Outputs for one DSpark training forward. Shape symbols: batch_size: number of samples in the batch seq_len: source sequence lengt

Overlaps DataLoader iteration and H2D transfer with compute. Uses a background thread and a dedicated CUDA stream so that both next(dataloade

Per-position ECE + AUROC + Brier for cumprod predictions.

RNN-based head that maintains recurrent state across positions within a block. Unlike the memoryless Markov heads, position k can access the full

Deterministic distributed sampler that streams across epoch boundaries. Each epoch uses the first ``total_size`` samples from a deterministic

Soft cross-entropy with fused Triton forward/backward. Returns ``sum_i(-sum_v target_p[i,v] * log_softmax(logits)[i,v]) / normalizer``, where