Struct GradWithLossOptions 

pub struct GradWithLossOptions {
    pub zero_missing_wrt: bool,
}

Compute the reverse-mode gradient graph and the loss value.

Returns a graph whose outputs are [loss, aux₁, aux₂, …, grad_wrt[0], grad_wrt[1], …].

The first forward output is treated as the scalar loss and is differentiated. Any additional forward outputs (aux₁ …) are mirrored from the forward graph and emitted unchanged — gradients are not propagated through them. This is the canonical hook for emitting training-side statistics (BatchNorm batch mean/variance, debug probes, …) alongside the loss in a single forward+backward pass.

The returned graph contains a copy of the entire forward graph so activations needed by gradient kernels are recomputed from inputs; it also exposes a new Op::Input named "d_output" which the caller seeds with the upstream gradient of the loss (typically a scalar 1.0 for “differentiate the loss directly”). Auxiliary outputs have no d_output-equivalent — by construction they don’t contribute to the gradient path.

Limitations

• Forward graph must have ≥ 1 output. The first is the loss.
• All ops in the forward graph must have an implemented VJP rule. Hitting an op without one is a panic, not a silent miscompute. Options for grad_with_loss_opts.

Fields

zero_missing_wrt: bool

When true, parameters in wrt with no gradient path receive an explicit zero tensor instead of panicking (e.g. unused logit_bias).

Implementations

impl GradWithLossOptions

pub const STRICT: GradWithLossOptions

pub const TRAINING: GradWithLossOptions

Trait Implementations

impl Clone for GradWithLossOptions

fn clone(&self) -> GradWithLossOptions

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for GradWithLossOptions

impl Debug for GradWithLossOptions

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

Formats the value using the given formatter.

impl Eq for GradWithLossOptions

impl PartialEq for GradWithLossOptions

fn eq(&self, other: &GradWithLossOptions) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for GradWithLossOptions