In speech separation, both CNN- and Transformer-based models have demonstrated robust separation capabilities, garnering significant attention within the research community. However, CNN-based methods have limited modelling capability for long-sequence audio, leading to suboptimal separation performance. Conversely, Transformer-based methods are limited in practical applications due to their high computational complexity. Notably, within computer vision, Mamba-based methods have been celebrated for their formidable performance and reduced computational requirements. In this paper, we propose a network architecture for speech separation using a state-space model, namely SPMamba. We adopt the TF-GridNet model as the foundational framework and substitute its Transformer component with a bidirectional Mamba module, aiming to capture a broader range of contextual information. Our experimental results reveal an important role in the performance aspects of Mamba-based models. SPMamba demonstrates superior performance with a significant advantage over existing separation models in a dataset built on Librispeech. Notably, SPMamba achieves a substantial improvement in separation quality, with a 2.42 dB enhancement in SI-SNRi compared to the TF-GridNet. The source code for SPMamba is publicly accessible at https://github.com/JusperLee/SPMamba .
Diffusion models have achieved great success in synthesizing high-quality images. However, generating high-resolution images with diffusion models is still challenging due to the enormous computational costs, resulting in a prohibitive latency for interactive applications. In this paper, we propose DistriFusion to tackle this problem by leveraging parallelism across multiple GPUs. Our method splits the model input into multiple patches and assigns each patch to a GPU. However, naively implementing such an algorithm breaks the interaction between patches and loses fidelity, while incorporating such an interaction will incur tremendous communication overhead. To overcome this dilemma, we observe the high similarity between the input from adjacent diffusion steps and propose displaced patch parallelism, which takes advantage of the sequential nature of the diffusion process by reusing the pre-computed feature maps from the previous timestep to provide context for the current step. Therefore, our method supports asynchronous communication, which can be pipelined by computation. Extensive experiments show that our method can be applied to recent Stable Diffusion XL with no quality degradation and achieve up to a 6.1$\times$ speedup on eight NVIDIA A100s compared to one. Our code is publicly available at https://github.com/mit-han-lab/distrifuser.
Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it's intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting finding not only highlights the potential redundancy of information added during fine-tuning, but also has significant implications for the multi-tenant serving and multi-tenant storage of fine-tuned models. By enabling the use of a single high-precision base model accompanied by multiple 1-bit deltas, BitDelta dramatically reduces GPU memory requirements by more than 10x, which can also be translated to enhanced generation latency in multi-tenant settings. We validate BitDelta through experiments across Llama-2 and Mistral model families, and on models up to 70B parameters, showcasing minimal performance degradation over all tested settings.
Audio-visual speech separation has gained significant traction in recent years due to its potential applications in various fields such as speech recognition, diarization, scene analysis and assistive technologies. Designing a lightweight audio-visual speech separation network is important for low-latency applications, but existing methods often require higher computational costs and more parameters to achieve better separation performance. In this paper, we present an audio-visual speech separation model called Top-Down-Fusion Net (TDFNet), a state-of-the-art (SOTA) model for audio-visual speech separation, which builds upon the architecture of TDANet, an audio-only speech separation method. TDANet serves as the architectural foundation for the auditory and visual networks within TDFNet, offering an efficient model with fewer parameters. On the LRS2-2Mix dataset, TDFNet achieves a performance increase of up to 10\% across all performance metrics compared with the previous SOTA method CTCNet. Remarkably, these results are achieved using fewer parameters and only 28\% of the multiply-accumulate operations (MACs) of CTCNet. In essence, our method presents a highly effective and efficient solution to the challenges of speech separation within the audio-visual domain, making significant strides in harnessing visual information optimally.
Multi-task reinforcement learning endeavors to accomplish a set of different tasks with a single policy. To enhance data efficiency by sharing parameters across multiple tasks, a common practice segments the network into distinct modules and trains a routing network to recombine these modules into task-specific policies. However, existing routing approaches employ a fixed number of modules for all tasks, neglecting that tasks with varying difficulties commonly require varying amounts of knowledge. This work presents a Dynamic Depth Routing (D2R) framework, which learns strategic skipping of certain intermediate modules, thereby flexibly choosing different numbers of modules for each task. Under this framework, we further introduce a ResRouting method to address the issue of disparate routing paths between behavior and target policies during off-policy training. In addition, we design an automatic route-balancing mechanism to encourage continued routing exploration for unmastered tasks without disturbing the routing of mastered ones. We conduct extensive experiments on various robotics manipulation tasks in the Meta-World benchmark, where D2R achieves state-of-the-art performance with significantly improved learning efficiency.
Machine learning (ML) inference platforms are tasked with balancing two competing goals: ensuring high throughput given many requests, and delivering low-latency responses to support interactive applications. Unfortunately, existing platform knobs (e.g., batch sizes) fail to ease this fundamental tension, and instead only enable users to harshly trade off one property for the other. This paper explores an alternate strategy to taming throughput-latency tradeoffs by changing the granularity at which inference is performed. We present Apparate, a system that automatically applies and manages early exits (EEs) in ML models, whereby certain inputs can exit with results at intermediate layers. To cope with the time-varying overhead and accuracy challenges that EEs bring, Apparate repurposes exits to provide continual feedback that powers several novel runtime monitoring and adaptation strategies. Apparate lowers median response latencies by 40.5-91.5% and 10.0-24.2% for diverse CV and NLP workloads, respectively, without affecting throughputs or violating tight accuracy constraints.
Deploying neural networks to different devices or platforms is in general challenging, especially when the model size is large or model complexity is high. Although there exist ways for model pruning or distillation, it is typically required to perform a full round of model training or finetuning procedure in order to obtain a smaller model that satisfies the model size or complexity constraints. Motivated by recent works on dynamic neural networks, we propose a simple way to train a large network and flexibly extract a subnetwork from it given a model size or complexity constraint during inference. We introduce a new way to allow a large model to be trained with dynamic depth and width during the training phase, and after the large model is trained we can select a subnetwork from it with arbitrary depth and width during the inference phase with a relatively better performance compared to training the subnetwork independently from scratch. Experiment results on a music source separation model show that our proposed method can effectively improve the separation performance across different subnetwork sizes and complexities with a single large model, and training the large model takes significantly shorter time than training all the different subnetworks.
This paper proposes a novel, data-agnostic, model poisoning attack on Federated Learning (FL), by designing a new adversarial graph autoencoder (GAE)-based framework. The attack requires no knowledge of FL training data and achieves both effectiveness and undetectability. By listening to the benign local models and the global model, the attacker extracts the graph structural correlations among the benign local models and the training data features substantiating the models. The attacker then adversarially regenerates the graph structural correlations while maximizing the FL training loss, and subsequently generates malicious local models using the adversarial graph structure and the training data features of the benign ones. A new algorithm is designed to iteratively train the malicious local models using GAE and sub-gradient descent. The convergence of FL under attack is rigorously proved, with a considerably large optimality gap. Experiments show that the FL accuracy drops gradually under the proposed attack and existing defense mechanisms fail to detect it. The attack can give rise to an infection across all benign devices, making it a serious threat to FL.