Devang Naik

Arnav Kundu, Yanzi Jin, Mohammad Sekhavat et al.

This paper delves into the challenging task of Active Speaker Detection (ASD), where the system needs to determine in real-time whether a person is speaking or not in a series of video frames. While previous works have made significant strides in improving network architectures and learning effective representations for ASD, a critical gap exists in the exploration of real-time system deployment. Existing models often suffer from high latency and memory usage, rendering them impractical for immediate applications. To bridge this gap, we present two scenarios that address the key challenges posed by real-time constraints. First, we introduce a method to limit the number of future context frames utilized by the ASD model. By doing so, we alleviate the need for processing the entire sequence of future frames before a decision is made, significantly reducing latency. Second, we propose a more stringent constraint that limits the total number of past frames the model can access during inference. This tackles the persistent memory issues associated with running streaming ASD systems. Beyond these theoretical frameworks, we conduct extensive experiments to validate our approach. Our results demonstrate that constrained transformer models can achieve performance comparable to or even better than state-of-the-art recurrent models, such as uni-directional GRUs, with a significantly reduced number of context frames. Moreover, we shed light on the temporal memory requirements of ASD systems, revealing that larger past context has a more profound impact on accuracy than future context. When profiling on a CPU we find that our efficient architecture is memory bound by the amount of past context it can use and that the compute cost is negligible as compared to the memory cost.

Mohammad Samragh, Mehrdad Farajtabar, Sachin Mehta et al. · utoronto

Training large transformer models from scratch for a target task requires lots of data and is computationally demanding. The usual practice of transfer learning overcomes this challenge by initializing the model with weights of a pretrained model of the same size and specification to increase the convergence and training speed. However, what if no pretrained model of the required size is available? In this paper, we introduce a simple yet effective technique to transfer the knowledge of a pretrained model to smaller variants. Our approach called weight subcloning expedites the training of scaled-down transformers by initializing their weights from larger pretrained models. Weight subcloning involves an operation on the pretrained model to obtain the equivalent initialized scaled-down model. It consists of two key steps: first, we introduce neuron importance ranking to decrease the embedding dimension per layer in the pretrained model. Then, we remove blocks from the transformer model to match the number of layers in the scaled-down network. The result is a network ready to undergo training, which gains significant improvements in training speed compared to random initialization. For instance, we achieve 4x faster training for vision transformers in image classification and language models designed for next token prediction.

Minsik Cho, Mohammad Rastegari, Devang Naik

Large Language Model or LLM inference has two phases, the prompt (or prefill) phase to output the first token and the extension (or decoding) phase to the generate subsequent tokens. In this work, we propose an efficient parallelization scheme, KV-Runahead to accelerate the prompt phase. The key observation is that the extension phase generates tokens faster than the prompt phase because of key-value cache (KV-cache). Hence, KV-Runahead parallelizes the prompt phase by orchestrating multiple processes to populate the KV-cache and minimizes the time-to-first-token (TTFT). Dual-purposing the KV-cache scheme has two main benefits. First, since KV-cache is designed to leverage the causal attention map, we minimize computation and computation automatically. Second, since it already exists for the extension phase, KV-Runahead is easy to implement. We further propose context-level load-balancing to handle uneven KV-cache generation (due to the causal attention) and to optimize TTFT. Compared with an existing parallelization scheme such as tensor or sequential parallelization where keys and values are locally generated and exchanged via all-gather collectives, our experimental results demonstrate that KV-Runahead can offer over 1.4x and 1.6x speedups for Llama 7B and Falcon 7B respectively.

Mohammad Samragh, Arnav Kundu, David Harrison et al.

Autoregressive language models are constrained by their inherently sequential nature, generating one token at a time. This paradigm limits inference speed and parallelism, especially during later stages of generation when the direction and semantics of text are relatively certain. In this work, we propose a novel framework that leverages the inherent knowledge of vanilla autoregressive language models about future tokens, combining techniques to realize this potential and enable simultaneous prediction of multiple subsequent tokens. Our approach introduces several key innovations: (1) a masked-input formulation where multiple future tokens are jointly predicted from a common prefix; (2) a gated LoRA formulation that preserves the original LLM's functionality, while equipping it for multi-token prediction; (3) a lightweight, learnable sampler module that generates coherent sequences from the predicted future tokens; (4) a set of auxiliary training losses, including a consistency loss, to enhance the coherence and accuracy of jointly generated tokens; and (5) a speculative generation strategy that expands tokens quadratically in the future while maintaining high fidelity. Our method achieves significant speedups through supervised fine-tuning on pretrained models. For example, it generates code and math nearly 5x faster, and improves general chat and knowledge tasks by almost 2.5x. These gains come without any loss in quality.

Kumari Nishu, Sachin Mehta, Samira Abnar et al.

Training large language models (LLMs) for different inference constraints is computationally expensive, limiting control over efficiency-accuracy trade-offs. Moreover, once trained, these models typically process tokens uniformly, regardless of their complexity, leading to static and inflexible behavior. In this paper, we introduce a post-training optimization framework, DynaMoE, that adapts a pre-trained dense LLM to a token-difficulty-driven Mixture-of-Experts model with minimal fine-tuning cost. This adaptation makes the model dynamic, with sensitivity control to customize the balance between efficiency and accuracy. DynaMoE features a token-difficulty-aware router that predicts the difficulty of tokens and directs them to the appropriate sub-networks or experts, enabling larger experts to handle more complex tokens and smaller experts to process simpler ones. Our experiments demonstrate that DynaMoE can generate a range of adaptive model variants of the existing trained LLM with a single fine-tuning step, utilizing only $10B$ tokens, a minimal cost compared to the base model's training. Each variant offers distinct trade-offs between accuracy and performance. Compared to the baseline post-training optimization framework, Flextron, our method achieves similar aggregated accuracy across downstream tasks, despite using only $\frac{1}{9}\text{th}$ of their fine-tuning cost.

Nikhil Bhendawade, Mahyar Najibi, Devang Naik et al.

Residual transformations enhance the representational depth and expressive power of large language models (LLMs). However, applying static residual transformations across all tokens in auto-regressive generation leads to a suboptimal trade-off between inference efficiency and generation fidelity. Existing methods, including Early Exiting, Skip Decoding, and Mixture-of-Depth address this by modulating the residual transformation based on token-level complexity. Nevertheless, these approaches predominantly consider the distance traversed by tokens through the model layers, neglecting the underlying velocity of residual evolution. We introduce Mixture of Multi-rate Residuals (M2R2), a framework that dynamically modulates residual velocity to improve early alignment, enhancing inference efficiency. Evaluations on reasoning oriented tasks such as Koala, Self-Instruct, WizardLM, and MT-Bench show M2R2 surpasses state-of-the-art distance-based strategies, balancing generation quality and speedup. In self-speculative decoding setup, M2R2 achieves up to 2.8x speedups on MT-Bench, outperforming methods like 2-model speculative decoding, Medusa, LookAhead Decoding, and DEED. In Mixture-of-Experts (MoE) architectures, integrating early residual alignment with ahead-of-time expert loading into high-bandwidth memory (HBM) accelerates decoding, reduces expert-switching bottlenecks, and achieves a 2.9x speedup, making it highly effective in resource-constrained environments.

Minsik Cho, Keivan A. Vahid, Qichen Fu et al.

Since Large Language Models or LLMs have demonstrated high-quality performance on many complex language tasks, there is a great interest in bringing these LLMs to mobile devices for faster responses and better privacy protection. However, the size of LLMs (i.e., billions of parameters) requires highly effective compression to fit into storage-limited devices. Among many compression techniques, weight-clustering, a form of non-linear quantization, is one of the leading candidates for LLM compression, and supported by modern smartphones. Yet, its training overhead is prohibitively significant for LLM fine-tuning. Especially, Differentiable KMeans Clustering, or DKM, has shown the state-of-the-art trade-off between compression ratio and accuracy regression, but its large memory complexity makes it nearly impossible to apply to train-time LLM compression. In this paper, we propose a memory-efficient DKM implementation, eDKM powered by novel techniques to reduce the memory footprint of DKM by orders of magnitudes. For a given tensor to be saved on CPU for the backward pass of DKM, we compressed the tensor by applying uniquification and sharding after checking if there is no duplicated tensor previously copied to CPU. Our experimental results demonstrate that \prjname can fine-tune and compress a pretrained LLaMA 7B model from 12.6 GB to 2.5 GB (3bit/weight) with the Alpaca dataset by reducing the train-time memory footprint of a decoder layer by 130$\times$, while delivering good accuracy on broader LLM benchmarks (i.e., 77.7% for PIQA, 66.1% for Winograde, and so on).

Pranay Dighe, Erik Marchi, Srikanth Vishnubhotla et al.

In this paper, we address the task of determining whether a given utterance is directed towards a voice-enabled smart-assistant device or not. An undirected utterance is termed as a "false trigger" and false trigger mitigation (FTM) is essential for designing a privacy-centric non-intrusive smart assistant. The directedness of an utterance can be identified by running automatic speech recognition (ASR) on it and determining the user intent by analyzing the ASR transcript. But in case of a false trigger, transcribing the audio using ASR itself is strongly undesirable. To alleviate this issue, we propose an LSTM-based FTM architecture which determines the user intent from acoustic features directly without explicitly generating ASR transcripts from the audio. The proposed models are small footprint and can be run on-device with limited computational resources. During training, the model parameters are optimized using a knowledge transfer approach where a more accurate self-attention graph neural network model serves as the teacher. Given the whole audio snippets, our approach mitigates 87% of false triggers at 99% true positive rate (TPR), and in a streaming audio scenario, the system listens to only 1.69s of the false trigger audio before rejecting it while achieving the same TPR.

Zakaria Aldeneh, Anushree Prasanna Kumar, Barry-John Theobald et al.

We present an introspection of an audiovisual speech enhancement model. In particular, we focus on interpreting how a neural audiovisual speech enhancement model uses visual cues to improve the quality of the target speech signal. We show that visual cues provide not only high-level information about speech activity, i.e., speech/silence, but also fine-grained visual information about the place of articulation. One byproduct of this finding is that the learned visual embeddings can be used as features for other visual speech applications. We demonstrate the effectiveness of the learned visual embeddings for classifying visemes (the visual analogy to phonemes). Our results provide insight into important aspects of audiovisual speech enhancement and demonstrate how such models can be used for self-supervision tasks for visual speech applications.

Siddharth Sigtia, Erik Marchi, Sachin Kajarekar et al.

Automatic speech transcription and speaker recognition are usually treated as separate tasks even though they are interdependent. In this study, we investigate training a single network to perform both tasks jointly. We train the network in a supervised multi-task learning setup, where the speech transcription branch of the network is trained to minimise a phonetic connectionist temporal classification (CTC) loss while the speaker recognition branch of the network is trained to label the input sequence with the correct label for the speaker. We present a large-scale empirical study where the model is trained using several thousand hours of labelled training data for each task. We evaluate the speech transcription branch of the network on a voice trigger detection task while the speaker recognition branch is evaluated on a speaker verification task. Results demonstrate that the network is able to encode both phonetic \emph{and} speaker information in its learnt representations while yielding accuracies at least as good as the baseline models for each task, with the same number of parameters as the independent models.