Hybrid models that combine the capabilities of Attention layers with the efficiency of recurrent layers (e.g., State Space Models) have gained traction in practically supporting long contexts in Large Language Model serving. Yet, the unique properties of these models complicate the usage of complementary efficiency optimizations such as prefix caching that skip redundant computations across requests. Most notably, their use of in-place state updates for recurrent layers precludes rolling back cache entries for partial sequence overlaps, and instead mandates only exact-match cache hits; the effect is a deluge of (large) cache entries per sequence, most of which yield minimal reuse opportunities. We present Marconi, the first system that supports efficient prefix caching with Hybrid LLMs. Key to Marconi are its novel admission and eviction policies that more judiciously assess potential cache entries based not only on recency, but also on (1) forecasts of their reuse likelihood across a taxonomy of different hit scenarios, and (2) the compute savings that hits deliver relative to memory footprints. Across diverse workloads and Hybrid models, Marconi achieves up to 34.4$\times$ higher token hit rates (71.1\% or 617 ms lower TTFT) compared to state-of-the-art prefix caching systems.

Transformers, driven by attention mechanisms, form the foundation of large language models (LLMs). As these models scale up, efficient GPU attention kernels become essential for high-throughput and low-latency inference. Diverse LLM applications demand flexible and high-performance attention solutions. We present FlashInfer: a customizable and efficient attention engine for LLM serving. FlashInfer tackles KV-cache storage heterogeneity using block-sparse format and composable formats to optimize memory access and reduce redundancy. It also offers a customizable attention template, enabling adaptation to various settings through Just-In-Time (JIT) compilation. Additionally, FlashInfer’s load-balanced scheduling algorithm adjusts to dynamism of user requests while maintaining compatibility with CUDAGraph which requires static configuration. FlashInfer have been integrated into leading LLM serving frameworks like SGLang, vLLM and MLC-Engine. Comprehensive kernel-level and end-to-end evaluations demonstrate FlashInfer’s ability to significantly boost kernel performance across diverse inference scenarios: compared to state-of-the-art LLM serving solutions, FlashInfer achieve 29-69% inter-token-latency reduction compared to compiler backends for LLM serving benchmark, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.

Software bloat refers to code and features that are not used by a software during runtime. For Machine Learning (ML) systems, bloat is a major contributor to their technical debt, leading to decreased performance and resource wastage. In this work, we present Negativa-ML, a novel tool to identify and remove bloat in ML frameworks by analyzing their shared libraries.Our approach includes novel techniques to detect and locate unnecessary code within GPU code - a key area overlooked by existing research.We evaluate Negativa-ML using four popular ML frameworks across ten workloads over 300 shared libraries.Our results demonstrate that ML frameworks are highly bloated on both the GPU and CPU code side, with GPU code being a primary source of bloat within ML frameworks.On average, Negativa-ML reduces the GPU code size by up to 75\% and the CPU code by up to 72\%, resulting in total file size reductions of up to 55\%.Through debloating, we achieve reductions in peak CPU memory usage, peak GPU memory usage, and execution time by up to 74.6\%, 69.6\%, and 44.6\%, respectively.

Mixture-of-experts (MoE) has been extensively employed to scale large language models to trillion-plus parameters while maintaining a fixed computational cost. The development of large MoE models in the distributed scenario encounters the problem of large communication overhead. The inter-device communication of a MoE layer can occupy 47% time of the entire model execution with popular models and frameworks. Therefore, existing methods suggest the communication in a MoE layer to be pipelined with the computation for overlapping. However, these coarse grained overlapping schemes introduce a notable impairment of computational efficiency and the latency concealing is sub-optimal.To this end, we present COMET, an optimized MoE system with fine-grained communication-computation overlapping. Leveraging data dependency analysis and task rescheduling, COMET achieves precise fine-grained overlapping of communication and computation. Through adaptive workload assignment, COMET effectively eliminates fine-grained communication bottlenecks and enhances its adaptability across various scenarios. Our evaluation shows that COMET accelerates the execution of a single MoE layer by $1.96\times$ and for end-to-end execution, COMET delivers a $1.71\times$ speedup on average. COMET has been adopted in the production environment of clusters with ten-thousand-scale of GPUs, achieving savings of millions of GPU hours.

To improve the efficiency of distributed large language model (LLM) inference, various parallelization strategies, such as tensor and pipeline parallelism, have been proposed. However, the distinct computational characteristics inherent in the two stages of LLM inference—prefilling and decoding—render a single static parallelization strategy insufficient for the effective optimization of both stages.In this work, we present Seesaw, an LLM inference engine optimized for throughput-oriented tasks. The key idea behind Seesaw is dynamic model re-sharding, a technique that facilitates the dynamic reconfiguration of parallelization strategies across stages, thereby maximizing throughput at both phases.To mitigate re-sharing overhead and optimize computational efficiency, we employ tiered KV cache buffering and transition-minimizing scheduling. These approaches work synergistically to reduce the overhead caused by frequent stage transitions while ensuring maximum batching efficiency.Our evaluation demonstrates that Seesaw achieves a throughput increase of up to 1.78$\times$ (1.36$\times$ on average) compared to vLLM, the most widely used state-of-the-art LLM inference engine.

Scheduling virtual machines (VMs) to hosts in cloud data centers dictates efficiency and is an NP-hard problem with incomplete information. Prior work improved VM scheduling with predicted VM lifetimes. Our work further improves lifetime-aware scheduling using repredictions with lifetime distributions vs. one-shot prediction. The approach repredicts and adjusts VM and host lifetimes when incorrect predictions emerge. We also present novel approaches for defragmentation and regular system maintenance, which are essential to our data center reliability and optimizations, and are unexplored in prior work. We show that repredictions deliver a fundamental advance in effectiveness over one-shot prediction.We call our novel combination of distribution-based lifetime predictions and scheduling algorithms Lifetime Aware VM Allocation (LAVA). LAVA improves resource stranding and the number of empty hosts, which are critical for large VM scheduling, cloud system updates, and reducing dynamic energy consumption. Our approach runs in production within AnonCorp’s hyperscale cloud data centers, where it improves efficiency by decreasing stranded compute and memory resources by ~3% and ~2% respectively, and increases availability for large VMs and cloud system updates by increasing empty hosts by 2.3-9.2 pp in production. We also show a reduction in VM migrations for host defragmentation and maintenance. In addition to our …

Large language models (LLMs) demonstrate remarkable capabilities but are notoriously memory-intensive during training, particularly with the popular AdamW optimizer. This memory burden often necessitates using more or higher-end GPUs or reducing batch sizes, limiting training scalability and throughput, respectively. To address this, various memory-efficient optimizers have been proposed to reduce optimizer memory usage. However, they face key challenges: (i) reliance on costly SVD operations (e.g., GaLore, Fira); (ii) significant performance trade-offs compared to AdamW (e.g., Flora); and (iii) still substantial memory overhead of optimization states in order to maintain competitive performance (e.g., 1/4 rank in GaLore, and full-rank first momentum in Adam-mini).In this work, we investigate the redundancy in AdamW's learning rate adaptation rule and identify that it can be coarsened as a structured learning rate update (channel-wise or tensor-wise). Based on this insight, we propose a novel approach, Approximated Gradient Scaling for Memory Efficient LLM Optimization (APOLLO), which approximates the channel-wise learning rate scaling with an auxiliary low-rank optimizer state based on pure random projection. The structured learning rate update rule makes APOLLO highly tolerant to further memory reduction with lower rank, halving the rank while delivering similar pre-training performance. Moreover, we further propose an extreme memory-efficient version, APOLLO-MINI, …