#DPO
2 related articles
·2 min
Random Pruning of the DPO Computation Graph: 1.4x Training Speedup and Lower Peak Memory
DPO training too slow and memory-hungry? This paper proposes a random pruning framework for the DPO computation graph, sampling independently across position, sequence, and branch dimensions; branch pruning achieves a 1.426x step-time speedup and cuts training time by about 30%, with in-depth attribution of memory and time gains.
Read more →
·2 min
The DPO Hyperparameter Landscape: How Scale Shifts the Optimum and Why It Doesn't Transfer Across Datasets
A systematic study uses dense hyperparameter grid experiments to reveal the DPO hyperparameter landscape: optimal learning rate shifts with model scale, hyperparameters do not transfer across datasets, validation loss is only weakly correlated with downstream benchmark optima, and short candidate training runs can significantly cut tuning costs.
Read more →