Achintya Kundu

Achintya Kundu, Rhui Dih Lee, Laura Wynter et al.

Padding is often used in tuning LLM models by adding special tokens to shorter training examples to match the length of the longest sequence in each batch. While this ensures uniformity for batch processing, it introduces inefficiencies by including irrelevant padding tokens in the computation and wastes GPU resources. Hugging Face SFT trainer has always offered the option to use packing to combine multiple training examples, allowing for maximal utilization of GPU resources. However, up till now, it did not offer proper masking of each packed training example. This capability has been added to Hugging Face Transformers 4.44. We analyse this new feature and show the benefits across different variations of packing.

Achintya Kundu, Laura Wynter, Rhui Dih Lee et al.

Weight-sharing neural architecture search aims to optimize a configurable neural network model (supernet) for a variety of deployment scenarios across many devices with different resource constraints. Existing approaches use evolutionary search to extract models of different sizes from a supernet trained on a very large data set, and then fine-tune the extracted models on the typically small, real-world data set of interest. The computational cost of training thus grows linearly with the number of different model deployment scenarios. Hence, we propose Transfer-Once-For-All (TOFA) for supernet-style training on small data sets with constant computational training cost over any number of edge deployment scenarios. Given a task, TOFA obtains custom neural networks, both the topology and the weights, optimized for any number of edge deployment scenarios. To overcome the challenges arising from small data, TOFA utilizes a unified semi-supervised training loss to simultaneously train all subnets within the supernet, coupled with on-the-fly architecture selection at deployment time.

Achintya Kundu, Fabian Lim, Aaron Chew et al.

Supernet training of LLMs is of great interest in industrial applications as it confers the ability to produce a palette of smaller models at constant cost, regardless of the number of models (of different size / latency) produced. We propose a new method called Multistage Low-rank Fine-tuning of Super-transformers (MLFS) for parameter-efficient supernet training. We show that it is possible to obtain high-quality encoder models that are suitable for commercial edge applications, and that while decoder-only models are resistant to a comparable degree of compression, decoders can be effectively sliced for a significant reduction in training time.

Achintya Kundu, Pengqian Yu, Laura Wynter et al.

We present a class of methods for robust, personalized federated learning, called Fed+, that unifies many federated learning algorithms. The principal advantage of this class of methods is to better accommodate the real-world characteristics found in federated training, such as the lack of IID data across parties, the need for robustness to outliers or stragglers, and the requirement to perform well on party-specific datasets. We achieve this through a problem formulation that allows the central server to employ robust ways of aggregating the local models while keeping the structure of local computation intact. Without making any statistical assumption on the degree of heterogeneity of local data across parties, we provide convergence guarantees for Fed+ for convex and non-convex loss functions under different (robust) aggregation methods. The Fed+ theory is also equipped to handle heterogeneous computing environments including stragglers without additional assumptions; specifically, the convergence results cover the general setting where the number of local update steps across parties can vary. We demonstrate the benefits of Fed+ through extensive experiments across standard benchmark datasets.