This paper presents MoE-Gen, a high-throughput MoE inference system optimized for single-GPU execution. Existing inference systems rely on model-based or continuous batching strategies, originally designed for interactive inference, which result in excessively small batches for MoE's key modules-attention and expert modules-leading to poor throughput. To address this, we introduce module-based batching, which accumulates tokens in host memory and dynamically launches large batches on GPUs to maximize utilization. Additionally, we optimize the choice of batch sizes for each module in an MoE to fully overlap GPU computation and communication, maximizing throughput. Evaluation demonstrates that MoE-Gen achieves 8-31x higher throughput compared to state-of-the-art systems employing model-based batching (FlexGen, MoE-Lightning, DeepSpeed), and offers even greater throughput improvements over continuous batching systems (e.g., vLLM and Ollama) on popular MoE models (DeepSeek and Mixtral) across offline inference tasks. MoE-Gen's source code is publicly available at https://github.com/EfficientMoE/MoE-Gen

Emerging AI accelerators increasingly adopt wafer-scale manufacturing technologies, integrating hundreds of thousands of AI cores in a mesh-based architecture with large distributed on-chip memory (tens of GB in total) and ultra-high on-chip memory bandwidth (tens of PB/s). However, current LLM inference systems, optimized for shared memory architectures like GPUs, fail to fully exploit these accelerators. We introduce WaferLLM, the first wafer-scale LLM inference system. WaferLLM is guided by a novel PLMR model (pronounced as "Plummer") that captures the unique hardware characteristics of wafer-scale architectures. Leveraging this model, WaferLLM pioneers wafer-scale LLM parallelism, optimizing the utilization of hundreds of thousands of on-chip cores. It also introduces MeshGEMM and MeshGEMV, the first GEMM and GEMV implementations designed to scale effectively on wafer-scale accelerators. Evaluations show that WaferLLM achieves 200$\times$ better wafer-scale accelerator utilization than state-of-the-art systems. On a commodity wafer-scale accelerator, WaferLLM delivers 606$\times$ faster and 22$\times$ more energy-efficient GEMV compared to an advanced GPU. For LLMs, based on 16-bit data type, WaferLLM achieves 2700 toks/sec/req decode speed on Llama3-8B model and 840 toks/sec/req decode speed on Qwen2-72B model, which enables 39$\times$ faster decoding with 1.7$\times$ better energy efficiency. We anticipate these numbers will grow significantly as wafer-scale AI models, software, and hardware continue to mature.

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently; however, MoE systems rely on heterogeneous compute and memory resources. These factors collectively influence the system's Cost, Accuracy, and Performance (CAP), creating a challenging trade-off. Current benchmarks often fail to provide precise estimates of these effects, complicating practical considerations for deploying MoE systems. To bridge this gap, we introduce MoE-CAP, a benchmark specifically designed to evaluate MoE systems. Our findings highlight the difficulty of achieving an optimal balance of cost, accuracy, and performance with existing hardware capabilities. MoE systems often necessitate compromises on one factor to optimize the other two, a dynamic we term the MoE-CAP trade-off. To identify the best trade-off, we propose novel performance evaluation metrics - Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU) - and develop cost models that account for the heterogeneous compute and memory hardware integral to MoE systems. This benchmark is publicly available on HuggingFace: https://huggingface.co/spaces/sparse-generative-ai/open-moe-llm-leaderboard.

Fourier features based positional encoding (PE) is commonly used in machine learning tasks that involve learning high-frequency features from low-dimensional inputs, such as 3D view synthesis and time series regression with neural tangent kernels. Despite their effectiveness, existing PEs require manual, empirical adjustment of crucial hyperparameters, specifically the Fourier features, tailored to each unique task. Further, PEs face challenges in efficiently learning high-frequency functions, particularly in tasks with limited data. In this paper, we introduce sinusoidal PE (SPE), designed to efficiently learn adaptive frequency features closely aligned with the true underlying function. Our experiments demonstrate that SPE, without hyperparameter tuning, consistently achieves enhanced fidelity and faster training across various tasks, including 3D view synthesis, Text-to-Speech generation, and 1D regression. SPE is implemented as a direct replacement for existing PEs. Its plug-and-play nature lets numerous tasks easily adopt and benefit from SPE.

Furthermore, LLM inference latency is difficult to predict because their response time depends on the output This paper presents ServerlessLLM, a locality-enhanced length, which can vary significantly [24, 39, 77], due to iterative serverless inference system for Large Language Models output token generation. To achieve low latency, processing (LLMs). ServerlessLLM exploits the substantial capacity and an LLM request often necessitates the use of several bandwidth of storage and memory devices available on GPU GPUs for durations ranging from seconds to minutes. In practice, servers, thereby reducing costly remote checkpoint downloads LLM service providers need to host a large number of and achieving efficient checkpoint loading. ServerlessLLM LLMs catered to different developers, leading to significant achieves this through three main contributions: (i) fast LLM GPU consumption [15] and impeding the sustainability of checkpoint loading via a novel loading-optimized checkpoint LLM services [19]. As a result, LLM inference services have format design, coupled with an efficient multi-tier checkpoint to impose strict caps on the number of requests sent to their loading system; (ii) locality-driven LLM inference with live services from their users (e.g., 40 messages per 3 hours for migration, which allows ServerlessLLM to effectively achieve ChatGPT [51]), showing the provider's current inability to locality-driven server allocation while preserving the low latency satisfy the LLM inference demand. Researchers [19] project of ongoing LLM inference; and (iii) locality-aware that LLM inference costs may increase by > 50 when it server allocation, enabling ServerlessLLM to evaluate the status reaches the popularity of Google Search. of each server in a cluster and effectively schedule model To reduce GPU consumption, LLM service providers are startup time to capitalize on local checkpoint placement. Our exploring serverless inference, as seen in systems like Amazon comprehensive experiments, which include microbenchmarks SageMaker [60], Azure [46], KServe [11] and Hugging-and real-world traces, show that ServerlessLLM surpasses Face [31].

This paper presents MoE-Infinity, a cost-efficient mixture-of-expert (MoE) serving system that realizes activation-aware expert offloading. MoE-Infinity features sequence-level expert activation tracing, a new approach adept at identifying sparse activations and capturing the temporal locality of MoE inference. By analyzing these traces, MoE-Infinity performs novel activation-aware expert prefetching and caching, substantially reducing the latency overheads usually associated with offloading experts for improved cost performance. Extensive experiments in a cluster show that MoE-Infinity outperforms numerous existing systems and approaches, reducing latency by 4 - 20X and decreasing deployment costs by over 8X for various MoEs. MoE-Infinity's source code is publicly available at https://github.com/TorchMoE/MoE-Infinity

Deep learning (DL) jobs use multi-dimensional parallelism, i.e they combine data, model, and pipeline parallelism, to use large GPU clusters efficiently. This couples jobs tightly to a set of GPU devices, but jobs may experience changes to the device allocation: (i) resource elasticity during training adds or removes devices; (ii) hardware maintenance may require redeployment on different devices; and (iii) device failures force jobs to run with fewer devices. Current DL frameworks lack support for these scenarios, as they cannot change the multi-dimensional parallelism of an already-running job in an efficient and model-independent way. We describe Tenplex, a state management library for DL frameworks that enables jobs to change the GPU allocation and job parallelism at runtime. Tenplex achieves this by externalizing the DL job state during training as a parallelizable tensor collection (PTC). When the GPU allocation for the DL job changes, Tenplex uses the PTC to transform the DL job state: for the dataset state, Tenplex repartitions it under data parallelism and exposes it to workers through a virtual file system; for the model state, Tenplex obtains it as partitioned checkpoints and transforms them to reflect the new parallelization configuration. For efficiency, these PTC transformations are executed in parallel with a minimum amount of data movement between devices and workers. Our experiments show that Tenplex enables DL jobs to support dynamic parallelization with low overhead.

This paper introduces a distributed, GPU-centric experience replay system, GEAR, designed to perform scalable reinforcement learning (RL) with large sequence models (such as transformers). With such models, existing systems such as Reverb face considerable bottlenecks in memory, computation, and communication. GEAR, however, optimizes memory efficiency by enabling the memory resources on GPU servers (including host memory and device memory) to manage trajectory data. Furthermore, it facilitates decentralized GPU devices to expedite various trajectory selection strategies, circumventing computational bottlenecks. GEAR is equipped with GPU kernels capable of collecting trajectories using zero-copy access to host memory, along with remote-directed-memory access over InfiniBand, improving communication efficiency. Cluster experiments have shown that GEAR can achieve performance levels up to 6x greater than Reverb when training state-of-the-art large RL models. GEAR is open-sourced at https://github.com/bigrl-team/gear.

Transformer architectures have facilitated the development of large-scale and general-purpose sequence models for prediction tasks in natural language processing and computer vision, e.g., GPT-3 and Swin Transformer. Although originally designed for prediction problems, it is natural to inquire about their suitability for sequential decision-making and reinforcement learning problems, which are typically beset by long-standing issues involving sample efficiency, credit assignment, and partial observability. In recent years, sequence models, especially the Transformer, have attracted increasing interest in the RL communities, spawning numerous approaches with notable effectiveness and generalizability. This survey presents a comprehensive overview of recent works aimed at solving sequential decision-making tasks with sequence models such as the Transformer, by discussing the connection between sequential decision-making and sequence modeling, and categorizing them based on the way they utilize the Transformer. Moreover, this paper puts forth various potential avenues for future research intending to improve the effectiveness of large sequence models for sequential decision-making, encompassing theoretical foundations, network architectures, algorithms, and efficient training systems. As this article has been accepted by the Frontiers of Computer Science, here is an early version, and the most up-to-date version can be found at https://journal.hep.com.cn/fcs/EN/10.1007/s11704-023-2689-5

Systems for serving inference requests on graph neural networks (GNN) must combine low latency with high throughout, but they face irregular computation due to skew in the number of sampled graph nodes and aggregated GNN features. This makes it challenging to exploit GPUs effectively: using GPUs to sample only a few graph nodes yields lower performance than CPU-based sampling; and aggregating many features exhibits high data movement costs between GPUs and CPUs. Therefore, current GNN serving systems use CPUs for graph sampling and feature aggregation, limiting throughput. We describe Quiver, a distributed GPU-based GNN serving system with low-latency and high-throughput. Quiver's key idea is to exploit workload metrics for predicting the irregular computation of GNN requests, and governing the use of GPUs for graph sampling and feature aggregation: (1) for graph sampling, Quiver calculates the probabilistic sampled graph size, a metric that predicts the degree of parallelism in graph sampling. Quiver uses this metric to assign sampling tasks to GPUs only when the performance gains surpass CPU-based sampling; and (2) for feature aggregation, Quiver relies on the feature access probability to decide which features to partition and replicate across a distributed GPU NUMA topology. We show that Quiver achieves up to 35 times lower latency with an 8 times higher throughput compared to state-of-the-art GNN approaches (DGL and PyG).