Large language models (LLMs) with billions of parameters and pretrained on massive amounts of data are now capable of near or better than state-of-the-art performance in a variety of downstream natural language processing tasks. Neural machine translation (NMT) is one such task that LLMs have been applied to with great success. However, little research has focused on applying LLMs to the more difficult subset of NMT called simultaneous translation (SimulMT), where translation begins before the entire source context is available to the model. In this paper, we address key challenges facing LLMs fine-tuned for SimulMT, validate classical SimulMT concepts and practices in the context of LLMs, explore adapting LLMs that are fine-tuned for NMT to the task of SimulMT, and introduce Simul-LLM, the first open-source fine-tuning and evaluation pipeline development framework for LLMs focused on SimulMT.
Simultaneous speech translation is an essential communication task difficult for humans whereby a translation is generated concurrently with oncoming speech inputs. For such a streaming task, transformers using block processing to break an input sequence into segments have achieved state-of-the-art performance at a reduced cost. Current methods to allow information to propagate across segments, including left context and memory banks, have faltered as they are both insufficient representations and unnecessarily expensive to compute. In this paper, we propose an Implicit Memory Transformer that implicitly retains memory through a new left context method, removing the need to explicitly represent memory with memory banks. We generate the left context from the attention output of the previous segment and include it in the keys and values of the current segment's attention calculation. Experiments on the MuST-C dataset show that the Implicit Memory Transformer provides a substantial speedup on the encoder forward pass with nearly identical translation quality when compared with the state-of-the-art approach that employs both left context and memory banks.
Transformer models using segment-based processing have been an effective architecture for simultaneous speech translation. However, such models create a context mismatch between training and inference environments, hindering potential translation accuracy. We solve this issue by proposing Shiftable Context, a simple yet effective scheme to ensure that consistent segment and context sizes are maintained throughout training and inference, even with the presence of partially filled segments due to the streaming nature of simultaneous translation. Shiftable Context is also broadly applicable to segment-based transformers for streaming tasks. Our experiments on the English-German, English-French, and English-Spanish language pairs from the MUST-C dataset demonstrate that when applied to the Augmented Memory Transformer, a state-of-the-art model for simultaneous speech translation, the proposed scheme achieves an average increase of 2.09, 1.83, and 1.95 BLEU scores across each wait-k value for the three language pairs, respectively, with a minimal impact on computation-aware Average Lagging.
Compactness in deep learning can be critical to a model's viability in low-resource applications, and a common approach to extreme model compression is quantization. We consider Iterative Product Quantization (iPQ) with Quant-Noise to be state-of-the-art in this area, but this quantization framework suffers from preventable inference quality degradation due to prevalent empty clusters. In this paper, we propose several novel enhancements aiming to improve the accuracy of iPQ with Quant-Noise by focusing on resolving empty clusters. Our contribution, which we call Partitioning-Guided k-means (PG k-means), is a heavily augmented k-means implementation composed of three main components. First, we propose a partitioning-based pre-assignment strategy that ensures no initial empty clusters and encourages an even weight-to-cluster distribution. Second, we propose an empirically superior empty cluster resolution heuristic executed via cautious partitioning of large clusters. Finally, we construct an optional optimization step that consolidates intuitively dense clusters of weights to ensure shared representation. The proposed approach consistently reduces the number of empty clusters in iPQ with Quant-Noise by 100x on average, uses 8x fewer iterations during empty cluster resolution, and improves overall model accuracy by up to 12%, when applied to RoBERTa on a variety of tasks in the GLUE benchmark.
Various natural language processing (NLP) tasks necessitate models that are efficient and small based on their ultimate application at the edge or in other resource-constrained environments. While prior research has reduced the size of these models, increasing computational efficiency without considerable performance impacts remains difficult, especially for autoregressive tasks. This paper proposes {modular linearized attention (MLA), which combines multiple efficient attention mechanisms, including cosFormer, to maximize inference quality while achieving notable speedups. We validate this approach on several autoregressive NLP tasks, including speech-to-text neural machine translation (S2T NMT), speech-to-text simultaneous translation (SimulST), and autoregressive text-to-spectrogram, noting efficiency gains on TTS and competitive performance for NMT and SimulST during training and inference.
Resource sharing between multiple workloads has become a prominent practice among cloud service providers, motivated by demand for improved resource utilization and reduced cost of ownership. Effective resource sharing, however, remains an open challenge due to the adverse effects that resource contention can have on high-priority, user-facing workloads with strict Quality of Service (QoS) requirements. Although recent approaches have demonstrated promising results, those works remain largely impractical in public cloud environments since workloads are not known in advance and may only run for a brief period, thus prohibiting offline learning and significantly hindering online learning. In this paper, we propose RAPID, a novel framework for fast, fully-online resource allocation policy learning in highly dynamic operating environments. RAPID leverages lightweight QoS predictions, enabled by domain-knowledge-inspired techniques for sample efficiency and bias reduction, to decouple control from conventional feedback sources and guide policy learning at a rate orders of magnitude faster than prior work. Evaluation on a real-world server platform with representative cloud workloads confirms that RAPID can learn stable resource allocation policies in minutes, as compared with hours in prior state-of-the-art, while improving QoS by 9.0x and increasing best-effort workload performance by 19-43%.
A growing number of service providers are exploring methods to improve server utilization, reduce power consumption, and reduce total cost of ownership by co-scheduling high-priority latency-critical workloads with best-effort workloads. This practice requires strict resource allocation between workloads to reduce resource contention and maintain Quality of Service (QoS) guarantees. Prior resource allocation works have been shown to improve server utilization under ideal circumstances, yet often compromise QoS guarantees or fail to find valid resource allocations in more dynamic operating environments. Further, prior works are fundamentally reliant upon QoS measurements that can, in practice, exhibit significant transient fluctuations, thus stable control behavior cannot be reliably achieved. In this paper, we propose a novel framework for dynamic resource allocation based on proactive QoS prediction. These predictions help guide a reinforcement-learning-based resource controller towards optimal resource allocations while avoiding transient QoS violations due to fluctuating workload demands. Evaluation shows that the proposed method incurs 4.3x fewer QoS violations, reduces severity of QoS violations by 3.7x, improves best-effort workload performance, and improves overall power efficiency compared with prior work.
Machine learning has enabled significant benefits in diverse fields, but, with a few exceptions, has had limited impact on computer architecture. Recent work, however, has explored broader applicability for design, optimization, and simulation. Notably, machine learning based strategies often surpass prior state-of-the-art analytical, heuristic, and human-expert approaches. This paper reviews machine learning applied system-wide to simulation and run-time optimization, and in many individual components, including memory systems, branch predictors, networks-on-chip, and GPUs. The paper further analyzes current practice to highlight useful design strategies and identify areas for future work, based on optimized implementation strategies, opportune extensions to existing work, and ambitious long term possibilities. Taken together, these strategies and techniques present a promising future for increasingly automated architectural design.