The inference process in Large Language Models (LLMs) is often limited due to the absence of parallelism in the auto-regressive decoding process, resulting in most operations being restricted by the memory bandwidth of accelerators. While methods such as speculative decoding have been suggested to address this issue, their implementation is impeded by the challenges associated with acquiring and maintaining a separate draft model. In this paper, we present Medusa, an efficient method that augments LLM inference by adding extra decoding heads to predict multiple subsequent tokens in parallel. Using a tree-based attention mechanism, Medusa constructs multiple candidate continuations and verifies them simultaneously in each decoding step. By leveraging parallel processing, Medusa introduces only minimal overhead in terms of single-step latency while substantially reducing the number of decoding steps required. We present two levels of fine-tuning procedures for Medusa to meet the needs of different use cases: Medusa-1: Medusa is directly fine-tuned on top of a frozen backbone LLM, enabling lossless inference acceleration. Medusa-2: Medusa is fine-tuned together with the backbone LLM, enabling better prediction accuracy of Medusa heads and higher speedup but needing a special training recipe that preserves the backbone model's capabilities. Moreover, we propose several extensions that improve or expand the utility of Medusa, including a self-distillation to handle situations where no training data is available and a typical acceptance scheme to boost the acceptance rate while maintaining generation quality. We evaluate Medusa on models of various sizes and training procedures. Our experiments demonstrate that Medusa-1 can achieve over 2.2x speedup without compromising generation quality, while Medusa-2 further improves the speedup to 2.3-3.6x.

Neural Architecture Search (NAS) has become a de facto approach in the recent trend of AutoML to design deep neural networks (DNNs). Efficient or near-zero-cost NAS proxies are further proposed to address the demanding computational issues of NAS, where each candidate architecture network only requires one iteration of backpropagation. The values obtained from the proxies are considered the predictions of architecture performance on downstream tasks. However, two significant drawbacks hinder the extended usage of Efficient NAS proxies. (1) Efficient proxies are not adaptive to various search spaces. (2) Efficient proxies are not extensible to multi-modality downstream tasks. Based on the observations, we design a Extensible proxy (Eproxy) that utilizes self-supervised, few-shot training (i.e., 10 iterations of backpropagation) which yields near-zero costs. The key component that makes Eproxy efficient is an untrainable convolution layer termed barrier layer that add the non-linearities to the optimization spaces so that the Eproxy can discriminate the performance of architectures in the early stage. Furthermore, to make Eproxy adaptive to different downstream tasks/search spaces, we propose a Discrete Proxy Search (DPS) to find the optimized training settings for Eproxy with only handful of benchmarked architectures on the target tasks. Our extensive experiments confirm the effectiveness of both Eproxy and Eproxy+DPS. Code is available at https://github.com/leeyeehoo/GenNAS-Zero.

Deep Neural Networks (DNNs) have achieved great success in a variety of machine learning (ML) applications, delivering high-quality inferencing solutions in computer vision, natural language processing, and virtual reality, etc. However, DNN-based ML applications also bring much increased computational and storage requirements, which are particularly challenging for embedded systems with limited compute/storage resources, tight power budgets, and small form factors. Challenges also come from the diverse application-specific requirements, including real-time responses, high-throughput performance, and reliable inference accuracy. To address these challenges, we introduce a series of effective design methodologies, including efficient ML model designs, customized hardware accelerator designs, and hardware/software co-design strategies to enable efficient ML applications on embedded systems.

Quantization for Convolutional Neural Network (CNN) has shown significant progress with the intention of reducing the cost of computation and storage with low-bitwidth data inputs. There are, however, no systematic studies on how an existing full-bitwidth processing unit, such as CPUs and DSPs, can be better utilized to carry out significantly higher computation throughput for convolution under various quantized bitwidths. In this study, we propose HiKonv, a unified solution that maximizes the compute throughput of a given underlying processing unit to process low-bitwidth quantized data inputs through novel bit-wise parallel computation. We establish theoretical performance bounds using a full-bitwidth multiplier for highly parallelized low-bitwidth convolution, and demonstrate new breakthroughs for high-performance computing in this critical domain. For example, a single 32-bit processing unit can deliver 128 binarized convolution operations (multiplications and additions) under one CPU instruction, and a single 27x18 DSP core can deliver eight convolution operations with 4-bit inputs in one cycle. We demonstrate the effectiveness of HiKonv on CPU and FPGA for both convolutional layers or a complete DNN model. For a convolutional layer quantized to 4-bit, HiKonv achieves a 3.17x latency improvement over the baseline implementation using C++ on CPU. Compared to the DAC-SDC 2020 champion model for FPGA, HiKonv achieves a 2.37x throughput improvement and 2.61x DSP efficiency improvement, respectively.

Most existing neural architecture search (NAS) algorithms are dedicated to the downstream tasks, e.g., image classification in computer vision. However, extensive experiments have shown that, prominent neural architectures, such as ResNet in computer vision and LSTM in natural language processing, are generally good at extracting patterns from the input data and perform well on different downstream tasks. These observations inspire us to ask: Is it necessary to use the performance of specific downstream tasks to evaluate and search for good neural architectures? Can we perform NAS effectively and efficiently while being agnostic to the downstream task? In this work, we attempt to affirmatively answer the above two questions and improve the state-of-the-art NAS solution by proposing a novel and generic NAS framework, termed Generic NAS (GenNAS). GenNAS does not use task-specific labels but instead adopts \textit{regression} on a set of manually designed synthetic signal bases for architecture evaluation. Such a self-supervised regression task can effectively evaluate the intrinsic power of an architecture to capture and transform the input signal patterns, and allow more sufficient usage of training samples. We then propose an automatic task search to optimize the combination of synthetic signals using limited downstream-task-specific labels, further improving the performance of GenNAS. We also thoroughly evaluate GenNAS's generality and end-to-end NAS performance on all search spaces, which outperforms almost all existing works with significant speedup.

The combination of Winograd's algorithm and systolic array architecture has demonstrated the capability of improving DSP efficiency in accelerating convolutional neural networks (CNNs) on FPGA platforms. However, handling arbitrary convolution kernel sizes in FPGA-based Winograd processing elements and supporting efficient data access remain underexplored. In this work, we are the first to propose an optimized Winograd processing element (WinoPE), which can naturally support multiple convolution kernel sizes with the same amount of computing resources and maintains high runtime DSP efficiency. Using the proposed WinoPE, we construct a highly efficient systolic array accelerator, termed WinoCNN. We also propose a dedicated memory subsystem to optimize the data access. Based on the accelerator architecture, we build accurate resource and performance modeling to explore optimal accelerator configurations under different resource constraints. We implement our proposed accelerator on multiple FPGAs, which outperforms the state-of-the-art designs in terms of both throughput and DSP efficiency. Our implementation achieves DSP efficiency up to 1.33 GOPS/DSP and throughput up to 3.1 TOPS with the Xilinx ZCU102 FPGA. These are 29.1\% and 20.0\% better than the best solutions reported previously, respectively.

Optimizing the quality of result (QoR) and the quality of service (QoS) of AI-empowered autonomous systems simultaneously is very challenging. First, there are multiple input sources, e.g., multi-modal data from different sensors, requiring diverse data preprocessing, sensor fusion, and feature aggregation. Second, there are multiple tasks that require various AI models to run simultaneously, e.g., perception, localization, and control. Third, the computing and control system is heterogeneous, composed of hardware components with varied features, such as embedded CPUs, GPUs, FPGAs, and dedicated accelerators. Therefore, autonomous systems essentially require multi-modal multi-task (MMMT) learning which must be aware of hardware performance and implementation strategies. While MMMT learning has been attracting intensive research interests, its applications in autonomous systems are still underexplored. In this paper, we first discuss the opportunities of applying MMMT techniques in autonomous systems and then discuss the unique challenges that must be solved. In addition, we discuss the necessity and opportunities of MMMT model and hardware co-design, which is critical for autonomous systems especially with power/resource-limited or heterogeneous platforms. We formulate the MMMT model and heterogeneous hardware implementation co-design as a differentiable optimization problem, with the objective of improving the solution quality and reducing the overall power consumption and critical path latency. We advocate for further explorations of MMMT in autonomous systems and software/hardware co-design solutions.

Artificial intelligence (AI) technologies have dramatically advanced in recent years, resulting in revolutionary changes in people's lives. Empowered by edge computing, AI workloads are migrating from centralized cloud architectures to distributed edge systems, introducing a new paradigm called edge AI. While edge AI has the promise of bringing significant increases in autonomy and intelligence into everyday lives through common edge devices, it also raises new challenges, especially for the development of its algorithms and the deployment of its services, which call for novel design methodologies catered to these unique challenges. In this paper, we provide a comprehensive survey of the latest enabling design methodologies that span the entire edge AI development stack. We suggest that the key methodologies for effective edge AI development are single-layer specialization and cross-layer co-design. We discuss representative methodologies in each category in detail, including on-device training methods, specialized software design, dedicated hardware design, benchmarking and design automation, software/hardware co-design, software/compiler co-design, and compiler/hardware co-design. Moreover, we attempt to reveal hidden cross-layer design opportunities that can further boost the solution quality of future edge AI and provide insights into future directions and emerging areas that require increased research focus.

Connectionist temporal classification (CTC) training criterion provides an alternative acoustic model (AM) training strategy for automatic speech recognition in an end-to-end fashion. Although CTC criterion benefits acoustic modeling without needs of time-aligned phonetics transcription, it remains in need of efforts of tweaking to convergence, especially in the resource-constrained scenario. In this paper, we proposed to improve CTC training by incorporating acoustic landmarks. We tailored a new set of acoustic landmarks to help CTC training converge more quickly while also reducing recognition error rates. We leveraged new target label sequences mixed with both phone and manner changes to guide CTC training. Experiments on TIMIT demonstrated that CTC based acoustic models converge faster and smoother significantly when they are augmented by acoustic landmarks. The models pretrained with mixed target labels can be finetuned furthermore, which reduced phone error rate by 8.72% on TIMIT. The consistent performance gain is also observed on reduced TIMIT and WSJ as well, in which case, we are the first to succeed in testing the effectiveness of acoustic landmark theory on mid-sized ASR tasks.