagrees with itself. Any other kind is the operating system’s, and means the file could not be read at all. Every error names the tensor and its storage.

pub fn transposed(&self) -> Option<Tensor>

The transpose of a 2-D tensor the checkpoint stores transposed.

A .t() view, which is how nn.Linear weights are often saved, is stored as the row-major bytes of its transpose. read has to gather those into the view’s own order, while the tensor returned here, with the reversed shape, reads them as a straight copy. A consumer that wants the transpose anyway can skip both passes.

None for every other tensor, including one whose transpose would itself need gathering. The returned tensor has the same name and reads and fails as read does; its own transposed is None.

pub fn into_parts( self, ) -> (String, DType, Vec<usize>, impl Fn() -> Result<Vec<u8>> + Send + Sync + 'static)

Split the tensor into its metadata and its read handle.

For a caller that keeps the metadata elsewhere and wants only the handle, without a second copy of the name and shape held alive with it. The handle reads and fails exactly as read does.

Trait Implementations

impl Clone for Tensor

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Tensor

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

Formats the value using the given formatter.