Low-rank training has emerged as a promising approach for reducing memory usage in training Large Language Models (LLMs). Previous methods either rely on decomposing weight matrices (e.g., LoRA), or seek to decompose gradient matrices (e.g., GaLore) to ensure reduced memory consumption. However, both of them constrain the training in a low-rank subspace, thus inevitably leading to sub-optimal performance. To resolve this, we propose a new plug-and-play training framework for LLMs called Fira, as the first attempt to consistently preserve the low-rank constraint for memory efficiency, while achieving full-rank training (i.e., training with full-rank gradients of full-rank weights) to avoid inferior outcomes. First, we observe an interesting phenomenon during LLM training: the scaling impact of adaptive optimizers (e.g., Adam) on the gradient norm remains similar from low-rank to full-rank training.

Diffusion bridges are a promising class of deep-learning methods for sampling from unnormalized distributions. Recent works show that the Log Variance (LV) loss consistently outperforms the reverse Kullback-Leibler (rKL) loss when using the reparametrization trick to compute rKL-gradients. While the on-policy LV loss yields identical gradients to the rKL loss when combined with the log-derivative trick for diffusion samplers with non-learnable forward processes, this equivalence does not hold for diffusion bridges or when diffusion coefficients are learned. Based on this insight we argue that for diffusion bridges the LV loss does not represent an optimization objective that can be motivated like the rKL loss via the data processing inequality. Our analysis shows that employing the rKL loss with the log-derivative trick (rKL-LD) does not only avoid these conceptual problems but also consistently outperforms the LV loss. Experimental results with different types of diffusion bridges on challenging benchmarks show that samplers trained with the rKL-LD loss achieve better performance. From a practical perspective we find that rKL-LD requires significantly less hyperparameter optimization and yields more stable training behavior.

Our 8BMoE model achieves stronger pre-training perplexity than its dense counterpart. However, a better perplexity does not always directly translate to downstream performance as demonstrated in Section 4.4. To this end, we compare fine-tuning performance of the 8B dense model and MoE model in Table 1. As shown in the table, our MoE model using expert choice routing consistently outperforms the dense model across the 11 tasks in GLUE and SuperGLUE. We evaluate the downstream task fine-tuning performance by varying the capacity factors.

Convolutional neural networks were the standard for solving many computer vision tasks until recently, when Transformers of MLP-based architectures have started to show competitive performance. These architectures typically have a vast number of weights and need to be trained on massive datasets; hence, they are not suitable for their use in low-data regimes. In this work, we propose a simple yet effective framework to improve generalization from small amounts of data. We augment modern CNNs with fully-connected (FC) layers and show the massive impact this architectural change has in low-data regimes. We further present an online joint knowledge-distillation method to utilize the extra FC layers at train time but avoid them during test time. This allows us to improve the generalization of a CNN-based model without any increase in the number of weights at test time. We perform classification experiments for a large range of network backbones and several standard datasets on supervised learning and active learning. Our experiments significantly outperform the networks without fully-connected layers, reaching a relative improvement of up to 16% validation accuracy in the supervised setting without adding any extra parameters during inference.

It is currently difficult to distill discrete diffusion models. In contrast, continuous diffusion literature has many distillation approaches methods that can reduce sampling steps to a handful. Our method, Discrete Moment Matching Distillation (D-MMD), leverages ideas that have been highly successful in the continuous domain. Whereas previous discrete distillation methods collapse, D-MMD maintains high quality and diversity (given sufficient sampling steps). This is demonstrated on both text and image datasets. Moreover, the newly distilled generators can outperform their teachers.

In many reinforcement learning (RL) applications, incorporating heuristic rewards alongside the task reward is crucial for achieving desirable performance. Heuristics encode prior human knowledge about how a task should be done, providing valuable hints for RL algorithms. However, such hints may not be optimal, limiting the performance of learned policies. The currently established way of using heuristics is to modify the heuristic reward in a manner that ensures that the optimal policy learned with it remains the same as the optimal policy for the task reward (i.e., optimal policy invariance). However, these methods often fail in practical scenarios with limited training data. We found that while optimal policy invariance ensures convergence to the best policy based on task rewards, it doesn't guarantee better performance than policies trained with biased heuristics under a finite data regime, which is impractical. In this paper, we introduce a new principle tailored for finite data settings. Instead of enforcing optimal policy invariance, we train a policy that combines task and heuristic rewards and ensures it outperforms the heuristic-trained policy. As such, we prevent policies from merely exploiting heuristic rewards without improving the task reward.