Transformers have emerged as the backbone of large language models (LLMs). However, generation remains inefficient due to the need to store in memory a cache of key-value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solution, we propose Dynamic Memory Compression (DMC), a method for on-line key-value cache compression at inference time. Most importantly, the model learns to apply different compression rates in different heads and layers. We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achieving up to ~3.7x throughput increase in auto-regressive inference on a NVIDIA H100 GPU. DMC is applied via continued pre-training on a negligible percentage of the original data without adding any extra parameters. We find that DMC preserves the original downstream performance with up to 4x cache compression, outperforming up-trained grouped-query attention (GQA). GQA and DMC can be even combined to obtain compounded gains. As a result DMC fits longer contexts and larger batches within any given memory budget.

Transformers achieve unrivalled performance in modelling language, but remain inefficient in terms of memory and time complexity. A possible remedy is to reduce the sequence length in the intermediate layers by pooling fixed-length segments of tokens. Nevertheless, natural units of meaning, such as words or phrases, display varying sizes. To address this mismatch, we equip language models with a dynamic-pooling mechanism, which predicts segment boundaries in an autoregressive fashion. We compare several methods to infer boundaries, including end-to-end learning through stochastic re-parameterisation, supervised learning (based on segmentations from subword tokenizers or spikes in conditional entropy), as well as linguistically motivated boundaries. We perform character-level evaluation on texts from multiple datasets and morphologically diverse languages. The results demonstrate that dynamic pooling, which jointly segments and models language, is both faster and more accurate than vanilla Transformers and fixed-length pooling within the same computational budget.

The success of deep learning comes from its ability to capture the hierarchical structure of data by learning high-level representations defined in terms of low-level ones. In this paper we explore self-supervised learning of hierarchical representations of speech by applying multiple levels of Contrastive Predictive Coding (CPC). We observe that simply stacking two CPC models does not yield significant improvements over single-level architectures. Inspired by the fact that speech is often described as a sequence of discrete units unevenly distributed in time, we propose a model in which the output of a low-level CPC module is non-uniformly downsampled to directly minimize the loss of a high-level CPC module. The latter is designed to also enforce a prior of separability and discreteness in its representations by enforcing dissimilarity of successive high-level representations through focused negative sampling, and by quantization of the prediction targets. Accounting for the structure of the speech signal improves upon single-level CPC features and enhances the disentanglement of the learned representations, as measured by downstream speech recognition tasks, while resulting in a meaningful segmentation of the signal that closely resembles phone boundaries.

In this paper we demonstrate methods for reliable and efficient training of discrete representation using Vector-Quantized Variational Auto-Encoder models (VQ-VAEs). Discrete latent variable models have been shown to learn nontrivial representations of speech, applicable to unsupervised voice conversion and reaching state-of-the-art performance on unit discovery tasks. For unsupervised representation learning, they became viable alternatives to continuous latent variable models such as the Variational Auto-Encoder (VAE). However, training deep discrete variable models is challenging, due to the inherent non-differentiability of the discretization operation. In this paper we focus on VQ-VAE, a state-of-the-art discrete bottleneck model shown to perform on par with its continuous counterparts. It quantizes encoder outputs with on-line $k$-means clustering. We show that the codebook learning can suffer from poor initialization and non-stationarity of clustered encoder outputs. We demonstrate that these can be successfully overcome by increasing the learning rate for the codebook and periodic date-dependent codeword re-initialization. As a result, we achieve more robust training across different tasks, and significantly increase the usage of latent codewords even for large codebooks. This has practical benefit, for instance, in unsupervised representation learning, where large codebooks may lead to disentanglement of latent representations.