torch.GradContext

export interface GradContext<_Op extends OpName = OpName> {
  /**
   * Save tensors for use in backward pass.
   *
   * @param tensors - Tensors to save
   */
  save_for_backward(...tensors: TensorLike[]): void;

  /**
   * Tensors saved during forward pass.
   * Available in backward().
   */
  readonly saved_tensors: readonly TensorLike[];

  /**
   * Additional data storage for custom values.
   * Use this for non-tensor data like shapes, indices, etc.
   */
  [key: string]: unknown;
}

_Opextends OpName

save_for_backward((...tensors: TensorLike[]) => void): – Save tensors for use in backward pass.
readonlysaved_tensors(readonly TensorLike[]): – Tensors saved during forward pass. Available in backward().

Context for gradient computation.

The context is passed to setup_context during forward pass and to backward during backward pass. It stores tensors needed for gradient computation.

torch.GradContext

export interface GradContext<_Op extends OpName = OpName> {
  /**
   * Save tensors for use in backward pass.
   *
   * @param tensors - Tensors to save
   */
  save_for_backward(...tensors: TensorLike[]): void;

  /**
   * Tensors saved during forward pass.
   * Available in backward().
   */
  readonly saved_tensors: readonly TensorLike[];

  /**
   * Additional data storage for custom values.
   * Use this for non-tensor data like shapes, indices, etc.
   */
  [key: string]: unknown;
}

_Opextends OpName

save_for_backward((...tensors: TensorLike[]) => void): – Save tensors for use in backward pass.
readonlysaved_tensors(readonly TensorLike[]): – Tensors saved during forward pass. Available in backward().

Context for gradient computation.

The context is passed to setup_context during forward pass and to backward during backward pass. It stores tensors needed for gradient computation.