arXiv subjects

Colin Cai

Publications and source records attributed to Colin Cai.

4 recordsLinked to original sources

arXiv2025

From Score Distributions to Balance: Plug-and-Play Mixture-of-Experts Routing

Mixture-of-Experts (MoE) models can scale parameter capacity by routing each token to a subset of experts through a learned gate function. While conditional routing reduces training costs, it shifts the burden on inference memory: expert parameters and activations consume memory, limiting the number of experts per device. As tokens are routed, some experts become overloaded while others are underutilized. Because experts are mapped to GPUs, this imbalance translates directly into degraded system performance in terms of latency, throughput, and cost. We present LASER, a plug-and-play, inference-time routing algorithm that balances load while preserving accuracy. LASER adapts to the shape of the gate's score distribution. When scores provide a clear preference, it routes to the strongest experts; when scores are more uniform, it broadens the set of viable experts and routes to the least-loaded among them. Because LASER relies only on gate scores from a trained model, it integrates directly into existing MoE inference pipelines without retraining or finetuning. We evaluate LASER on Mixtral-8x7B and DeepSeek-MoE-16b-chat across four datasets (ARC-Easy, ARC-Challenge, MMLU, and GSM8K). LASER improves load balancing, translating into lower latency and higher throughput, while keeping the accuracy changes negligible.

cs.LG↗

arXiv2025

Nexus:Proactive Intra-GPU Disaggregation of Prefill and Decode in LLM Serving

Monolithic serving with chunked prefill improves GPU utilization by batching prefill and decode together, but suffers from fine-grained phase interference. Engine-level prefill-decode (PD) disaggregation avoids interference but incurs higher hardware and coordination overhead. Prior intra-GPU disaggregation approaches multiplex prefill and decode within a single GPU, using SLO-based tuning guided by heuristics from offline profiling or reactive feedback loops. However, these methods respond reactively to performance issues rather than anticipating them, limiting adaptability under dynamic workloads. We ask: can we achieve proactive intra-GPU disaggregation that adapts effectively to dynamic workloads? The key challenge lies in managing the conflicting resource demands of prefill and decode under varying conditions. We first show that GPU resources exhibit diminishing returns -- beyond a saturation point, more allocation yields minimal latency benefit. Second, we observe that memory bandwidth contention becomes a critical bottleneck. These insights motivate a design that dynamically partitions GPU resources across prefill and decode phases, while jointly considering compute capacity, memory footprint, and bandwidth contention. Evaluated on diverse LLMs and workloads, our system Nexus achieves up to 2.2x higher throughput, 20x lower TTFT, and 2.5x lower TBT than vLLM; outperforms SGLang by up to 2x; and matches or exceeds disaggregated vLLM.

cs.DC↗

arXiv2025

Autellix: An Efficient Serving Engine for LLM Agents as General Programs

Large language model (LLM) applications are evolving beyond simple chatbots into dynamic, general-purpose agentic programs, which scale LLM calls and output tokens to help AI agents reason, explore, and solve complex tasks. However, existing LLM serving systems ignore dependencies between programs and calls, missing significant opportunities for optimization. Our analysis reveals that programs submitted to LLM serving engines experience long cumulative wait times, primarily due to head-of-line blocking at both the individual LLM request and the program. To address this, we introduce Autellix, an LLM serving system that treats programs as first-class citizens to minimize their end-to-end latencies. Autellix intercepts LLM calls submitted by programs, enriching schedulers with program-level context. We propose two scheduling algorithms-for single-threaded and distributed programs-that preempt and prioritize LLM calls based on their programs' previously completed calls. Our evaluation demonstrates that across diverse LLMs and agentic workloads, Autellix improves throughput of programs by 4-15x at the same latency compared to state-of-the-art systems, such as vLLM.

cs.LG↗

arXiv2024

Tenspiler: A Verified Lifting-Based Compiler for Tensor Operations (Extended Version)

Tensor processing infrastructures such as deep learning frameworks and specialized hardware accelerators have revolutionized how computationally intensive code from domains such as deep learning and image processing is executed and optimized. These infrastructures provide powerful and expressive abstractions while ensuring high performance. However, to utilize them, code must be written specifically using the APIs / ISAs of such software frameworks or hardware accelerators. Importantly, given the fast pace of innovation in these domains, code written today quickly becomes legacy as new frameworks and accelerators are developed, and migrating such legacy code manually is a considerable effort. To enable developers in leveraging such DSLs while preserving their current programming paradigm, we introduce Tenspiler, a verified lifting-based compiler that uses program synthesis to translate sequential programs written in general-purpose programming languages (e.g., C++ or Python code) into tensor operations. Central to Tenspiler is our carefully crafted yet simple intermediate language, named TensIR, that expresses tensor operations. TensIR enables efficient lifting, verification, and code generation. Currently, Tenspiler already supports $\textbf{six}$ DSLs, spanning a broad spectrum of software and hardware environments. Furthermore, we show that new backends can be easily supported by Tenspiler by adding simple pattern-matching rules for TensIR. Using 10 real-world code benchmark suites, our experimental evaluation shows that by translating code to be executed on $\textbf{6}$ different software frameworks and hardware devices, Tenspiler offers on average 105$\times$ kernel and 9.65$\times$ end-to-end execution time improvement over the fully-optimized sequential implementation of the same benchmarks.

cs.PL↗