Yejin Lee

Yejin Lee, Anna Sun, Basil Hosmer et al. · meta-ai, stanford

Generative artificial intelligence (AI) technology is revolutionizing the computing industry. Not only its applications have broadened to various sectors but also poses new system design and optimization opportunities. The technology is capable of understanding and responding in multiple modalities. However, the advanced capability currently comes with significant system resource demands. To sustainably scale generative AI capabilities to billions of users in the world, inference must be fast and efficient. This paper pinpoints key system design and optimization opportunities by characterizing a family of emerging multi-modal generation models on real systems. Auto-regressive token generation is a critical latency performance bottleneck, typically dominated by GPU idle time. In addition to memory-intensive attention across the generative AI models, linear operations constitute significant inference latency due to the feed forward networks in Transformer-based models. We demonstrate that state-of-the-art optimization levers, spanning from applications to system software and hardware, set a 3.88x better baseline.

Ruiqi Wang, Zichen Wang, Peiqi Gao et al.

With advancements in computer vision and deep learning, video-based human action recognition (HAR) has become practical. However, due to the complexity of the computation pipeline, running HAR on live video streams incurs excessive delays on embedded platforms. This work tackles the real-time performance challenges of HAR with four contributions: 1) an experimental study identifying a standard Optical Flow (OF) extraction technique as the latency bottleneck in a state-of-the-art HAR pipeline, 2) an exploration of the latency-accuracy tradeoff between the standard and deep learning approaches to OF extraction, which highlights the need for a novel, efficient motion feature extractor, 3) the design of Integrated Motion Feature Extractor (IMFE), a novel single-shot neural network architecture for motion feature extraction with drastic improvement in latency, 4) the development of RT-HARE, a real-time HAR system tailored for embedded platforms. Experimental results on an Nvidia Jetson Xavier NX platform demonstrated that RT-HARE realizes real-time HAR at a video frame rate of 30 frames per second while delivering high levels of recognition accuracy.

Saurabh Agarwal, Bilge Acun, Basil Hosmer et al.

Large Language Models (LLMs) with hundreds of billions of parameters have transformed the field of machine learning. However, serving these models at inference time is both compute and memory intensive, where a single request can require multiple GPUs and tens of Gigabytes of memory. Multi-Head Attention is one of the key components of LLMs, which can account for over 50% of LLMs memory and compute requirement. We observe that there is a high amount of redundancy across heads on which tokens they pay attention to. Based on this insight, we propose Clustered Head Attention (CHAI). CHAI combines heads with a high amount of correlation for self-attention at runtime, thus reducing both memory and compute. In our experiments, we show that CHAI is able to reduce the memory requirements for storing K,V cache by up to 21.4% and inference time latency by up to 1.73x without any fine-tuning required. CHAI achieves this with a maximum 3.2% deviation in accuracy across 3 different models (i.e. OPT-66B, LLAMA-7B, LLAMA-33B) and 5 different evaluation datasets.

Alicia Golden, Samuel Hsia, Fei Sun et al.

As the development of large-scale Generative AI models evolve beyond text (1D) generation to include image (2D) and video (3D) generation, processing spatial and temporal information presents unique challenges to quality, performance, and efficiency. We present the first work towards understanding this new system design space for multi-modal text-to-image (TTI) and text-to-video (TTV) generation models. Current model architecture designs are bifurcated into 2 categories: Diffusion- and Transformer-based models. Our systematic performance characterization on a suite of eight representative TTI/TTV models shows that after state-of-the-art optimization techniques such as Flash Attention are applied, Convolution accounts for up to 44% of execution time for Diffusion-based TTI models, while Linear layers consume up to 49% of execution time for Transformer-based models. We additionally observe that Diffusion-based TTI models resemble the Prefill stage of LLM inference, and benefit from 1.1-2.5x greater speedup from Flash Attention than Transformer-based TTI models that resemble the Decode phase. Since optimizations designed for LLMs do not map directly onto TTI/TTV models, we must conduct a thorough characterization of these workloads to gain insights for new optimization opportunities. In doing so, we define sequence length in the context of TTI/TTV models and observe sequence length can vary up to 4x in Diffusion model inference. We additionally observe temporal aspects of TTV workloads pose unique system bottlenecks, with Temporal Attention accounting for over 60% of total Attention time. Overall, our in-depth system performance characterization is a critical first step towards designing efficient and deployable systems for emerging TTI/TTV workloads.

Bangsheng Tang, Carl Chengyan Fu, Fei Kou et al.

Speculative decoding is a standard method for accelerating the inference speed of large language models. However, scaling it for production environments poses several engineering challenges, including efficiently implementing different operations (e.g., tree attention and multi-round speculative decoding) on GPU. In this paper, we detail the training and inference optimization techniques that we have implemented to enable EAGLE-based speculative decoding at a production scale for Llama models. With these changes, we achieve a new state-of-the-art inference latency for Llama models. For example, Llama4 Maverick decodes at a speed of about 4 ms per token (with a batch size of one) on 8 NVIDIA H100 GPUs, which is 10% faster than the previously best known method. Furthermore, for EAGLE-based speculative decoding, our optimizations enable us to achieve a speed-up for large batch sizes between 1.4x and 2.0x at production scale.

Yejin Lee, Jaehoon Kang, Kyuhong Shim

While persona-driven large language models (LLMs) and prompt-based text-to-speech (TTS) systems have advanced significantly, a usability gap arises when users attempt to generate voices matching their desired personas from implicit descriptions. Most users lack specialized knowledge to specify detailed voice attributes, which often leads TTS systems to misinterpret their expectations. To address these gaps, we introduce Persona-to-Voice-Attribute (P2VA), the first framework enabling voice generation automatically from persona descriptions. Our approach employs two strategies: P2VA-C for structured voice attributes, and P2VA-O for richer style descriptions. Evaluation shows our P2VA-C reduces WER by 5% and improves MOS by 0.33 points. To the best of our knowledge, P2VA is the first framework to establish a connection between persona and voice synthesis. In addition, we discover that current LLMs embed societal biases in voice attributes during the conversion process. Our experiments and findings further provide insights into the challenges of building persona-voice systems.