WarmServe: Enabling One-for-Many GPU Prewarming for Multi-LLM Serving

Deploying multiple models within shared GPU clusters is promising for improving resource efficiency in large language model (LLM) serving. Existing multi-LLM serving systems optimize GPU utilization at the cost of worse inference performance, especially time-to-first-token (TTFT). We identify the root cause of such compromise as their unawareness of future workload characteristics. In contrast, recent analysis on real-world traces has shown the high periodicity and long-term predictability of LLM serving workloads. We propose universal GPU workers to enable one-for-many GPU prewarming that loads models with knowledge of future workloads. Based on universal GPU workers, we design and build WarmServe, a multi-LLM serving system that (1) mitigates cluster-wide prewarming interference by adopting an evict-aware model placement strategy, (2) prepares universal GPU workers in advance by proactive prewarming, and (3) manages GPU memory with a zero-overhead memory switching mechanism. Evaluation under real-world datasets shows that WarmServe improves TTFT by up to 50.8$\times$ compared to the state-of-the-art autoscaling-based system, while being capable of serving up to 2.5$\times$ more requests compared to the GPU-sharing system.

TokenLake: A Unified Segment-level Prefix Cache Pool for Fine-grained Elastic Long-Context LLM Serving

Prefix caching is crucial to accelerate multi-turn interactions and requests with shared prefixes. At the cluster level, existing prefix caching systems are tightly coupled with request scheduling to optimize cache efficiency and computation performance together, leading to load imbalance, data redundancy, and memory fragmentation of caching systems across instances. To address these issues, memory pooling is promising to shield the scheduler from the underlying cache management so that it can focus on the computation optimization. However, because existing prefix caching systems only transfer increasingly longer prefix caches between instances, they cannot achieve low-latency memory pooling. To address these problems, we propose a unified segment-level prefix cache pool, TokenLake. It uses a declarative cache interface to expose requests' query tensors, prefix caches, and cache-aware operations to TokenLake for efficient pooling. Powered by this abstraction, TokenLake can manage prefix cache at the segment level with a heavy-hitter-aware load balancing algorithm to achieve better cache load balance, deduplication, and defragmentation. TokenLake also transparently minimizes the communication volume of query tensors and new caches. Based on TokenLake, the scheduler can schedule requests elastically by using existing techniques without considering prefix cache management. Evaluations on real-world workloads show that TokenLake can improve throughput by up to 2.6$\times$ and 2.0$\times$ and boost hit rate by 2.0$\times$ and 2.1$\times$, compared to state-of-the-art cache-aware routing and cache-centric PD-disaggregation solutions, respectively.

FinMME: Benchmark Dataset for Financial Multi-Modal Reasoning Evaluation

Multimodal Large Language Models (MLLMs) have experienced rapid development in recent years. However, in the financial domain, there is a notable lack of effective and specialized multimodal evaluation datasets. To advance the development of MLLMs in the finance domain, we introduce FinMME, encompassing more than 11,000 high-quality financial research samples across 18 financial domains and 6 asset classes, featuring 10 major chart types and 21 subtypes. We ensure data quality through 20 annotators and carefully designed validation mechanisms. Additionally, we develop FinScore, an evaluation system incorporating hallucination penalties and multi-dimensional capability assessment to provide an unbiased evaluation. Extensive experimental results demonstrate that even state-of-the-art models like GPT-4o exhibit unsatisfactory performance on FinMME, highlighting its challenging nature. The benchmark exhibits high robustness with prediction variations under different prompts remaining below 1%, demonstrating superior reliability compared to existing datasets. Our dataset and evaluation protocol are available at https://huggingface.co/datasets/luojunyu/FinMME and https://github.com/luo-junyu/FinMME.

MegaScale-MoE: Large-Scale Communication-Efficient Training of Mixture-of-Experts Models in Production

We present MegaScale-MoE, a production system tailored for the efficient training of large-scale mixture-of-experts (MoE) models. MoE emerges as a promising architecture to scale large language models (LLMs) to unprecedented sizes, thereby enhancing model performance. However, existing MoE training systems experience a degradation in training efficiency, exacerbated by the escalating scale of MoE models and the continuous evolution of hardware. Recognizing the pivotal role of efficient communication in enhancing MoE training, MegaScale-MoE customizes communication-efficient parallelism strategies for attention and FFNs in each MoE layer and adopts a holistic approach to overlap communication with computation at both inter- and intra-operator levels. Additionally, MegaScale-MoE applies communication compression with adjusted communication patterns to lower precision, further improving training efficiency. When training a 352B MoE model on 1,440 NVIDIA Hopper GPUs, MegaScale-MoE achieves a training throughput of 1.41M tokens/s, improving the efficiency by 1.88$\times$ compared to Megatron-LM. We share our operational experience in accelerating MoE training and hope that by offering our insights in system design, this work will motivate future research in MoE systems.

MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism

Mixture-of-Experts (MoE) showcases tremendous potential to scale large language models (LLMs) with enhanced performance and reduced computational complexity. However, its sparsely activated architecture shifts feed-forward networks (FFNs) from being compute-intensive to memory-intensive during inference, leading to substantially lower GPU utilization and increased operational costs. We present MegaScale-Infer, an efficient and cost-effective system for serving large-scale MoE models. MegaScale-Infer disaggregates attention and FFN modules within each model layer, enabling independent scaling, tailored parallelism strategies, and heterogeneous deployment for both modules. To fully exploit disaggregation in the presence of MoE's sparsity, MegaScale-Infer introduces ping-pong pipeline parallelism, which partitions a request batch into micro-batches and shuttles them between attention and FFNs for inference. Combined with distinct model parallelism for each module, MegaScale-Infer effectively hides communication overhead and maximizes GPU utilization. To adapt to disaggregated attention and FFN modules and minimize data transmission overhead (e.g., token dispatch), MegaScale-Infer provides a high-performance M2N communication library that eliminates unnecessary GPU-to-CPU data copies, group initialization overhead, and GPU synchronization. Experimental results indicate that MegaScale-Infer achieves up to 1.90x higher per-GPU throughput than state-of-the-art solutions.

Benchmarking Bias in Large Language Models during Role-Playing

Large Language Models (LLMs) have become foundational in modern language-driven applications, profoundly influencing daily life. A critical technique in leveraging their potential is role-playing, where LLMs simulate diverse roles to enhance their real-world utility. However, while research has highlighted the presence of social biases in LLM outputs, it remains unclear whether and to what extent these biases emerge during role-playing scenarios. In this paper, we introduce BiasLens, a fairness testing framework designed to systematically expose biases in LLMs during role-playing. Our approach uses LLMs to generate 550 social roles across a comprehensive set of 11 demographic attributes, producing 33,000 role-specific questions targeting various forms of bias. These questions, spanning Yes/No, multiple-choice, and open-ended formats, are designed to prompt LLMs to adopt specific roles and respond accordingly. We employ a combination of rule-based and LLM-based strategies to identify biased responses, rigorously validated through human evaluation. Using the generated questions as the benchmark, we conduct extensive evaluations of six advanced LLMs released by OpenAI, Mistral AI, Meta, Alibaba, and DeepSeek. Our benchmark reveals 72,716 biased responses across the studied LLMs, with individual models yielding between 7,754 and 16,963 biased responses, underscoring the prevalence of bias in role-playing contexts. To support future research, we have publicly released the benchmark, along with all scripts and experimental results.

Empowering 1000 tokens/second on-device LLM prefilling with mllm-NPU

On-device large language models (LLMs) are catalyzing novel mobile applications such as UI task automation and personalized email auto-reply, without giving away users' private data. However, on-device LLMs still suffer from unacceptably long inference latency, especially the time to first token (prefill stage) due to the need of long context for accurate, personalized content generation, as well as the lack of parallel computing capacity of mobile CPU/GPU. To enable practical on-device LLM, we present mllm-NPU, the first-of-its-kind LLM inference system that efficiently leverages on-device Neural Processing Unit (NPU) offloading. Essentially, mllm-NPU is an algorithm-system co-design that tackles a few semantic gaps between the LLM architecture and contemporary NPU design. Specifically, it re-constructs the prompt and model in three levels: (1) At prompt level, it divides variable-length prompts into multiple fixed-sized chunks while maintaining data dependencies; (2) At tensor level, it identifies and extracts significant outliers to run on the CPU/GPU in parallel with minimal overhead; (3) At block level, it schedules Transformer blocks in an out-of-order manner to the CPU/GPU and NPU based on their hardware affinity and sensitivity to accuracy. Compared to competitive baselines, mllm-NPU achieves 22.4x faster prefill speed and 30.7x energy savings on average, and up to 32.8x speedup in an end-to-end real-world application. For the first time, mllm-NPU achieves more than 1,000 tokens/sec prefilling for a billion-sized model (Qwen1.5-1.8B), paving the way towards practical on-device LLM.

RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose Thoth, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design Thoth, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. Thoth proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement Thoth and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that Thoth reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

LoongServe: Efficiently Serving Long-context Large Language Models with Elastic Sequence Parallelism

The context window of large language models (LLMs) is rapidly increasing, leading to a huge variance in resource usage between different requests as well as between different phases of the same request. Restricted by static parallelism strategies, existing LLM serving systems cannot efficiently utilize the underlying resources to serve variable-length requests in different phases. To address this problem, we propose a new parallelism paradigm, elastic sequence parallelism (ESP), to elastically adapt to the variance between different requests and phases. Based on ESP, we design and build LoongServe, an LLM serving system that (1) improves computation efficiency by elastically adjusting the degree of parallelism in real-time, (2) improves communication efficiency by reducing key-value cache migration overhead and overlapping partial decoding communication with computation, and (3) improves GPU memory efficiency by reducing key-value cache fragmentation across instances. Our evaluation under diverse real-world datasets shows that LoongServe improves the maximum throughput by up to 3.85$\times$ compared to the chunked prefill and 5.81$\times$ compared to the prefill-decoding disaggregation.

Exploring the Impact of In-Browser Deep Learning Inference on Quality of User Experience and Performance

Deep Learning (DL) is increasingly being integrated into Web applications through a method known as "in-browser inference", where the DL processes occur directly within Web browsers. However, the actual performance of this method and its effect on user experience quality (QoE) is not well-understood. This gap in knowledge necessitates new forms of QoE measurement, going beyond traditional metrics such as page load time. To address this, we conducted the first extensive performance evaluation of in-browser inference. We introduced new metrics for this purpose: responsiveness, smoothness, and inference accuracy. Our thorough study included 9 widely-used DL models and tested them across 50 popular PC Web browsers. The findings show a significant latency issue with in-browser inference: it's on average 16.9 times slower on CPU and 4.9 times slower on GPU than native inference methods. Several factors contribute to this latency, including underused hardware instruction sets, inherent delays in the runtime environment, resource competition within the browser, and inefficiencies in software libraries and GPU abstractions. Moreover, in-browser inference demands a lot of memory, sometimes up to 334.6 times more than the size of the DL models themselves. This excessive memory usage is partly due to suboptimal memory management. Additionally, we noticed that in-browser inference increases the time it takes for graphical user interface (GUI) components to load in web browsers by a significant 67.2\%, which severely impacts the overall QoE for users of web applications that depend on this technology.