class

AutogradMeta

AutogradMeta

source edit

Per-tensor autograd metadata — gradient buffer, backward edge, and version counter for in-place safety.

requires_grad controls whether operations involving this tensor build an autograd graph. is_leaf is true for parameters created by the user (model weights) and false for intermediate tensors produced by ops. Leaf tensors accumulate gradients into grad; non-leaf tensors do not accumulate unless the engine later honours a retain_grad opt-in (tracked in the sibling core::AutogradMeta, not yet exposed in this struct).

grad_fn points to the backward Node that produced this tensor. It is nullptr for leaf tensors (which use AccumulateGrad installed lazily by ensure_grad_fn) and for tensors detached from the graph via detach.

See Also

AutogradNode — backward node type referenced by grad_fn. AccumulateGrad — terminal node that writes into grad for leaf tensors.

Attributes

requires_gradbool

Whether autograd tracks this tensor. When true, downstream ops emit an AutogradNode and link it into grad_fn of their outputs.

is_leafbool

true for user-created parameters (no producing op); false for op outputs. Leaves install AccumulateGrad as their grad_fn on first backward and write into grad.

versionstd::int64_t

Monotonically increasing in-place mutation counter. Bumped by TensorImpl::bump_version on every in-place op or optimizer step. Saved tensors snapshot this value at capture time and compare it during backward.

grad_fnNodePtr

Backward function (autograd graph edge) that produced this tensor. nullptr for leaves and detached tensors.

gradstd::optional<Storage>

Accumulated gradient storage, populated on leaves after a normal backward call. Lazily allocated by AccumulateGrad on first invocation; cleared by zero_grad.

grad_implstd::shared_ptr<TensorImpl>

Full TensorImpl view of the gradient — populated instead of (or in addition to) grad when backward(create_graph=true) is used. Allows the gradient itself to participate in further autograd operations (second-order derivatives, MAML, Hessian-vector products, …).

Notes

Invariants enforced by the autograd layer:

is_leaf == true ⇒ grad_fn == nullptr.
- is_leaf == false ⇒ grad_fn != nullptr.
- version is monotonically non-decreasing for the lifetime of the owning TensorImpl.

Lifecycle: created at first parameter use (when requires_grad is set), persists through forward() / backward() cycles, and grad is cleared by optimizer.zero_grad() between iterations.