Function dsparkLoss

dsparkLoss(
    out: DflashTrainOut,
    targetLogits: MlxArray,
    xStar: MlxArray,
    gamma: number,
    w: MlxArray,
    weights?: LossWeights,
): { ce: MlxArray; conf: MlxArray; loss: MlxArray; tv: MlxArray }
Full DSpark loss. Returns the scalar to differentiate plus the three term scalars (for logging — they share the graph; read them with .toFloat32() AFTER the backward, or eval them alongside the value).
- out.draftLogits [A,γ,V] (U+B, pre-softmax, float32)
- out.conf [A,γ] (sigmoid output ∈ (0,1))
- targetLogits [A,γ,V] (frozen target logits; constant)
- xStar [A,γ] int32 ground-truth next tokens
Parameters
- out: DflashTrainOut
- targetLogits: MlxArray
- xStar: MlxArray
- gamma: number
- w: MlxArray
- weights: LossWeights = DEFAULT_LOSS_WEIGHTS
Returns { ce: MlxArray; conf: MlxArray; loss: MlxArray; tv: MlxArray }
- Defined in packages/training/src/dspark/loss.ts:75