Struct TensorSnapshot 

pub struct TensorSnapshot {
    pub dtype: DType,
    pub shape: Vec<usize>,
    pub path_stack: Option<Vec<String>>,
    pub container_stack: Option<Vec<String>>,
    pub tensor_id: Option<ParamId>,
    /* private fields */
}

A lightweight snapshot of a tensor that can lazily produce TensorData.

TensorSnapshot stores a cloned tensor internally (which is cheap due to reference counting) and only materializes the actual data when to_data() is called. This allows efficient inspection of module structure without the overhead of copying all tensor data.

The dtype and shape are cached for efficient access without requiring data materialization, which is particularly useful for serialization formats that need metadata upfront.

Fields

dtype: DType

Data type of the tensor (cached for efficient access)

shape: Vec<usize>

Shape of the tensor (cached for efficient access)

path_stack: Option<Vec<String>>

Path stack representing the module hierarchy

container_stack: Option<Vec<String>>

Container stack representing the container types at each level

tensor_id: Option<ParamId>

Unique identifier for the tensor parameter

Implementations

impl TensorSnapshot

pub fn from_float<B: Backend, const D: usize>( tensor: &Tensor<B, D>, path_stack: Vec<String>, container_stack: Vec<String>, tensor_id: ParamId, ) -> Self

Create a new tensor snapshot from a float tensor

pub fn from_int<B: Backend, const D: usize>( tensor: &Tensor<B, D, Int>, path_stack: Vec<String>, container_stack: Vec<String>, tensor_id: ParamId, ) -> Self

Create a new tensor snapshot from an int tensor

pub fn from_bool<B: Backend, const D: usize>( tensor: &Tensor<B, D, Bool>, path_stack: Vec<String>, container_stack: Vec<String>, tensor_id: ParamId, ) -> Self

Create a new tensor snapshot from a bool tensor

pub fn to_data(&self) -> Result<TensorData, TensorSnapshotError>

Convert to TensorData (this is where actual data copy happens)

pub fn full_path(&self) -> String

Get the full path by joining the path stack

pub fn container_path(&self) -> String

Get the full container path by joining the container stack

pub fn module_type(&self) -> Option<String>

Get the module type (last Struct/Enum in the hierarchy)

Returns the last user-defined module type, skipping primitive containers like “Vec”, “Array”. This is useful for determining which user-defined module a tensor belongs to.

Examples

• Linear.weight → Some("Struct:Linear")
• Vec<Linear>[0].weight → Some("Struct:Linear")
• Linear.bias (Optional) → Some("Struct:Linear")
• Vec<Param>[0] (no module) → None

pub fn container_type(&self) -> String

Get the immediate container type (last in the container stack)

Returns the last element in the container stack, which could be a user-defined type (“Struct:”, “Enum:”) or a collection type (“Vec”, “Array”). This is useful for understanding the full container hierarchy.

Examples

• Linear.weight → "Struct:Linear"
• Vec<Linear>[0].weight → "Struct:Linear" (the Linear, not the Vec)
• Vec<Param>[0] → "Vec"

pub fn from_closure( data_fn: Rc<dyn Fn() -> Result<TensorData, TensorSnapshotError>>, dtype: DType, shape: Vec<usize>, path_stack: Vec<String>, container_stack: Vec<String>, tensor_id: ParamId, ) -> Self

Create a TensorSnapshot from a closure that produces TensorData This is used internally for lazy loading

pub fn from_data( data: TensorData, path_stack: Vec<String>, container_stack: Vec<String>, tensor_id: ParamId, ) -> Self

Create a TensorSnapshot from TensorData directly

pub fn data_len(&self) -> usize

Get the size of the tensor data in bytes without materializing it.

For regular (non-quantized) types, this is simply shape.product() * dtype.size().

For quantized types (QFloat), this accounts for:

• The quantized values (packed according to the quantization scheme)
• Alignment padding (values are aligned to 4-byte boundary)
• Quantization parameters (scale values appended to the data)

pub fn clone_data_fn( &self, ) -> Rc<dyn Fn() -> Result<TensorData, TensorSnapshotError>>

Clone the data function for lazy composition

Trait Implementations

impl Clone for TensorSnapshot

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 TensorSnapshot

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

Formats the value using the given formatter.