Struct Tensor 

pub struct Tensor { /* private fields */ }

Tensor with automatic differentiation support

Data is stored behind Rc for O(1) clone. Backward ops clone input tensors to hold references for gradient computation — with Rc, this is a reference-count bump instead of a full memcpy. For Qwen3-4B with batch_size=8, this eliminates ~16 GB of redundant frozen-weight copies per training step (KAIZEN-019).

data_mut() uses Rc::make_mut for copy-on-write: mutation only copies when there are multiple Rc references to the same data.

Implementations

impl Tensor

pub fn new(data: Array1<f32>, requires_grad: bool) -> Self

Create a new tensor with data

pub fn from_vec(data: Vec<f32>, requires_grad: bool) -> Self

Create a tensor from a vector

pub fn zeros(size: usize, requires_grad: bool) -> Self

Create a tensor filled with zeros

pub fn ones(size: usize, requires_grad: bool) -> Self

Create a tensor filled with ones

pub fn data(&self) -> &Array1<f32>

Get reference to data

pub fn data_mut(&mut self) -> &mut Array1<f32>

Get mutable reference to data (copy-on-write via Rc::make_mut)

If this is the only Rc reference, returns a mutable reference to the existing data with no copy. If there are other references (e.g. backward ops holding clones), clones the data first so mutations don’t affect other holders.

pub fn grad(&self) -> Option<Array1<f32>>

Get gradient (if computed)

pub fn set_grad(&self, grad: Array1<f32>)

Set gradient

pub fn accumulate_grad(&self, grad: Array1<f32>)

Accumulate gradient (for when tensor is used multiple times)

pub fn zero_grad(&self)

Zero out gradient

pub fn scale_grad(&self, factor: f32)

Scale gradient in-place (KAIZEN-037: zero-allocation gradient scaling)

Contract (C-SCALE-GRAD-001)

• Precondition: factor is finite
• Postcondition: grad[i] *= factor for all i
• Invariant: No heap allocation (mutates existing Array1 in-place)

pub fn requires_grad(&self) -> bool

Check if requires gradient

pub fn set_requires_grad(&mut self, requires_grad: bool)

Set requires gradient flag

Use this to enable gradient tracking on loaded tensors before training.

pub fn grad_cell(&self) -> Rc<RefCell<Option<Array1<f32>>>>

Get reference to gradient cell (for backward operations)

pub fn set_backward_op(&mut self, op: Rc<dyn BackwardOp>)

Set backward operation

pub fn backward_op(&self) -> Option<Rc<dyn BackwardOp>>

Get backward operation

pub fn len(&self) -> usize

Get size

pub fn is_empty(&self) -> bool

Check if empty

Trait Implementations

impl Clone for Tensor

fn clone(&self) -> Tensor

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.