Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks, from conversational AI to code generation and content creation [1, 2, 3]. However, the deployment of these models in production environments presents significant engineering challenges. The computational demands of autoregressive text generation, combined with the massive parameter counts of modern LLMs, necessitate specialized serving infrastructure that can efficiently manage GPU resources while meeting application-specific performance requirements. The serving infrastructure for LLMs must address several competing objectives: maximizing throughput to serve many concurrent users, minimizing latency for responsive user experiences, and efficiently utilizing expensive GPU resources. Different applications prioritize these objectives differently--a chatbot requires low latency for individual requests, while a batch document processing system prioritizes throughput. This variation in requirements has led to the development of specialized serving frameworks, each making different design trade-offs. Among the available open-source solutions, vLLM [4] and HuggingFace Text Generation Inference (TGI) [5] have emerged as leading frameworks, widely adopted in both research and production settings.

This work elaborates on a High performance computing (HPC) architecture based on Simple Linux Utility for Resource Management (SLURM) [1] for deploying heterogeneous Large Language Models (LLMs) into a scalable inference engine. Dynamic resource scheduling and seamless integration of containerized microservices have been leveraged herein to manage CPU, GPU, and memory allocations efficiently in multi-node clusters. Extensive experiments, using Llama 3.2 (1B and 3B parameters) [2] and Llama 3.1 (8B and 70B) [3], probe throughput, latency, and concurrency and show that small models can handle up to 128 concurrent requests at sub-50 ms latency, while for larger models, saturation happens with as few as two concurrent users, with a latency of more than 2 seconds. This architecture includes Representational State Transfer Application Programming Interfaces (REST APIs) [4] endpoints for single and bulk inferences, as well as advanced workflows such as multi-step "tribunal" refinement. Experimental results confirm minimal overhead from container and scheduling activities and show that the approach scales reliably both for batch and interactive settings. We further illustrate real-world scenarios, including the deployment of chatbots with retrievalaugmented generation, which helps to demonstrate the flexibility and robustness of the architecture. The obtained results pave ways for significantly more efficient, responsive, and fault-tolerant LLM inference on large-scale HPC infrastructures.

Conventional operating system scheduling algorithms are largely content-ignorant, making decisions based on factors such as latency or fairness without considering the actual intents or semantics of processes. Consequently, these algorithms often do not prioritize tasks that require urgent attention or carry higher importance, such as in emergency management scenarios. However, recent advances in language models enable semantic analysis of processes, allowing for more intelligent and context-aware scheduling decisions. In this paper, we introduce the concept of semantic scheduling in scheduling of requests from large language models (LLM), where the semantics of the process guide the scheduling priorities. We present a novel scheduling algorithm with optimal time complexity, designed to minimize the overall waiting time in LLM-based prompt scheduling. Large language models (LLMs) are increasingly prevalent in a variety of domains, serving millions of users worldwide (Y u et al., 2024; Atkinson et al., 2020). Recent efforts to enhance LLM performance have focused on efficient serving architectures (Kwon et al., 2023; Dao et al., 2022; Hua et al., 2024), with the primary objectives of lowering latency and enhancing throughput. However, as LLM applications expand into areas such as medicine (Y u et al., 2024) and law (Atkinson et al., 2020), it becomes clear that the semantics (Mei et al., 2024) of each request ( e.g., the urgency or importance of the request content) can be critical to scheduling decisions. Most LLM services currently employ a first-come-first-served (FCFS) scheduling strategy, largely because the running time for each user request is unknown.

LLM decoding is bottlenecked for large batches and long contexts by loading the key-value (KV) cache from high-bandwidth memory, which inflates per-token latency, while the sequential nature of decoding limits parallelism. We analyze the interplay among arithmetic intensity, parallelization, and model quality and question whether current architectures fully exploit modern hardware. This work redesigns attention to perform more computation per byte loaded from memory to maximize hardware efficiency without trading off parallel scalability. We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfers without compromising model quality. We then introduce Grouped Latent Attention (GLA), a parallel-friendly latent attention paired with low-level optimizations for fast decoding while maintaining high model quality. Experiments show that GTA matches Grouped-Query Attention (GQA) quality while using roughly half the KV cache and that GLA matches Multi-head Latent Attention (MLA) and is easier to shard. Our optimized GLA kernel is up to 2$\times$ faster than FlashMLA, for example, in a speculative decoding setting when the query length exceeds one. Furthermore, by fetching a smaller KV cache per device, GLA reduces end-to-end latency and increases throughput in online serving benchmarks by up to 2$\times$.

In recent years,Large Language Models (LLMs) have significantly improved in generating high-quality code, enabling their integration into developers' Integrated Development Environments (IDEs) as code assistants. These assistants, such as GitHub Copilot, deliver real-time code suggestions and can greatly enhance developers' productivity. However, the environmental impact of these tools, in particular their energy consumption, remains a key concern. This paper investigates the energy consumption of LLM-based code assistants by simulating developer interactions with GitHub Copilot and analyzing various configuration factors. We collected a dataset of development traces from 20 developers and conducted extensive software project development simulations to measure energy usage under different scenarios. Our findings reveal that the energy consumption and performance of code assistants are influenced by various factors, such as the number of concurrent developers, model size, quantization methods, and the use of streaming. Notably, a substantial portion of generation requests made by GitHub Copilot is either canceled or rejected by developers, indicating a potential area for reducing wasted computations. Based on these findings, we share actionable insights into optimizing configurations for different use cases, demonstrating that careful adjustments can lead to significant energy savings.

Large language models (LLMs) have surged in popularity and are extensively used in commercial applications, where the efficiency of model serving is crucial for the user experience. Most current research focuses on optimizing individual sub-procedures, e.g. local inference and communication, however, there is no comprehensive framework that provides a holistic system view for optimizing LLM serving in an end-to-end manner. In this work, we conduct a detailed analysis to identify major bottlenecks that impact end-to-end latency in LLM serving systems. Our analysis reveals that a comprehensive LLM serving endpoint must address a series of efficiency bottlenecks that extend beyond LLM inference. We then propose ScaleLLM, an optimized system for resource-efficient LLM serving. Our extensive experiments reveal that with 64 concurrent requests, ScaleLLM achieves a 4.3x speed up over vLLM and outperforms state-of-the-arts with 1.5x higher throughput.

A lot of people were interested in the architecture behind Cartoonizer. So Tejas and I (Niraj) have tried explaining the process we employed to make it work. In today's fast paced world, ML experts are expected to put on multiple hats in the ML workflow. One of the critical tasks in the workflow is to serve the models in production! This seemingly important piece in the pipeline tends to get overlooked thus faltering to provide value to the customers.