Daily Papers

Training large language models requires jointly configuring two interdependent aspects of the system: the global batch size, which governs statistical efficiency, and the 3D parallelism strategy, which governs hardware throughput. Existing approaches make these decisions independently: optimization work adapts the batch size to track the evolving critical batch size while keeping parallelism fixed, and systems work selects the fastest parallelism for a given fixed batch size without anticipating that the optimal batch size could change. We show that these decisions are tightly coupled: the throughput-optimal parallelism strategy may shift as the global batch size changes, so any method that fixes one while adapting the other operates with a suboptimal configuration for part of the training run. We present COPUS, a system that adaptively tunes the global batch size, parallelism strategy, and micro-batch size as training evolves. COPUS is guided by Goodput, the product of throughput and statistical efficiency, which models both hardware and statistical effects jointly and directly measures useful convergence per unit of wall-clock time. The system combines online gradient noise scale estimation under 3D parallelism with throughput-aware evaluation of candidate configurations, and supports efficient reconfiguration of both batch size and parallelism during training. We evaluate COPUS on LLM pre-training workloads across 1-4 nodes of 8xH100 and 8xMI210 GPUs and model sizes from 3B to 32B parameters, demonstrating average time-to-convergence speedups of 3.9-8.0% over the fastest baseline across four configurations, with peak gains up to 11.1%, including system overheads.

The increasing usage of Large Language Models (LLMs) has resulted in a surging demand for planet-scale serving systems, where tens of thousands of GPUs continuously serve hundreds of millions of users. Consequently, throughput (under reasonable latency constraints) has emerged as a key metric that determines serving systems' performance. To boost throughput, various methods of inter-device parallelism (e.g., data, tensor, pipeline) have been explored. However, existing methods do not consider overlapping the utilization of different resources within a single device, leading to underutilization and sub-optimal performance. We propose NanoFlow, a novel serving framework that exploits intra-device parallelism, which overlaps the usage of resources including compute, memory, and network within a single device through operation co-scheduling. To exploit intra-device parallelism, NanoFlow introduces two key innovations: First, NanoFlow splits requests into nano-batches at the granularity of operations, which breaks the dependency of sequential operations in LLM inference and enables overlapping; then, to get benefit from overlapping, NanoFlow uses an operation-level pipeline with execution unit scheduling, which partitions the device's functional units and simultaneously executes different operations in each unit. NanoFlow automates the pipeline setup using a parameter search algorithm, which enables easily porting NanoFlow to different models. We implement NanoFlow on NVIDIA GPUs and evaluate end-to-end serving throughput on several popular models such as LLaMA-2-70B, Mixtral 8x7B, LLaMA-3-8B, etc.. With practical workloads, NanoFlow provides 1.91x throughput boost compared to state-of-the-art serving systems achieving 59% to 72% of optimal throughput across ported models.

Efficient parallelism is necessary for achieving low-latency, high-throughput inference with large language models (LLMs). Tensor parallelism (TP) is the state-of-the-art method for reducing LLM response latency, however GPU communications reduces combined token throughput. On the other hand, data parallelism (DP) obtains a higher throughput yet is slow in response latency. Best of both worlds does not exist, and it is not possible to combine TP and DP because of the KV cache variance across the parallelisms. We notice Sequence Parallelism (SP - Ulysses in training) has similar properties as DP but with KV cache invariance. We adapt SP to inference, and combine it with TP to get the best of both worlds. Our solution: Shift Parallelism. Shift Parallelism dynamically switches across TP and SP, and minimizes latency in low traffic without losing throughput in high traffic. The efficient GPU communications of Shift Parallelism yields up to i) 1.51x faster response in interactive workloads and ii) 50% higher throughput in batch workloads, compared to a TP-only solution. We evaluate Shift Parallelism with real-world production traces with dynamic traffic patterns as well as synthetic benchmarking patterns across models, context sizes, and arrival rates. All results affirm the same: Shift Parallelism has a better the latency vs. throughput tradeoff than TP or DP, and hence obtains low latency without degrading throughput in dynamic workloads.

AI agents execute tens to hundreds of chained LLM calls per task, yet GPU schedulers treat each call as independent, discarding gigabytes of intermediate state between steps and inflating end-to-end latency by 3-8x. We argue that this request-level abstraction is fundamentally mismatched to compound AI workloads, and propose a shift to program-level scheduling: treating the entire agent workflow (not individual inference calls) as the first-class schedulable unit. We present SAGA, a distributed scheduler that implements this abstraction through three mechanisms: (1) Agent Execution Graphs that capture workflow structure to predict KV cache reuse across tool-call boundaries, achieving within 1.31x of Bélády's optimal offline policy; (2) session-affinity batching with work stealing that co-locates correlated requests while maintaining global load balance; and (3) Agent Fair Share, a task-completion-time fairness metric with provable bounded-deviation guarantees. On a 64-GPU cluster serving SWE-bench coding agents and WebArena browser tasks, SAGA reduces task completion time by 1.64x (geometric mean, p < 0.001) over vLLM v0.15.1 with prefix caching and affinity routing, while improving GPU memory utilization by 1.22x and achieving 99.2% SLO attainment under multi-tenant interference. These latency gains come at a quantified cost: approximately 30% lower peak throughput than throughput-optimal batch scheduling, a tradeoff appropriate for the latency-sensitive interactive deployments that dominate compound AI usage. Our results demonstrate that workflow-aware scheduling is essential for efficient compound AI serving.

The scaling of Large Language Models (LLMs) drives interest in matrix-based optimizers (e.g., Shampoo, Muon, SOAP) for their convergence efficiency; yet their requirement for holistic updates conflicts with the tensor fragmentation in distributed frameworks like Megatron. Existing solutions are suboptimal: synchronous approaches suffer from computational redundancy, while layer-wise partitioning fails to reconcile this conflict without violating the geometric constraints of efficient communication primitives. To bridge this gap, we propose Canzona, a Unified, Asynchronous, and Load-Balanced framework that decouples logical optimizer assignment from physical parameter distribution. For Data Parallelism, we introduce an alpha-Balanced Static Partitioning strategy that respects atomicity while neutralizing the load imbalance. For Tensor Parallelism, we design an Asynchronous Compute pipeline utilizing Micro-Group Scheduling to batch fragmented updates and hide reconstruction overhead. Extensive evaluations on the Qwen3 model family (up to 32B parameters) on 256 GPUs demonstrate that our approach preserves the efficiency of established parallel architectures, achieving a 1.57x speedup in end-to-end iteration time and reducing optimizer step latency by 5.8x compared to the baseline.

Concurrent local LLM serving on unified-memory desktops must preserve memory headroom and output fidelity, which speed-only rankings overlook. We introduce SiliconBench, which evaluates nine Apple Silicon serving engines through three lenses: speed, memory, and fidelity. We evaluate chat and agent serving on Qwen3, Qwen3.5, and Gemma 4. We use a classification task to check for quality regressions against an NVIDIA reference. DGX Spark provides a complementary serving-performance reference. Three desiderata guide interpretation: serving architecture readiness, memory discipline, and multi-node scaling. On Qwen3-0.6B, vllm-metal alone more than doubles throughput on both workloads from concurrency 1 to 16. CUDA vLLM and SGLang show stronger concurrency scaling on the same prompts. Explicit memory budgets do not guarantee memory headroom: two stacks complete every request while memory use approaches physical capacity and throughput declines. The newer model architectures have narrower engine support. Their evaluated implementations match the fidelity reference. Only three stacks satisfy the completion, fidelity, and model-coverage gates. Comparisons on larger dense and MoE models reinforce the importance of scheduling prompt processing alongside ongoing generation: vllm-metal's packed prefill-decode path maintains lower first-token latency than omlx under concurrent load. In the tested two-machine configurations, tensor parallelism over Thunderbolt RDMA scales while pipeline parallelism over TCP regresses. We release benchmark code, per-run results, and maintenance journals, supported by a workflow combining bounded agent fixes with human review.

Efficient data transfers over high-speed, long-distance shared networks require proper utilization of available network bandwidth. Using parallel TCP streams enables an application to utilize network parallelism and can improve transfer throughput; however, finding the optimum number of parallel TCP streams is challenging due to nondeterministic background traffic sharing the same network. Additionally, the non-stationary, multi-objectiveness, and partially-observable nature of network signals in the host systems add extra complexity in finding the current network condition. In this work, we present a novel approach to finding the optimum number of parallel TCP streams using deep reinforcement learning (RL). We devise a learning-based algorithm capable of generalizing different network conditions and utilizing the available network bandwidth intelligently. Contrary to rule-based heuristics that do not generalize well in unknown network scenarios, our RL-based solution can dynamically discover and adapt the parallel TCP stream numbers to maximize the network bandwidth utilization without congesting the network and ensure fairness among contending transfers. We extensively evaluated our RL-based algorithm's performance, comparing it with several state-of-the-art online optimization algorithms. The results show that our RL-based algorithm can find near-optimal solutions 40% faster while achieving up to 15% higher throughput. We also show that, unlike a greedy algorithm, our devised RL-based algorithm can avoid network congestion and fairly share the available network resources among contending transfers.

Offline batch inference, which leverages the flexibility of request batching to achieve higher throughput and lower costs, is becoming more popular for latency-insensitive applications. Meanwhile, recent progress in model capability and modality makes requests more diverse in compute and memory demands, creating unique opportunities for throughput improvement by resource overlapping. However, a request schedule that maximizes resource overlapping can conflict with the schedule that maximizes prefix sharing, a widely-used performance optimization, causing sub-optimal inference throughput. We present BlendServe, a system that maximizes resource utilization of offline batch inference by combining the benefits of resource overlapping and prefix sharing using a resource-aware prefix tree. BlendServe exploits the relaxed latency requirements in offline batch inference to reorder and overlap requests with varied resource demands while ensuring high prefix sharing. We evaluate BlendServe on a variety of synthetic multi-modal workloads and show that it provides up to 1.44times throughput boost compared to widely-used industry standards, vLLM and SGLang.

We present a cross-architecture evaluation of production LLM inference on AMD Instinct MI325X GPUs, benchmarking four models spanning 235B to 1 trillion parameters across three architectural families (MoE+MLA, Dense+GQA, MoE+GQA) on an 8-GPU cluster with 2TB aggregate HBM3e using vLLM v0.14.1. Our results demonstrate that architecture-aware optimization is essential: MLA models require block size 1 and cannot use KV cache offloading, while GQA models benefit from both. The AMD AITER runtime is required for competitive MLA inference throughput and must be selectively disabled for architectures with incompatible attention head configurations. A controlled AITER ablation on Llama-3.1-405B (n=5 per condition) reveals a modest 3-5% throughput benefit at high concurrency but 2-16x higher measurement variability, confirming that AITER's large speedups target MoE/MLA kernels specifically. Under text-only workloads, Llama-405B and DeepSeek V3.2 achieve comparable peak throughput (15,944 and 15,343 tok/s) despite an order-of-magnitude difference in active parameters. Under vision workloads, Qwen3-VL-235B reaches 47,873 tok/s, 6.5x higher than Kimi-K2.5 (7,327 tok/s). Active parameter count per token is associated with inference throughput, though confounded by differences in quantization, AITER acceleration, and tensor parallelism. All four models exhibit a common throughput saturation point consistent with a memory-bandwidth bottleneck (~500 concurrent for short sequences, ~100-200 for longer sequences). All models maintain 100% HTTP-level success rates through 1,000 concurrent users, processing 18.9 million tokens across 17,406 requests without failures.

This paper presents a comprehensive comparison of three dominant parallel programming models in High Performance Computing (HPC): Message Passing Interface (MPI), Open Multi-Processing (OpenMP), and Compute Unified Device Architecture (CUDA). Selecting optimal programming approaches for modern heterogeneous HPC architectures has become increasingly critical. We systematically analyze these models across multiple dimensions: architectural foundations, performance characteristics, domain-specific suitability, programming complexity, and recent advancements. We examine each model's strengths, weaknesses, and optimization techniques. Our investigation demonstrates that MPI excels in distributed memory environments with near-linear scalability for communication-intensive applications, but faces communication overhead challenges. OpenMP provides strong performance and usability in shared-memory systems and loop-centric tasks, though it is limited by shared memory contention. CUDA offers substantial performance gains for data-parallel GPU workloads, but is restricted to NVIDIA GPUs and requires specialized expertise. Performance evaluations across scientific simulations, machine learning, and data analytics reveal that hybrid approaches combining two or more models often yield optimal results in heterogeneous environments. The paper also discusses implementation challenges, optimization best practices, and emerging trends such as performance portability frameworks, task-based programming, and the convergence of HPC and Big Data. This research helps developers and researchers make informed decisions when selecting programming models for modern HPC applications, emphasizing that the best choice depends on application requirements, hardware, and development constraints.

Pipeline parallelism has been widely explored, but most existing schedules lack a systematic methodology. In this paper, we propose a framework to decompose pipeline schedules as repeating a building block and we show that the lifespan of the building block decides the peak activation memory of the pipeline schedule. Guided by the observations, we find that almost all existing pipeline schedules, to the best of our knowledge, are memory inefficient. To address this, we introduce a family of memory efficient building blocks with controllable activation memory, which can reduce the peak activation memory to 1/2 of 1F1B without sacrificing efficiency, and even to 1/3 with comparable throughput. We can also achieve almost zero pipeline bubbles while maintaining the same activation memory as 1F1B. Our evaluations demonstrate that in pure pipeline parallelism settings, our methods outperform 1F1B by from 7% to 55% in terms of throughput. When employing a grid search over hybrid parallelism hyperparameters in practical scenarios, our proposed methods demonstrate a 16% throughput improvement over the 1F1B baseline for large language models.

Triangle centrality is introduced for finding important vertices in a graph based on the concentration of triangles surrounding each vertex. It has the distinct feature of allowing a vertex to be central if it is in many triangles or none at all. We show experimentally that triangle centrality is broadly applicable to many different types of networks. Our empirical results demonstrate that 30% of the time triangle centrality identified central vertices that differed with those found by five well-known centrality measures, which suggests novelty without being overly specialized. It is also asymptotically faster to compute on sparse graphs than all but the most trivial of these other measures. We introduce optimal algorithms that compute triangle centrality in O(mbarδ) time and O(m+n) space, where barδle O(m) is the average degeneracy introduced by Burkhardt, Faber, and Harris (2020). In practical applications, barδ is much smaller than m so triangle centrality can be computed in nearly linear time. On a Concurrent Read Exclusive Write (CREW) Parallel Random Access Machine (PRAM), we give a near work-optimal parallel algorithm that takes O(log n) time using O(mm) CREW PRAM processors. In MapReduce, we show it takes four rounds using O(mm) communication bits and is therefore optimal. We also derive a linear algebraic formulation of triangle centrality which can be computed in O(mbarδ) time on sparse graphs.

Acceleration of graph applications on GPUs has found large interest due to the ubiquitous use of graph processing in various domains. The inherent irregularity in graph applications leads to several challenges for parallelization. A key challenge, which we address in this paper, is that of load-imbalance. If the work-assignment to threads uses node-based graph partitioning, it can result in skewed task-distribution, leading to poor load-balance. In contrast, if the work-assignment uses edge-based graph partitioning, the load-balancing is better, but the memory requirement is relatively higher. This makes it unsuitable for large graphs. In this work, we propose three techniques for improved load-balancing of graph applications on GPUs. Each technique brings in unique advantages, and a user may have to employ a specific technique based on the requirement. Using Breadth First Search and Single Source Shortest Paths as our processing kernels, we illustrate the effectiveness of each of the proposed techniques in comparison to the existing node-based and edge-based mechanisms.

Large language models have led to state-of-the-art accuracies across a range of tasks. However, training these models efficiently is challenging for two reasons: a) GPU memory capacity is limited, making it impossible to fit large models on even a multi-GPU server, and b) the number of compute operations required to train these models can result in unrealistically long training times. Consequently, new methods of model parallelism such as tensor and pipeline parallelism have been proposed. Unfortunately, naive usage of these methods leads to fundamental scaling issues at thousands of GPUs, e.g., due to expensive cross-node communication or devices spending significant time waiting on other devices to make progress. In this paper, we show how different types of parallelism methods (tensor, pipeline, and data parallelism) can be composed to scale to thousands of GPUs and models with trillions of parameters. We survey techniques for pipeline parallelism and propose a novel interleaved pipeline parallelism schedule that can improve throughput by 10+% with memory footprint comparable to existing approaches. We quantitatively study the trade-offs between tensor, pipeline, and data parallelism, and provide intuition as to how to configure distributed training of a large model. Our approach allows us to perform training iterations on a model with 1 trillion parameters at 502 petaFLOP/s on 3072 GPUs with achieved per-GPU throughput of 52% of theoretical peak. Our code is open sourced at https://github.com/nvidia/megatron-lm.

The rapid growth of large-language models (LLMs) is driving a new wave of specialized hardware for inference. This paper presents the first workload-centric, cross-architectural performance study of commercial AI accelerators, spanning GPU-based chips, hybrid packages, and wafer-scale engines. We compare memory hierarchies, compute fabrics, and on-chip interconnects, and observe up to 3.7x performance variation across architectures as batch size and sequence length change. Four scaling techniques for trillion-parameter models are examined; expert parallelism offers an 8.4x parameter-to-compute advantage but incurs 2.1x higher latency variance than tensor parallelism. These findings provide quantitative guidance for matching workloads to accelerators and reveal architectural gaps that next-generation designs must address.

Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models.

Fine-tuning large language models (LLMs) remains resource-intensive due to their sheer scale. While zeroth-order (ZO) optimization provides a memory-efficient alternative by eliminating backward passes, its application to multi-hundred-billion-parameter models is constrained by GPU memory and compute throughput. The ZO2 framework addresses the memory bottleneck by offloading model parameters to CPU memory and overlapping transformer block transfer with dual forward computation on a single GPU. However, ZO2 remains limited by its single-device execution and achieves modest throughput. In this work, we present DistZO2, a high-throughput, memory-efficient framework for distributed zeroth-order fine-tuning of LLMs. DistZO2 introduces three parallel strategies: (1) Perturbation Parallelism (PertP), which parallelizes the two perturbed forward passes across devices; (2) Distributed Data Parallelism (DDP), adapted to the scalar-gradient nature of ZO training; and (3) a unified 2D Parallelism design that combines PertP and DDP. To further mitigate communication bottlenecks introduced by parameter offloading, we propose a hardware-aware communication strategy that slices parameter blocks and redistributes them across GPUs via high-speed interconnects such as NVLink. DistZO2 scales zeroth-order fine-tuning to modern multi-GPU systems, preserving ZO2's memory efficiency while substantially improving training throughput. In our experiments on OPT-175B, DistZO2 achieves a 3x speedup over ZO2 with distributed computing. DistZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.

The deployment of deep neural networks (DNNs) on resource-constrained edge devices is frequently hindered by their significant computational and memory requirements. While partitioning and distributing a DNN across multiple devices is a well-established strategy to mitigate this challenge, prior research has largely focused on single-objective optimization, such as minimizing latency or maximizing throughput. This paper challenges that view by reframing DNN partitioning as a multi-objective optimization problem. We argue that in real-world scenarios, a complex trade-off between latency and throughput exists, which is further complicated by network variability. To address this, we introduce ParetoPipe, an open-source framework that leverages Pareto front analysis to systematically identify optimal partitioning strategies that balance these competing objectives. Our contributions are threefold: we benchmark pipeline partitioned inference on a heterogeneous testbed of Raspberry Pis and a GPU-equipped edge server; we identify Pareto-optimal points to analyze the latency-throughput trade-off under varying network conditions; and we release a flexible, open-source framework to facilitate distributed inference and benchmarking. This toolchain features dual communication backends, PyTorch RPC and a custom lightweight implementation, to minimize overhead and support broad experimentation.

Efficiently serving Large Language Models (LLMs) requires selecting an optimal parallel execution plan, balancing computation, memory, and communication overhead. However, determining the best strategy is challenging due to varying parallelism techniques (data, pipeline, tensor) and workload characteristics (e.g., compute-intensive tasks with long prompts vs. memory-intensive tasks with long generation). We propose APEX, an LLM serving system simulator that efficiently identifies optimal parallel execution plans by considering key factors of LLM serving systems, such as memory usage, batching behavior, etc. APEX performs dynamism-aware simulation to model iteration-level batching, and leverages LLMs' repetitive structure to reduce design space, scaling efficiently to trillion-scale models. APEX abstracts the key components of LLM serving systems, including the model, batching module, quantization formats, and device clusters, enabling the simulator to be general and extensible. Simulating on a CPU, APEX evaluates execution plans for various device clusters, covering diverse LLMs and workloads. APEX finds plans up to 3.37x faster than heuristics, and also plans that reduce energy consumption by up to 45% compared to latency-optimal plans. APEX performs comprehensive evaluations, reporting key system metrics like time per output token and time to first token, which can help service providers meet SLOs. APEX identifies an optimal plan within 15 minutes on a CPU, making it 71x faster and 1234x more cost-effective than cloud-based GPU deployment. APEX can be accessed at https://github.com/microsoft/apex_plus

AlltoAll dispatch is the dominant bottleneck of MoE expert parallelism, and the interconnect community has responded with four families of mitigations: predictive sample placement, adaptive expert relayout, hierarchical collectives, and EP-aware topology. All four rest on two assumptions about the workload: that routing imbalance is correctable by the system layer, and that the mock-token benchmarks evaluating them faithfully represent production routing. We introduce DODOCO to test both, instrumenting five open MoE checkpoints that span today's sequence-mixer designs (MHA, MLA, GQA, Gated DeltaNet and Mamba-2 SSM) under a factorial grid of six data conditions and a matched expert-parallelism scan on H100 clusters. Both assumptions fail. Scaling EP leaves per-expert load concentration essentially unchanged: the straggler is intrinsic to the routing decision the model makes, not to how its experts land on ranks. Mock tokens overestimate routing imbalance by up to a factor of 2.35, and the error is a level offset rather than a trend: it stays flat across a 32times batch-size sweep. Skewing the synthetic distribution toward realism (Zipf) widens the gap instead of closing it. A third pattern organizes the results: the architectures separate into a data-resilient band (MHA, Mamba-2), whose routing approaches uniform on real text, and a persistently concentrated band (MLA, GDN), with GQA intermediate. These bands, not the EP degree or the mock-data profile, are the right workload input to AlltoAll-aware interconnect and dispatch design.

Training billion-parameter models requires distributing model states across GPUs using fully sharded data parallel (i.e., ZeRO-3). While ZeRO-3 succeeds on clusters with high-bandwidth NVLink and InfiniBand interconnects, researchers with commodity hardware face severe inter-node all-gather bottlenecks. Existing optimizations take two approaches: GPU memory caching (MiCS, ZeRO++) trades memory capacity for reduced communication, triggering out-of-memory failures on large models; host memory offloading (ZeRO-Offload, ZeRO-Infinity) extends capacity but degrades throughput due to PCIe overhead. We observe that on bandwidth-limited clusters, host memory can serve not as an overflow tier but as a fast caching layer that outperforms inter-node communication. Based on this insight, we propose FCDP, which eliminates redundant inter-node communication while preserving ZeRO-3's minimal GPU memory footprint. FCDP caches forward-pass parameters in host memory and reuses them during the backward pass via fast intra-node all-gather, reducing inter-node all-gather by 50%. For parameter-efficient fine-tuning (PEFT), FCDP selectively communicates only trainable parameters to maximize caching, reducing inter-node traffic by over 99%. In our commodity cluster setup, FCDP achieves up to 100x higher throughput than ZeRO-3 and 51x higher than ZeRO++, while maintaining ZeRO-3's maximum batch size.

Pipeline parallelism (PP) is widely used for training large language models (LLMs), yet its scalability is often constrained by high activation memory consumption as the number of in-flight microbatches grows with the degree of PP. In this paper, we focus on addressing this challenge by leveraging the under-explored memory offload strategy in PP. With empirical study, we discover that in the majority of standard configurations, at least half, and potentially all, of the activations can be offloaded with negligible overhead. In the cases where full overload is not possible, we introduce a novel selective offload strategy that decreases peak activation memory in a better-than-linear manner. Furthermore, we integrate memory offload with other techniques to jointly consider overall throughput and memory limitation. Our experiments proves that the per-device activation memory effectively reduces with the total number of stages, making PP a stronger alternative than TP, offering up to a 19\% acceleration with even lower memory consumption. The implementation is open-sourced at https://github.com/sail-sg/zero-bubble-pipeline-parallelism{this url}.

Test-time compute scaling is a primary driver of performance in large reasoning models (LRMs), but extreme inefficiency bounds current approaches, shifting the critical question from how much compute to spend, to where to allocate it. We formalize test-time reasoning as a constrained compute allocation problem over partial trajectories. Under a fixed hardware budget, existing paradigms fail to actively allocate the compute to the most promising partial progress: traditional parallel sampling treats traces independently and induces severe memory bottlenecks, while subtractive pruning starves hardware and fails to actively and sufficiently shift the output distribution. To overcome this dichotomy, we introduce Gambit, an inference algorithm that executes thought-level beam search. By periodically pruning unpromising trajectories and immediately branching from high-quality prefixes, Gambit dynamically concentrates compute onto the most promising reasoning traces via a light-weight scorer probing hidden states while maintaining continuous high hardware utilization. Extensive evaluations across multiple models and benchmarks demonstrate that Gambit strictly dominates existing baselines. Under identical hardware constraints, our method yields up to a +6.7\% absolute accuracy gain on HMMT-24 and +3.3\% on AIME-25 over pruning baselines, delivers >2times higher throughput on trace completion, and reduces total token consumption by up to 68.5\% relative to standard parallel sampling.

Large language model (LLM) routing distributes inference work across different models, advancing the cost-quality Pareto frontier of LLM serving. While coarse-grained routing at the session or query level has been widely adopted in production systems, recent algorithmic work shows that fine-grained token-level routing can yield substantial efficiency and quality gains. However, efficiently serving token-level routed inference poses significant challenges to existing systems. Built on single-LLM assumptions, current systems suffer from severe step desynchronization and frequent batch admission delays under token-level routing, and they also impose high implementation complexity on developers. To address these challenges, we design TokenRouter, an efficient and developer-friendly serving system for token-level routed LLM inference. TokenRouter follows the principle of request-centric programming, model-centric execution: developers describe routing logic from the perspective of a single request, while the runtime launches a subserver for each LLM and dispatches requests asynchronously. Each subserver employs a delayed-batching scheduler, whose optimal hyperparameters are derived from a mathematical throughput model of the system. Across diverse routing algorithms, workloads, and model pairs, TokenRouter achieves 2.01-64.15x higher decoding throughput than existing systems, substantially advancing the serving efficiency of token-level LLM routing. Our code is available at https://github.com/thu-nics/TokenRouter.

We present the design of a new large scale orchestration layer for accelerators. Our system, Pathways, is explicitly designed to enable exploration of new systems and ML research ideas, while retaining state of the art performance for current models. Pathways uses a sharded dataflow graph of asynchronous operators that consume and produce futures, and efficiently gang-schedules heterogeneous parallel computations on thousands of accelerators while coordinating data transfers over their dedicated interconnects. Pathways makes use of a novel asynchronous distributed dataflow design that lets the control plane execute in parallel despite dependencies in the data plane. This design, with careful engineering, allows Pathways to adopt a single-controller model that makes it easier to express complex new parallelism patterns. We demonstrate that Pathways can achieve performance parity (~100% accelerator utilization) with state-of-the-art systems when running SPMD computations over 2048 TPUs, while also delivering throughput comparable to the SPMD case for Transformer models that are pipelined across 16 stages, or sharded across two islands of accelerators connected over a data center network.

The emergence of large-scale Mixture of Experts (MoE) models has marked a significant advancement in artificial intelligence, offering enhanced model capacity and computational efficiency through conditional computation. However, the deployment and inference of these models present substantial challenges in terms of computational resources, latency, and energy efficiency. This comprehensive survey systematically analyzes the current landscape of inference optimization techniques for MoE models across the entire system stack. We first establish a taxonomical framework that categorizes optimization approaches into model-level, system-level, and hardware-level optimizations. At the model level, we examine architectural innovations including efficient expert design, attention mechanisms, various compression techniques such as pruning, quantization, and knowledge distillation, as well as algorithm improvement including dynamic routing strategies and expert merging methods. At the system level, we investigate distributed computing approaches, load balancing mechanisms, and efficient scheduling algorithms that enable scalable deployment. Furthermore, we delve into hardware-specific optimizations and co-design strategies that maximize throughput and energy efficiency. This survey not only provides a structured overview of existing solutions but also identifies key challenges and promising research directions in MoE inference optimization. Our comprehensive analysis serves as a valuable resource for researchers and practitioners working on large-scale deployment of MoE models in resource-constrained environments. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE inference optimization research, we have established a repository accessible at https://github.com/MoE-Inf/awesome-moe-inference/.

We present Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for interactive deployment. We designed the model to maximize server throughput under high user throughput constraints. In interactive serving workloads on a single 8xB200 node, Puzzle-75B-A9B achieves approximately 2x higher server throughput than Nemotron-3-Super at matched user throughput constraints. In ultra-long-context deployment on a single H100 GPU, the compressed model increases 1M-token concurrency from 1 request to 8 requests. Puzzle-75B-A9B is constructed using a multi-stage pipeline that combines the Iterative Puzzle compression framework with knowledge distillation, reinforcement learning, quantization, and a Multi-Token Prediction head. The compression process jointly optimizes heterogeneous MoE pruning, active parameter budget, and Mamba pruning to improve inference efficiency while preserving model quality. We evaluate Puzzle-75B-A9B on a broad suite of reasoning, coding, multilingual, long-context, and agentic benchmarks. Despite substantial compression, the model retains strong downstream accuracy relative to the parent model across a wide range of tasks. These results demonstrate that large hybrid MoE models can be substantially optimized for deployment efficiency while maintaining strong downstream capability.

Graphics processing unit (GPU) architectures are growing in size to meet the increasing compute and memory requirements. As GPU sizes increase, intra-socket wire transfer delay increases significantly. While previous research has optimized for compute and memory locality within a socket, the spatial impact on inter-GPU communication has not been well-studied. We introduce the term non-uniform network access (NUNA) to describe this emerging optimization dimension in multi-GPU systems. We specifically focus on latency-sensitive collective communication, common in machine learning inference. First, we highlight the need for NUNA-aware routing (NAR), choosing optimized, spatially-aware inter-GPU paths in large scale-up network topologies. Second, we introduce NUNA-aware placement (NAP), placing threadblocks and data near I/O to optimize the inter-GPU traffic. We demonstrate that the NAP optimizations alone offer up to 1.5x collective speedups over a locality-unaware baseline. Combining NAP with NAR yields up to 1.8x faster collectives over the locality-unaware baseline. This leads to 7% mean (28% max) time per output token speedup in machine learning inference.

Modern large-scale LLM pretraining benefits from utilizing Pipeline Parallelism; however, synchronous implementations leave GPUs idle during pipeline bubbles, wasting computational resources. Asynchronous Pipeline Parallelism eliminates these bubbles, maximizing throughput at the cost of gradient staleness. Among asynchronous schedules, PipeDream-2BW is particularly appealing: unlike the original PipeDream schedule, it ensures a constant one-step gradient delay regardless of pipeline depth. However, its adoption remains limited due to the common belief that optimizing under staleness is fundamentally unstable. In this work, we challenge this assumption, demonstrating that degradation under one-step delay depends strongly on optimizer choice rather than being an intrinsic limitation. We provide the first comprehensive empirical analysis showing that while AdamW, the predominant optimizer at the time when PipeDream-2BW was introduced, indeed suffers from severe degradation, recent methods like Muon exhibit strong robustness under a one-step delay. We introduce an optimizer-agnostic Error Feedback-inspired correction to further mitigate delay effects. We provide supporting theoretical analysis demonstrating convergence for Muon with and without this correction. Extensive evaluation on models up to 10B parameters confirms that our strategies bridge the performance gap with synchronous training, highlighting the practical potential of asynchronous pipeline parallelism at scale.

We consider online scheduling on unrelated (heterogeneous) machines in a speed-oblivious setting, where an algorithm is unaware of the exact job-dependent processing speeds. We show strong impossibility results for clairvoyant and non-clairvoyant algorithms and overcome them in models inspired by practical settings: (i) we provide competitive learning-augmented algorithms, assuming that (possibly erroneous) predictions on the speeds are given, and (ii) we provide competitive algorithms for the speed-ordered model, where a single global order of machines according to their unknown job-dependent speeds is known. We prove strong theoretical guarantees and evaluate our findings on a representative heterogeneous multi-core processor. These seem to be the first empirical results for scheduling algorithms with predictions that are evaluated in a non-synthetic hardware environment.

The surge in generative AI workloads has created a need for scalable inference systems that can flexibly harness both GPUs and specialized accelerators while containing operational costs. This paper proposes a hardware-agnostic control loop that adaptively allocates requests across heterogeneous accelerators based on real-time cost and capacity signals. The approach sustains low latency and high throughput by dynamically shifting between cost-optimized and capacity-optimized modes, ensuring the most efficient use of expensive compute resources under fluctuating availability. Evaluated using the Stable Diffusion model, the framework consistently meets latency targets, automatically redirects traffic during capacity shortfalls, and capitalizes on lower-cost accelerators when possible. These results highlight how a feedback-driven deployment strategy, spanning the entire software and hardware stack, can help organizations efficiently scale generative AI workloads while maintaining resilience in the face of limited accelerator capacity.

In recent years, large language models have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, deploying these models for real-world applications often requires efficient inference solutions to handle the computational demands. In this paper, we explore the utilization of CPUs for accelerating the inference of large language models. Specifically, we introduce a parallelized approach to enhance throughput by 1) Exploiting the parallel processing capabilities of modern CPU architectures, 2) Batching the inference request. Our evaluation shows the accelerated inference engine gives an 18-22x improvement in the generated token per sec. The improvement is more with longer sequence and larger models. In addition to this, we can also run multiple workers in the same machine with NUMA node isolation to further improvement in tokens/s. Table 2, we have received 4x additional improvement with 4 workers. This would also make Gen-AI based products and companies environment friendly, our estimates shows that CPU usage for Inference could reduce the power consumption of LLMs by 48.9% while providing production ready throughput and latency.

Serving numerous users and requests concurrently requires good fairness in Large Language Models (LLMs) serving system. This ensures that, at the same cost, the system can meet the Service Level Objectives (SLOs) of more users , such as time to first token (TTFT) and time between tokens (TBT), rather than allowing a few users to experience performance far exceeding the SLOs. To achieve better fairness, the preemption-based scheduling policy dynamically adjusts the priority of each request to maintain balance during runtime. However, existing systems tend to overly prioritize throughput, overlooking the overhead caused by preemption-induced context switching, which is crucial for maintaining fairness through priority adjustments. In this work, we identify three main challenges that result in this overhead. 1) Inadequate I/O utilization. 2) GPU idleness. 3) Unnecessary I/O transmission during multi-turn conversations. Our key insight is that the block-based KV cache memory policy in existing systems, while achieving near-zero memory waste, leads to discontinuity and insufficient granularity in the KV cache memory. To respond, we introduce FastSwitch, a fairness-aware serving system that not only aligns with existing KV cache memory allocation policy but also mitigates context switching overhead. Our evaluation shows that FastSwitch outperforms the state-of-the-art LLM serving system vLLM with speedups of 1.4-11.2x across different tail TTFT and TBT.

The emergence of Large Language Models (LLMs) has necessitated the adoption of parallel training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, we have found that the efficiency of current parallel training is often suboptimal, largely due to the following two main issues. Firstly, hardware failures are inevitable, leading to interruptions in the training tasks. The inability to quickly identify the faulty components results in a substantial waste of GPU resources. Secondly, since GPUs must wait for parameter synchronization to complete before proceeding to the next round of computation, network congestions can greatly increase the waiting time for GPUs. To address these challenges, this paper introduces a communication-driven solution, namely the C4. The key insights of C4 are two folds. First, in parallel training, collective communication exhibits periodic and homogeneous characteristics, so any anomalies are certainly due to some form of hardware malfunction. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving few large flows, allows C4 to efficiently execute traffic planning, substantially reducing network congestion. C4 has been extensively implemented across our production systems, cutting error-induced overhead by roughly 30% and enhancing runtime performance by about 15% for certain applications with moderate communication costs.

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-sharding 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 1.78x (1.36x on average) compared to vLLM, the most widely used state-of-the-art LLM inference engine.

TeleChat3-MoE is the latest series of TeleChat large language models, featuring a Mixture-of-Experts (MoE) architecture with parameter counts ranging from 105 billion to over one trillion,trained end-to-end on Ascend NPU cluster. This technical report mainly presents the underlying training infrastructure that enables reliable and efficient scaling to frontier model sizes. We detail systematic methodologies for operator-level and end-to-end numerical accuracy verification, ensuring consistency across hardware platforms and distributed parallelism strategies. Furthermore, we introduce a suite of performance optimizations, including interleaved pipeline scheduling, attention-aware data scheduling for long-sequence training,hierarchical and overlapped communication for expert parallelism, and DVM-based operator fusion. A systematic parallelization framework, leveraging analytical estimation and integer linear programming, is also proposed to optimize multi-dimensional parallelism configurations. Additionally, we present methodological approaches to cluster-level optimizations, addressing host- and device-bound bottlenecks during large-scale training tasks. These infrastructure advancements yield significant throughput improvements and near-linear scaling on clusters comprising thousands of devices, providing a robust foundation for large-scale language model development on hardware ecosystems.

Test-time compute scaling has emerged as a powerful paradigm for enhancing mathematical reasoning in large language models (LLMs) by allocating additional computational resources during inference. However, current methods employ uniform resource distribution across all reasoning sub-problems, creating fundamental bottlenecks where challenging sub-problems receive insufficient attention while routine operations consume disproportionate resources. This uniform allocation creates performance bottlenecks where additional computational resources yield diminishing returns. Inspired by dual-process theory, we propose SCALE (Selective Resource Allocation), a framework that selectively allocates computational resources based on sub-problem difficulty. SCALE operates through four stages: (1) problem decomposition into sequential reasoning sub-problems, (2) difficulty assessment of each sub-problem to distinguish between routine operations and computationally challenging sub-problems, (3) selective processing mode assignment between System 1 for simple sub-problems and System 2 for complex ones, and (4) sequential execution with context propagation. By concentrating resources on challenging sub-problems while processing routine operations efficiently, SCALE achieves substantial performance improvements with superior resource utilization. Extensive experiments demonstrate that SCALE significantly outperforms uniform scaling baselines, achieving accuracy improvements of up to 13.75 percentage points (57.50% to 71.25% on AIME25) while reducing computational costs by 33%-53%, representing a major advance in test-time scaling that addresses fundamental limitations of current approaches.

We rethink test-time scaling laws from a practical efficiency perspective, revealing that the effectiveness of smaller models is significantly overestimated. Prior work, grounded in compute-optimality, overlooks critical memory access bottlenecks introduced by inference-time strategies (e.g., Best-of-N, long CoTs). Our holistic analysis, spanning models from 0.6B to 32B parameters, reveals a new Kinetics Scaling Law that better guides resource allocation by incorporating both computation and memory access costs. Kinetics Scaling Law suggests that test-time compute is more effective when used on models above a threshold than smaller ones. A key reason is that in TTS, attention, rather than parameter count, emerges as the dominant cost factor. Motivated by this, we propose a new scaling paradigm centered on sparse attention, which lowers per-token cost and enables longer generations and more parallel samples within the same resource budget. Empirically, we show that sparse attention models consistently outperform dense counterparts, achieving over 60 points gains in low-cost regimes and over 5 points gains in high-cost regimes for problem-solving accuracy on AIME, encompassing evaluations on state-of-the-art MoEs. These results suggest that sparse attention is essential for realizing the full potential of test-time scaling because, unlike training, where parameter scaling saturates, test-time accuracy continues to improve through increased generation. The code is available at https://github.com/Infini-AI-Lab/Kinetics.

Training large language models (LLMs) is known to be challenging because of the huge computational and memory capacity requirements. To address these issues, it is common to use a cluster of GPUs with 3D parallelism, which splits a model along the data batch, pipeline stage, and intra-layer tensor dimensions. However, the use of 3D parallelism produces the additional challenge of finding the optimal number of ways on each dimension and mapping the split models onto the GPUs. Several previous studies have attempted to automatically find the optimal configuration, but many of these lacked several important aspects. For instance, the heterogeneous nature of the interconnect speeds is often ignored. While the peak bandwidths for the interconnects are usually made equal, the actual attained bandwidth varies per link in real-world clusters. Combined with the critical path modeling that does not properly consider the communication, they easily fall into sub-optimal configurations. In addition, they often fail to consider the memory requirement per GPU, often recommending solutions that could not be executed. To address these challenges, we propose Pipette, which is an automatic fine-grained LLM training configurator for real-world clusters. By devising better performance models along with the memory estimator and fine-grained individual GPU assignment, Pipette achieves faster configurations that satisfy the memory constraints. We evaluated Pipette on large clusters to show that it provides a significant speedup over the prior art. The implementation of Pipette is available at https://github.com/yimjinkyu1/date2024_pipette.

Large language models (LLMs) require vast amounts of GPU compute to train, but limited availability and high costs of GPUs make homogeneous clusters impractical for many organizations. Instead, assembling heterogeneous clusters by pooling together GPUs of different generations allows them to achieve higher aggregate compute and make use of all available GPUs. However, training on heterogeneous clusters presents several challenges, including load balancing across GPUs, optimizing memory usage to accommodate varying memory capacities, and ensuring communication-efficient training over diverse network interconnects potentially spanning multiple datacenters. In this paper, we make the case that efficient training on heterogeneous clusters requires (1) the integration of pipeline parallelism and data parallelism in a manner that is both communication- and memory-efficient, and (2) a more adaptable configuration of pipeline and data parallelism, which includes the capability to flexibly partition GPUs into asymmetric pipeline parallel stages and to incorporate heterogeneous GPUs within the same data parallelism group. We propose Zorse, the first system to unify all these capabilities while incorporating a planner that automatically configures training strategies for a given workload. Our evaluation shows that Zorse significantly outperforms state-of-the-art systems in heterogeneous training scenarios.

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

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt to produce one output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler Sarathi-Serve inspired by the techniques we originally proposed for optimizing throughput in Sarathi. Sarathi-Serve leverages chunked-prefills from Sarathi to create stall-free schedules that can add new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Our evaluation shows that Sarathi-Serve improves serving throughput within desired latency SLOs of Mistral-7B by up to 2.6x on a single A100 GPU and up to 6.9x for Falcon-180B on 8 A100 GPUs over Orca and vLLM.

Batch inference for agentic workloads stresses the GPU key-value (KV) cache in a sustained and cumulative manner, often causing severe throughput degradation well before memory capacity is exhausted. We identify this phenomenon as middle-phase thrashing, a previously under-characterized pathology in which cache efficiency collapses as long-lived agents accumulate state over time. We argue that mitigating this pathology requires moving beyond reactive, request-level cache management to proactive, agent-level admission control. Drawing inspiration from congestion control in distributed systems, we view the KV cache as a shared resource whose efficient utilization depends on feedback-driven regulation. Based on this insight, we present CONCUR, a lightweight control layer that regulates agent admission to bound aggregate cache pressure while preserving execution continuity. CONCUR adapts a cache-aware control algorithm to dynamically adjust the number of active agents using runtime cache signals. Across large models and real-world agent workloads, CONCUR prevents middle-phase thrashing and improves batch inference throughput by up to 4.09x on Qwen3-32B and 1.9x on DeepSeek-V3, while remaining compatible with existing LLM serving systems.

We introduce Breadth-First Pipeline Parallelism, a novel training schedule which optimizes the combination of pipeline and data parallelism. Breadth-First Pipeline Parallelism lowers training time, cost and memory usage by combining a high GPU utilization with a small batch size per GPU, and by making use of fully sharded data parallelism. Experimentally, we observed an increase of up to 43% in training throughput for a 52 billion-parameter model using a small batch size per GPU compared to Megatron-LM, which would reduce the training time and cost by the same amount on a large GPU cluster.

Mixture-of-Experts (MoE) models are typically pre-trained with explicit load-balancing constraints to ensure statistically balanced expert routing. Despite this, we observe that even well-trained MoE models exhibit significantly imbalanced routing. This behavior is arguably natural-and even desirable - as imbalanced routing allows models to concentrate domain-specific knowledge within a subset of experts. Expert parallelism (EP) is designed to scale MoE models by distributing experts across multiple devices, but with a less-discussed assumption of balanced routing. Under extreme imbalance, EP can funnel a disproportionate number of tokens to a small number of experts, leading to compute- and memory-bound failures on overloaded devices during post-training or inference, where explicit load balancing is often inapplicable. We propose Least-Loaded Expert Parallelism (LLEP), a novel EP algorithm that dynamically reroutes excess tokens and associated expert parameters from overloaded devices to underutilized ones. This ensures that all devices complete their workloads within the minimum collective latency while respecting memory constraints. Across different model scales, LLEP achieves up to 5x speedup and 4x reduction in peak memory usage compared to standard EP. This enables faster and higher-throughput post-training and inference, with ~1.9x faster for gpt-oss-120b. We support our method with extensive theoretical analysis and comprehensive empirical evaluations, including ablation studies. These results illuminate key trade-offs and enable a principled framework for hardware-specific hyper-parameter tuning to achieve optimal performance.

The recent surge of open-source large language models (LLMs) enables developers to create AI-based solutions while maintaining control over aspects such as privacy and compliance, thereby providing governance and ownership of the model deployment process. To utilize these LLMs, inference engines are needed. These engines load the model's weights onto available resources, such as GPUs, and process queries to generate responses. The speed of inference, or performance, of the LLM, is critical for real-time applications, as it computes millions or billions of floating point operations per inference. Recently, advanced inference engines such as vLLM have emerged, incorporating novel mechanisms such as efficient memory management to achieve state-of-the-art performance. In this paper, we analyze the performance, particularly the throughput (tokens generated per unit of time), of 20 LLMs using two inference libraries: vLLM and HuggingFace's pipelines. We investigate how various hyperparameters, which developers must configure, influence inference performance. Our results reveal that throughput landscapes are irregular, with distinct peaks, highlighting the importance of hyperparameter optimization to achieve maximum performance. We also show that applying hyperparameter optimization when upgrading or downgrading the GPU model used for inference can improve throughput from HuggingFace pipelines by an average of 9.16% and 13.7%, respectively.

We introduce stale synchronous parallel as a mode of execution in parallel sparse triangular linear system solve and present a general directed-acyclic-graph scheduler capable of producing such schedules. Stale-synchronous-parallel schedules allow the overlap of synchronisation and compute which results in a geometric-mean speed-up of 7-30% of our scheduler, ElasticDivide, over state-of-the-art synchronous scheduler GrowLocal on an ARM machine using 48 cores. On an x86 machine using 48 cores, we report geometric-mean speed-ups of 19-60% over SpMP.

Many LLM tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix. The KV context that is about to be reused may prematurely be evicted with the implicit cache management. Even if not evicted, the lifetime of the shared KV context is extended since requests sharing the same context are not scheduled together, resulting in larger memory usage. These streaming oriented systems schedule the requests in the first-come-first-serve or similar order. As a result, the requests with larger ratio of decoding steps may be scheduled too late to be able to mix with the prefill chunks to increase the hardware utilization. Besides, the token and request number based batching can limit the size of token-batch, which keeps the GPU from saturating for the iterations dominated by decoding tokens. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best, which also shrinks the lifetime of common KV memory. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Extensive evaluation shows that BatchLLM outperforms vLLM by 1.1x to 2x on a set of microbenchmarks and two typical industry workloads.

In recent years, the training requirements of many state-of-the-art Deep Learning (DL) models have scaled beyond the compute and memory capabilities of a single processor, and necessitated distribution among processors. Training such massive models necessitates advanced parallelism strategies to maintain efficiency. However, such distributed DL parallelism strategies require a varied mixture of collective and point-to-point communication operations across a broad range of message sizes and scales. Examples of models using advanced parallelism strategies include Deep Learning Recommendation Models (DLRM) and Mixture-of-Experts (MoE). Communication libraries' performance varies wildly across different communication operations, scales, and message sizes. We propose MCR-DL: an extensible DL communication framework that supports all point-to-point and collective operations while enabling users to dynamically mix-and-match communication backends for a given operation without deadlocks. MCR-DL also comes packaged with a tuning suite for dynamically selecting the best communication backend for a given input tensor. We select DeepSpeed-MoE and DLRM as candidate DL models and demonstrate a 31% improvement in DS-MoE throughput on 256 V100 GPUs on the Lassen HPC system. Further, we achieve a 20% throughput improvement in a dense Megatron-DeepSpeed model and a 25% throughput improvement in DLRM on 32 A100 GPUs with the Theta-GPU HPC system.

Cache partitioning techniques have been successfully adopted to mitigate interference among concurrently executing real-time tasks on multi-core processors. Considering that the execution time of a cache-sensitive task strongly depends on the cache available for it to use, co-optimizing cache partitioning and task allocation improves the system's schedulability. In this paper, we propose a hybrid multi-layer design space exploration technique to solve this multi-resource management problem. We explore the interplay between cache partitioning and schedulability by systematically interleaving three optimization layers, viz., (i) in the outer layer, we perform a breadth-first search combined with proactive pruning for cache partitioning; (ii) in the middle layer, we exploit a first-fit heuristic for allocating tasks to cores; and (iii) in the inner layer, we use the well-known recurrence relation for the schedulability analysis of non-preemptive fixed-priority (NP-FP) tasks in a uniprocessor setting. Although our focus is on NP-FP scheduling, we evaluate the flexibility of our framework in supporting different scheduling policies (NP-EDF, P-EDF) by plugging in appropriate analysis methods in the inner layer. Experiments show that, compared to the state-of-the-art techniques, the proposed framework can improve the real-time schedulability of NP-FP task sets by an average of 15.2% with a maximum improvement of 233.6% (when tasks are highly cache-sensitive) and a minimum of 1.6% (when cache sensitivity is low). For such task sets, we found that clustering similar-period (or mutually compatible) tasks often leads to higher schedulability (on average 7.6%) than clustering by cache sensitivity. In our evaluation, the framework also achieves good results for preemptive and dynamic-priority scheduling policies.

Enterprise LLM deployment faces a critical scalability challenge: organizations must optimize models systematically to scale AI initiatives within constrained compute budgets, yet the specialized expertise required for manual optimization remains a niche and scarce skillset. This challenge is particularly evident in managing GPU utilization across heterogeneous infrastructure while enabling teams with diverse workloads and limited LLM optimization experience to deploy models efficiently. We present OptiKIT, a distributed LLM optimization framework that democratizes model compression and tuning by automating complex optimization workflows for non-expert teams. OptiKIT provides dynamic resource allocation, staged pipeline execution with automatic cleanup, and seamless enterprise integration. In production, it delivers more than 2x GPU throughput improvement while empowering application teams to achieve consistent performance improvements without deep LLM optimization expertise. We share both the platform design and key engineering insights into resource allocation algorithms, pipeline orchestration, and integration patterns that enable large-scale, production-grade democratization of model optimization. Finally, we open-source the system to enable external contributions and broader reproducibility.

Modern Intel AI PCs ship capable integrated GPUs and NPUs with 16+ GB of unified memory, and they spend considerable time idle. That is not enough memory to fit a large model such as a 70B-parameter LLM. We show that a handful of AIPCs, working together over an ordinary network, can serve models beyond the capability of any single one. We use pipeline parallelism: a model is split by layer into per-stage shards, each pre-compiled into an OpenVINO graph, so that every machine runs one shard and passes activations to the next. Three techniques make this fast enough to be useful. First, we recover the speed of the unsplit model: a naive per-stage export runs well below monolithic inference because it misses an OpenVINO GPU optimization, and injecting a beam_idx Gather into each shard triggers that optimization (the IndirectKVCache fusion) and brings the shards to parity. Second, we leverage speculative decoding on stateful OpenVINO models. Third, the pipeline serves several users at once by interleaving their requests across the stages, each request carrying its own cache (micro-batching). Together, a two-node Llama 3.1 8B INT4 pipeline serves two concurrent users at 1.79x the single-user throughput of the unsplit model on the same hardware, and the gap widens under simulated wide-area latency. The same design scales to a 70B model that no single fleet member can hold: a four-node deployment of Lunar Lake AI PCs on Intel Tiber Cloud serves a single user at interactive speed, with output token-for-token identical to the same four-node pipeline decoding without speculation. Code, raw benchmark logs, and reproduction scripts ship as a self-contained package at https://github.com/labscommunity/pipeline-sharded-inference-paper (in the top-level reproduction/ directory).

Resource constraints increasingly determine what can be trained, fine-tuned, and deployed in large language models (LLMs), yet efficiency is often studied through isolated techniques rather than as an interacting system of limits. This survey adopts a constraint-centric perspective and organizes recent progress around three coupled bottlenecks: data efficiency (what to train on), memory efficiency (how to fit training), and compute budget awareness (when and where to spend FLOPs). On the data axis, we review selection and pruning methods that maximize learning per token, ranging from scalable proxy signals based on learning dynamics to gradient- and influence-based scoring, as well as difficulty-aware and curriculum-style strategies. We highlight emerging evidence that different notions of good data dominate in different regimes, implying that optimal subsets depend on the task objective and resource budget rather than being universal. On the systems side, we show that GPU memory, not raw compute, is often the dominant bottleneck in fine-tuning, and that effective scaling requires jointly reducing weight storage, optimizer states, and activation memory rather than optimizing any single component in isolation. Beyond memory, we frame training and inference as compute-governed processes in which optimization, data selection, and decoding must explicitly account for finite FLOP budgets. We review evidence for compute-optimal allocation and stopping rules, where computation should be halted or reallocated once marginal performance gains fall below a budget-dependent threshold. Together, these results unify compute-aware data selection, scaling laws, and adaptive inference under a common principle of resource-conditioned decision-making.

Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.

Solving Mixed-Integer Programming (MIP) problems often requires substantial computational resources due to their combinatorial nature. Parallelization has emerged as a critical strategy to accelerate solution times and enhance scalability to tackle large, complex instances. This paper investigates the parallelization capabilities of Balans, a recently proposed multi-armed bandits-based adaptive large neighborhood search for MIPs. While Balans's modular architecture inherently supports parallel exploration of diverse parameter configurations, this potential has not been thoroughly examined. To address this gap, we introduce ParBalans, an extension that leverages both solver-level and algorithmic-level parallelism to improve performance on challenging MIP instances. Our experimental results demonstrate that ParBalans exhibits competitive performance compared to the state-of-the-art commercial solver Gurobi, particularly on hard optimization benchmarks.

Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to Autoregressive Models (ARMs), utilizing parallel decoding to overcome sequential bottlenecks. However, existing research focuses primarily on kernel-level optimizations, lacking a holistic serving framework that addresses the unique memory dynamics of diffusion processes in production. We identify a critical "memory footprint crisis" specific to dLLMs, driven by monolithic logit tensors and the severe resource oscillation between compute-bound "Refresh" phases and bandwidth-bound "Reuse" phases. To bridge this gap, we present dLLM-Serve, an efficient dLLM serving system that co-optimizes memory footprint, computational scheduling, and generation quality. dLLM-Serve introduces Logit-Aware Activation Budgeting to decompose transient tensor peaks, a Phase-Multiplexed Scheduler to interleave heterogeneous request phases, and Head-Centric Sparse Attention to decouple logical sparsity from physical storage. We evaluate dLLM-Serve on diverse workloads (LiveBench, Burst, OSC) and GPUs (RTX 4090, L40S). Relative to the state-of-the-art baseline, dLLM-Serve improves throughput by 1.61times-1.81times on the consumer-grade RTX 4090 and 1.60times-1.74times on the server-grade NVIDIA L40S, while reducing tail latency by nearly 4times under heavy contention. dLLM-Serve establishes the first blueprint for scalable dLLM inference, converting theoretical algorithmic sparsity into tangible wall-clock acceleration across heterogeneous hardware.

GPU-initiated communication lets GPU threads post RDMA operations directly to the NIC. It underpins NVSHMEM, NCCL GIN, and DeepEP, which serve the fine-grained, latency-critical communication of Mixture-of-Experts (MoE) models, yet its performance characteristics and optimizations remain scarcely documented beyond source code, and library comparisons fail to separate the costs of the hardware mechanism from those of the library around it. This paper dissects GPU-initiated communication at the GPU-NIC boundary. We first detail the GPU-side network path: queue placement, work-request construction, doorbell ordering, and completion semantics. We then introduce mini-gda and mini-proxy, minimal transports for the GPU and CPU-proxy submission paths, and measure them alongside NVSHMEM IBGDA, NCCL GIN, DeepEP, UCCL-EP, MSCCL++, and fabric-lib on NVIDIA H100, H200, B200, and GB200 platforms. A minimal GPU path issues an operation in 0.7 μs and completes in 4.0 μs; libraries add up to 4.6 μs of issue time through queue management, memory ordering, and completion scope, and issue time scales with the SM clock. A tuned CPU proxy matches or beats the GPU path at idle, at the cost of a dedicated core whose operating state sets its latency and throughput. On either path, sharing a queue with bulk traffic raises latency by one to three orders of magnitude. Reaching the 260 M msg/s ceiling of our InfiniBand platform requires doorbell batching and queue parallelism, and both have resource costs: communication code can reduce GPU block residency even when unused, and all-to-all traffic loses 59% of its NIC message rate at about 3,000 active connections. The submission path alone therefore does not predict communication performance. Our experiment code and results are available at https://github.com/ParCoreLab/Dissecting-GPU-Communication-Experiments.

This paper introduces bucket calculus, a novel mathematical framework that fundamentally transforms the computational complexity landscape of parallel machine scheduling optimization. We address the strongly NP-hard problem P2|r_j|C_{max} through an innovative adaptive temporal discretization methodology that achieves exponential complexity reduction from O(T^n) to O(B^n) where B ll T, while maintaining near-optimal solution quality. Our bucket-indexed mixed-integer linear programming (MILP) formulation exploits dimensional complexity heterogeneity through precision-aware discretization, reducing decision variables by 94.4\% and achieving a theoretical speedup factor 2.75 times 10^{37} for 20-job instances. Theoretical contributions include partial discretization theory, fractional bucket calculus operators, and quantum-inspired mechanisms for temporal constraint modeling. Empirical validation on instances with 20--400 jobs demonstrates 97.6\% resource utilization, near-perfect load balancing (σ/μ= 0.006), and sustained performance across problem scales with optimality gaps below 5.1\%. This work represents a paradigm shift from fine-grained temporal discretization to multi-resolution precision allocation, bridging the fundamental gap between exact optimization and computational tractability for industrial-scale NP-hard scheduling problems.

I/O performance is crucial to efficiency in data-intensive scientific computing; but tuning large-scale storage systems is complex, costly, and notoriously manpower-intensive, making it inaccessible for most domain scientists. To address this problem, we propose STELLAR, an autonomous tuner for high-performance parallel file systems. Our evaluations show that STELLAR almost always selects near-optimal parameter configurations for parallel file systems within the first five attempts, even for previously unseen applications. STELLAR differs fundamentally from traditional autotuning methods, which often require hundreds of thousands of iterations to converge. Powered by large language models (LLMs), STELLAR enables autonomous end-to-end agentic tuning by (1) accurately extracting tunable parameters from software manuals, (2) analyzing I/O trace logs generated by applications, (3) selecting initial tuning strategies, (4) rerunning applications on real systems and collecting I/O performance feedback, (5) adjusting tuning strategies and repeating the tuning cycle, and (6) reflecting on and summarizing tuning experiences into reusable knowledge for future optimizations. STELLAR integrates retrieval-augmented generation (RAG), tool execution, LLM-based reasoning, and a multiagent design to stabilize reasoning and combat hallucinations. We evaluate the impact of each component on optimization outcomes, providing design insights for similar systems in other optimization domains. STELLAR's architecture and empirical results highlight a promising approach to complex system optimization, especially for problems with large search spaces and high exploration costs, while making I/O tuning more accessible to domain scientists with minimal added resources.

Decoder-only Transformer models such as GPT have demonstrated superior performance in text generation, by autoregressively predicting the next token. However, the performance of GPT is bounded by low compute-to-memory-ratio and high memory access. Throughput-oriented architectures such as GPUs target parallel processing rather than sequential token generation, and are not efficient for GPT acceleration, particularly on-device inference applications. Process-in-memory (PIM) architectures can significantly reduce data movement and provide high computation parallelism, and are promising candidates to accelerate GPT inference. In this work, we propose PIM-GPT that aims to achieve high throughput, high energy efficiency and end-to-end acceleration of GPT inference. PIM-GPT leverages DRAM-based PIM solutions to perform multiply-accumulate (MAC) operations on the DRAM chips, greatly reducing data movement. A compact application-specific integrated chip (ASIC) is designed and synthesized to initiate instructions to PIM chips and support data communication along with necessary arithmetic computations. At the software level, the mapping scheme is designed to maximize data locality and computation parallelism by partitioning a matrix among DRAM channels and banks to utilize all in-bank computation resources concurrently. We develop an event-driven clock-cycle accurate simulator to validate the efficacy of the proposed PIM-GPT architecture. Overall, PIM-GPT achieves 41-137times, 631-1074times speedup and 339-1085times, 890-1632times energy efficiency over GPU and CPU baseline, respectively, on 8 GPT models with up to 1.4 billion parameters.

When a Mixture-of-Experts (MoE) model does not fit in GPU memory, batch-1 decoding is dominated by copying expert weights from host memory. Cache-aware routing reduces these copies by biasing the router toward resident experts, but combined with asynchronous prefetching it makes the model's outputs depend on transfer timing, so a harness that does not simulate that timing no longer evaluates the deployed model. We propose prefetch-steered routing (PSR): routing conditions on a deterministic logical cache, into which at most n experts predicted by the next layer's router are inserted per layer and token; physical copies affect only latency. On OLMoE-1B-7B with 50% and 25% of experts resident on an RTX 4090, PSR decodes 2.2-2.3x faster than exact routing with LRU and 25-30% faster than the Cache-Prior operating point, with mean benchmark differences of +0.1 and +0.5 points (95% lower bounds -0.8 and -0.3). In greedy generation its lead over Cache-Prior depends on the workload and reverses on HumanEval at 50% residency. On Moonlight-16B-A3B, PSR is 1.9x faster than exact routing with LRU at 37.5% residency and, at the latency of Cache-Prior, significantly better on three of eight benchmarks. On Qwen3.6-35B-A3B, where copies are cheap relative to computation, PSR keeps its advantage over Cache-Prior on the likelihood benchmarks but not its speed. Distillation-trained gates did not help. Speculative decoding with block-level expert selection was slower than PSR with a 1B autoregressive drafter on OLMoE, but with a block drafter on Qwen3.6 it was 2.1-2.9x faster than any autoregressive configuration with at most its GPU memory on the 64 prompts we timed.

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

The rapid advancement of Large Language Models (LLMs) has driven the need for more efficient serving strategies. In this context, efficiency refers to the proportion of requests that meet their Service Level Objectives (SLOs), particularly for Time To First Token (TTFT) and Time Between Tokens (TBT). However, existing systems often prioritize one metric at the cost of the other. We present Ascendra, an LLM serving system designed to meet both TTFT and TBT SLOs simultaneously. The core insight behind Ascendra is that a request's urgency evolves as it approaches its deadline. To leverage this, Ascendra partitions GPU resources into two types of instances: low-priority and high-priority. Low-priority instances maximize throughput by processing requests out of arrival order, but at the risk of request starvation. To address this, Ascendra employs a performance model to predict requests at risk of missing their SLOs and proactively offloads them to high-priority instances. High-priority instances are optimized for low-latency execution and handle urgent requests nearing their deadlines. This partitioned architecture enables Ascendra to effectively balance high throughput and low latency. Extensive evaluation shows that Ascendra improves system throughput by up to 1.7x compared to vLLM and Sarathi-Serve while meeting both TTFT and TBT SLOs.

Modern applications increasingly rely on inference serving systems to provide low-latency insights with a diverse set of machine learning models. Existing systems often utilize resource elasticity to scale with demand. However, many applications cannot rely on hardware scaling when deployed at the edge or other resource-constrained environments. In this work, we propose a model selection and scheduling algorithm that implements accuracy scaling to increase efficiency for these more constrained deployments. We show that existing schedulers that make decisions using profiled model accuracy are biased toward the label distribution present in the test dataset. To address this problem, we propose using ML models -- which we call SneakPeek models -- to dynamically adjust estimates of model accuracy, based on the underlying data. Furthermore, we greedily incorporate inference batching into scheduling decisions to improve throughput and avoid the overhead of swapping models in and out of GPU memory. Our approach employs a new notion of request priority, which navigates the trade-off between attaining high accuracy and satisfying deadlines. Using data and models from three real-world applications, we show that our proposed approaches result in higher-utility schedules and higher accuracy inferences in these hardware-constrained environments.

Parallel thinking enhances LLM reasoning by multi-path sampling and aggregation. In system-level evaluations, a global parallelism level N is allocated to all samples, typically set large to maximize overall dataset accuracy. However, due to sample heterogeneity, some samples can achieve comparable performance with a smaller N'< N, causing budget redundancy. This incompatibility between system-level efficacy and sample-level efficiency constitutes the overscaling curse. In this paper, we formalize and quantify the overscaling curse, showing its universality and severity in practice, and analyze its trigger mechanism. We then propose a lightweight method, T2, to break the overscaling curse, which utilizes latent representations to estimate the optimal parallelism level for each sample before decoding. Experiments show that T2 significantly reduces cost while maintaining comparable performance, enabling more efficient parallel thinking.

Multi-agent applications often execute complex tasks as multi-stage workflows, where each stage is an LLM call whose output becomes part of context for subsequent steps. Existing LLM serving systems largely assume homogeneous clusters with identical model replicas. This design overlooks the potential of heterogeneous deployments, where models of different sizes and capabilities enable finer trade-offs between latency and performance. However, heterogeneity introduces new challenges in scheduling across models with diverse throughput and performance. We present Chimera, a predictive scheduling system for multi-agent workflow serving on heterogeneous LLM clusters that jointly improves end-to-end latency and task performance. Chimera applies semantic routing to estimate per-model confidence scores for each request, predicts the total remaining output length of the workflow, and estimates per-model congestion using in-flight predicted token volumes for load balancing. We evaluate Chimera on representative agentic workflows for code generation and math reasoning using multiple heterogeneous LLM configurations. Across comparable settings, Chimera traces the best latency-performance frontier, reducing end-to-end latency by 1.2--2.4times and improving task performance by 8.0-9.5 percentage points on average over competitive baselines including vLLM.

Running large language models on consumer devices such as laptops and desktops is challenging because model weights often exceed GPU memory capacity, making offloading inference necessary to extend effective model capacity with CPU memory. Existing offloading systems, however, typically rely on coarse layer-level or expert-level scheduling, which overlooks substantial heterogeneity among tensors within the same layer and adapts poorly to changing hardware load conditions on such devices. This paper presents ATSInfer, a hybrid CPU-GPU inference system for consumer devices that performs offloading at tensor granularity. ATSInfer combines static tensor placement with load-aware dynamic transfer, and introduces asynchronous CPU-GPU coordination to efficiently schedule hardware storage, data movement, and computation across heterogeneous backends. We implement ATSInfer and evaluate it on representative consumer platforms using both dense and MoE models. Compared with existing systems, ATSInfer improves prefill throughput by up to 1.94times and decode throughput by up to 3.29times, while also increasing GPU utilization and making more effective use of PCIe bandwidth. These results show that ATSInfer can substantially improve the user experience of local LLM deployment on personal consumer devices.

Large Language Models increasingly rely on distributed techniques for their training and inference. These techniques require communication across devices which can reduce scaling efficiency as the number of devices increases. While some distributed techniques can overlap, and thus, hide this communication with independent computations, techniques such as Tensor Parallelism (TP) inherently serialize communication with model execution. One approach to hide this serialized communication is to interleave it with the producer operation (of the communicated data) in a fine-grained manner. However, this fine-grained interleaving of communication and computation in software can be difficult. Furthermore, as with any concurrent execution, it requires compute and memory resources to be shared between computation and communication, causing resource contention that reduces overlapping efficacy. To overcome these challenges, we propose T3 which applies hardware-software co-design to transparently overlap serialized communication while minimizing resource contention with compute. T3 transparently fuses producer operations with the subsequent communication via a simple configuration of the producer's output address space and requires minor software changes. At the hardware level, T3 adds a lightweight track and trigger mechanism to orchestrate the producer's compute, and communication. It further uses compute-enhanced memories for communication's attendant compute. As a result, T3 reduces resource contention, and efficiently overlaps serialized communication with computation. For important Transformer models like T-NLG, T3 speeds up communication-heavy sublayers by 30% geomean (max 47%) and reduces data movement by 22% geomean (max 36%). Furthermore, T3's benefits persist as models scale: geomean 29% for sublayers in sim500-billion parameter models, PALM and MT-NLG.

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4times with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm

In modern agentic AI training, the system bottleneck is shifting from rollout to update. Emerging sampling strategies such as tree-structured and step-wise RL greatly increase training sample volume while incurring relatively low marginal rollout cost, causing the update phase to dominate the end-to-end execution time. Crucially, this shift exposes a new optimization opportunity, as production traces reveal substantial prefix redundancy across training samples. In this paper, we propose psRL (prefix sharing for RL), a new training system for agentic AI designed to exploit prefix redundancy among training samples. Leveraging the global visibility and data immutability inherent to the update phase, psRL achieves efficient workload scheduling and memory management for distributed training. Specifically, psRL introduces two novel prefix-sharing mechanisms that enable flexible, fine-grained workload distribution across GPU workers, simultaneously optimizing prefix reuse and achieving load balancing. Moreover, psRL implements a new underlying KV cache manager that facilitates adaptable block-size allocation and dynamic KV caching, maximizing memory utilization while maintaining a high prefix hit rate. Evaluations using production traces demonstrate that psRL outperforms existing systems by up to 5.2x in throughput. The source code will be publicly available soon.

Training large language models requires distributing computation across many accelerators, yet practitioners select parallelism strategies (data, tensor, pipeline, ZeRO) through trial and error because no unified systematic framework predicts their behavior. We introduce placement semantics: each strategy is specified by how it places four training states (parameters, optimizer, gradients, activations) across devices using five modes (replicated, sharded, sharded-with-gather, materialized, offloaded). From placement alone, without implementation details, we derive memory consumption and communication volume. Our predictions match published results exactly: ZeRO-3 uses 8x less memory than data parallelism at 1.5x communication cost, as reported in the original paper. We prove two conditions (gradient integrity, state consistency) are necessary and sufficient for distributed training to match single-device results, and provide composition rules for combining strategies safely. The framework unifies ZeRO Stages 1-3, Fully Sharded Data Parallel (FSDP), tensor parallelism, and pipeline parallelism as instances with different placement choices.

While scaling laws guide compute allocation for LLM pre-training, analogous prescriptions for reinforcement learning (RL) post-training of large language models (LLMs) remain poorly understood. We study the compute-optimal allocation of sampling compute for on-policy RL methods in LLMs, framing scaling as a compute-constrained optimization over three resources: parallel rollouts per problem, number of problems per batch, and number of update steps. We find that the compute-optimal number of parallel rollouts per problem increases predictably with compute budget and then saturates. This trend holds across both easy and hard problems, though driven by different mechanisms: solution sharpening on easy problems and coverage expansion on hard problems. We further show that increasing the number of parallel rollouts mitigates interference across problems, while the number of problems per batch primarily affects training stability and can be chosen within a broad range. Validated across base models and data distributions, our results recast RL scaling laws as prescriptive allocation rules and provide practical guidance for compute-efficient LLM RL post-training.

Large language model (LLM) serving requires substantial memory to store model weights and KV caches. As models grow larger and contexts become longer, memory capacity and bandwidth increasingly become bottlenecks for serving performance. Agentic workloads compound this pressure through repeated interactions over growing contexts, making it increasingly important to retain KV state for reuse. High-bandwidth flash (HBF) offers a way to expand accelerator memory capacity for large language model (LLM) serving, but its access costs and limited write endurance complicate its use. We evaluate HBF for high-throughput agentic serving across system design and scheduling choices to understand when additional capacity improves serving performance and energy efficiency. We introduce an HBM-HBF-host hierarchical storage system and buffered cache-aware scheduling, and use trace-driven simulations to analyze their effects on performance, energy consumption, and HBF write lifetime. Across the evaluated workloads, the fastest HBF-augmented systems reduce completion time by 36.1-87.7% relative to HBM-only systems. Modeled energy savings reach 59.1%, with benefits depending on the workload and weight placement. Buffered cache-aware scheduling extends estimated HBF write lifetime from 1.21 to 14.82 years in the evaluated configuration. These results demonstrate the importance of coordinating data placement and scheduling to improve serving efficiency while sustaining a practical HBF write lifetime.

Multi-turn LLM serving accumulates dialogue history whose Key-Value (KV) cache grows with every turn and every user, quickly exceeding the model weights themselves and making memory -- not compute -- the binding constraint on throughput. Non-uniform KV compression, which allocates heterogeneous budgets across attention heads, preserves accuracy far better than uniform schemes, yet remains impractical: modern serving stacks assume identical KV lengths across heads, so heterogeneity traps freed memory as page fragmentation, spends up to 25% of prefill time reclaiming scattered pages, and skews GPU workloads that inflate decode latency by up to 1.7times or burn 15--20% of each decode step on re-planning. We observe that this heterogeneity need not be discovered at runtime: head-wise retention follows a two-level structural regularity -- an input-invariant head ranking with narrowly bounded per-head ratios -- that can be calibrated offline from as few as 50 samples. Building on this insight, we present Tangram, a serving framework that statically resolves what prior systems handle dynamically: Budget Reservation fixes each head's post-compression footprint at scheduling time, eliminating page reclamation; Ragged Paging clusters similar-budget heads into independent page tables, turning fragmentation into reclaimable memory; and Ahead-of-Time Load Balancing precomputes balanced GPU partitions with zero runtime planning. Implemented on vLLM, Tangram serves as a drop-in substrate for existing non-uniform compression methods, matching their accuracy while improving end-to-end throughput by up to 2.6times over the full-KV baseline. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

Scheduling precedence-constrained tasks under shared renewable resources is central to modern computing platforms. The Resource Investment Problem (RIP) models this setting by minimizing the cost of provisioned renewable resources under precedence and timing constraints. Exact mixed-integer programming and constraint programming become impractically slow on large instances, and dynamic updates require schedule revisions under tight latency budgets. We present iScheduler, a reinforcement-learning-driven iterative scheduling framework that formulates RIP solving as a Markov decision process over decomposed subproblems and constructs schedules through sequential process selection. The framework accelerates optimization and supports reconfiguration by reusing unchanged process schedules and rescheduling only affected processes. We also release L-RIPLIB, an industrial-scale benchmark derived from cloud-platform workloads with 1,000 instances of 2,500-10,000 tasks. Experiments show that iScheduler attains competitive resource costs while reducing time to feasibility by up to 43times against strong commercial baselines.

Inference of Large Language Models (LLMs) across computer clusters has become a focal point of research in recent times, with many acceleration techniques taking inspiration from CPU speculative execution. These techniques reduce bottlenecks associated with memory bandwidth, but also increase end-to-end latency per inference run, requiring high speculation acceptance rates to improve performance. Combined with a variable rate of acceptance across tasks, speculative inference techniques can result in reduced performance. Additionally, pipeline-parallel designs require many user requests to maintain maximum utilization. As a remedy, we propose PipeInfer, a pipelined speculative acceleration technique to reduce inter-token latency and improve system utilization for single-request scenarios while also improving tolerance to low speculation acceptance rates and low-bandwidth interconnects. PipeInfer exhibits up to a 2.15times improvement in generation speed over standard speculative inference. PipeInfer achieves its improvement through Continuous Asynchronous Speculation and Early Inference Cancellation, the former improving latency and generation speed by running single-token inference simultaneously with several speculative runs, while the latter improves speed and latency by skipping the computation of invalidated runs, even in the middle of inference.

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.

Two widely adopted techniques for LLM inference serving systems today are hybrid batching and disaggregated serving. A hybrid batch combines prefill and decode tokens of different requests in the same batch to improve resource utilization and throughput at the cost of increased latency per token. In contrast, disaggregated serving decouples compute-bound prefill and bandwidth-bound decode phases to optimize for service level objectives (SLOs) at the cost of resource under-utilization and KV-cache transfer overheads. To address the limitations of these techniques, we propose RAPID-Serve: a technique to concurrently execute prefill and decode on the same GPU(s) to meet latency SLOs while maintaining high throughput and efficient resource utilization. Furthermore, we propose Adaptive Resource Management for runtime compute resource allocation, optionally leveraging CU masking (a fine-grained Compute Unit partitioning feature on AMD InstinctTM GPUs). RAPID-Serve provides up to 4.1x (average 1.7x) unconstrained throughput improvement and 32x and higher (average 4.9x) throughput improvement under SLO constraints, showing it as an effective strategy compared to the state-of-the-art approaches, particularly in resource-constrained environments.

Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Recent studies on loss scaling have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops of computation. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer dedicated to open science. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of 38.38%, 36.14%, and 31.96%, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved 100% weak scaling efficiency on 1024 and 3072 MI250X GPUs, respectively. We also achieved strong scaling efficiencies of 89% and 87% for these two models.

This paper presents a detailed analysis of the scalability and parallelization of local search algorithms for the Satisfiability problem. We propose a framework to estimate the parallel performance of a given algorithm by analyzing the runtime behavior of its sequential version. Indeed, by approximating the runtime distribution of the sequential process with statistical methods, the runtime behavior of the parallel process can be predicted by a model based on order statistics. We apply this approach to study the parallel performance of two SAT local search solvers, namely Sparrow and CCASAT, and compare the predicted performances to the results of an actual experimentation on parallel hardware up to 384 cores. We show that the model is accurate and predicts performance close to the empirical data. Moreover, as we study different types of instances (random and crafted), we observe that the local search solvers exhibit different behaviors and that their runtime distributions can be approximated by two types of distributions: exponential (shifted and non-shifted) and lognormal.

GPU collective communication is typically optimized for bandwidth, yet many emerging workloads are increasingly limited by latency. Long-context decode-heavy large language model (LLM) inference is a prime example, where serving large models requires multiple GPUs, and many small collectives lie directly on the critical path of token generation. Therefore, even microsecond of overhead can impact performance and cost. In this work, we study how to approach the hardware Speed-of-Light (SoL) lower bound for GPU collectives within a scale-up network. We identify key principles for near-optimal designs, including barrier-free synchronization and efficient use of symmetric memory and multicast. Building on NCCL's device-side API, we develop low-latency interfaces for constructing custom collective kernels and use them to implement new symmetric collectives in NCCL. Microbenchmarks show substantial latency reductions for small and medium messages, reducing overhead to within 7% of the absolute SoL lower bound. When integrated into real applications, these kernels improve inter-token latency and throughput in LLM inference and accelerate cuSOLVERMp, demonstrating benefits for both AI inference and traditional HPC workloads.

Efficient distributed training serves as a powerful catalyst and an essential foundation for the development of large-scale neural networks. In distributed training scenarios, various pipeline parallelism methods are cleverly designed and widely employed. In this paper, we propose ZeroPP, a highly efficient and flexible pipeline parallelism method that trades off pipeline bubbles, memory usage, and communication through adaptive scheduling units. ZeroPP achieves minimal pipeline bubbles by carefully staggering the computation tasks of forward, input gradient, and weight gradient within a scheduling unit. Additionally, ZeroPP optimizes the combination of pipeline parallelism and fully sharded data parallelism using a blockwise schedule. We conduct experiments with popular GPT-style models and observe up to a 30% increase in throughput compared to the state-of-the-art breath-first pipeline parallelism. Besides, our evaluation also demonstrates up to a 68% increase in throughput and a 10% reduction in memory consumption compared to the memory-efficient 1F1B method.

Scaling test-time compute has emerged as an effective strategy for improving the performance of large language models. However, existing methods typically allocate compute uniformly across all queries, overlooking variation in query difficulty. To address this inefficiency, we formulate test-time compute allocation as a novel bandit learning problem and propose adaptive algorithms that estimate query difficulty on the fly and allocate compute accordingly. Compared to uniform allocation, our algorithms allocate more compute to challenging queries while maintaining accuracy on easier ones. Among challenging queries, our algorithms further learn to prioritize solvable instances, effectively reducing excessive computing on unsolvable queries. We theoretically prove that our algorithms achieve better compute efficiency than uniform allocation and empirically validate their effectiveness on math and code benchmarks. Specifically, our algorithms achieve up to an 11.10% performance improvement (15.04% relative) on the MATH-500 dataset and up to a 7.41% performance improvement (14.40% relative) on LiveCodeBench.

We study an online caching problem in which requests can be served by a local cache to avoid retrieval costs from a remote server. The cache can update its state after a batch of requests and store an arbitrarily small fraction of each file. We study no-regret algorithms based on Online Mirror Descent (OMD) strategies. We show that bounds for the regret crucially depend on the diversity of the request process, provided by the diversity ratio R/h, where R is the size of the batch, and h is the maximum multiplicity of a request in a given batch. We characterize the optimality of OMD caching policies w.r.t. regret under different diversity regimes. We also prove that, when the cache must store the entire file, rather than a fraction, OMD strategies can be coupled with a randomized rounding scheme that preserves regret guarantees, even when update costs cannot be neglected. We provide a formal characterization of the rounding problem through optimal transport theory, and moreover we propose a computationally efficient randomized rounding scheme.

Serverless computing has emerged as a compelling new paradigm of cloud computing models in recent years. It promises the user services at large scale and low cost while eliminating the need for infrastructure management. On cloud provider side, flexible resource management is required to meet fluctuating demand. It can be enabled through automated provisioning and deprovisioning of resources. A common approach among both commercial and open source serverless computing platforms is workload-based auto-scaling, where a designated algorithm scales instances according to the number of incoming requests. In the recently evolving serverless framework Knative a request-based policy is proposed, where the algorithm scales resources by a configured maximum number of requests that can be processed in parallel per instance, the so-called concurrency. As we show in a baseline experiment, this predefined concurrency level can strongly influence the performance of a serverless application. However, identifying the concurrency configuration that yields the highest possible quality of service is a challenging task due to various factors, e.g. varying workload and complex infrastructure characteristics, influencing throughput and latency. While there has been considerable research into intelligent techniques for optimizing auto-scaling for virtual machine provisioning, this topic has not yet been discussed in the area of serverless computing. For this reason, we investigate the applicability of a reinforcement learning approach, which has been proven on dynamic virtual machine provisioning, to request-based auto-scaling in a serverless framework. Our results show that within a limited number of iterations our proposed model learns an effective scaling policy per workload, improving the performance compared to the default auto-scaling configuration.

Machine learning applications are increasingly deployed not only to serve predictions using static models, but also as tightly-integrated components of feedback loops involving dynamic, real-time decision making. These applications pose a new set of requirements, none of which are difficult to achieve in isolation, but the combination of which creates a challenge for existing distributed execution frameworks: computation with millisecond latency at high throughput, adaptive construction of arbitrary task graphs, and execution of heterogeneous kernels over diverse sets of resources. We assert that a new distributed execution framework is needed for such ML applications and propose a candidate approach with a proof-of-concept architecture that achieves a 63x performance improvement over a state-of-the-art execution framework for a representative application.

Current heterogeneity-aware training systems for large models balance workloads by assigning unequal numbers of layers across pipeline stages, a strategy aimed at equalizing execution time through a unique pipeline configuration. However, this approach fails to address critical memory constraints and suffers from inherently coarse balancing granularity. In this paper, we propose Heterogeneous Pipeline Parallelism (HPP), a novel approach enabling efficient and fine-grained data balance across heterogeneous pipelines. To address HPP's key challenges, prohibitive search overhead for optimal configurations and high communication overhead from asymmetric All-Reduce, we design the ETH training system. ETH employs a hierarchical workload balance framework that jointly optimizes balance across heterogeneous pipelines and within individual pipeline stages. To mitigate asymmetric communication costs, ETH implements block-wise gradient aggregation. Furthermore, it integrates specialized optimizations that dramatically reduce HPP configuration search overhead. Evaluations demonstrate ETH achieves up to 2x training speedup while reducing configuration generation overhead by up to 56x compared to state-of-the-art systems. The code is available at https://github.com/YuMJie/ETH.

Agentic systems solve complex tasks by coordinating multiple agents that iteratively reason, invoke tools, and exchange intermediate results. To improve robustness and solution quality, recent approaches deploy multiple agent teams running in parallel to explore diverse reasoning trajectories. However, parallel execution comes at a significant computational cost: when different teams independently reason about similar sub-problems or execute analogous steps, they repeatedly perform substantial overlapping computation. To address these limitations, in this paper, we propose Learning to Share (LTS), a learned shared-memory mechanism for parallel agentic frameworks that enables selective cross-team information reuse while controlling context growth. LTS introduces a global memory bank accessible to all teams and a lightweight controller that decides whether intermediate agent steps should be added to memory or not. The controller is trained using stepwise reinforcement learning with usage-aware credit assignment, allowing it to identify information that is globally useful across parallel executions. Experiments on the AssistantBench and GAIA benchmarks show that LTS significantly reduces overall runtime while matching or improving task performance compared to memory-free parallel baselines, demonstrating that learned memory admission is an effective strategy for improving the efficiency of parallel agentic systems. Project page: https://joefioresi718.github.io/LTS_webpage/

Standard transport protocols like TCP operate as a blind, FIFO conveyor belt for data, a model that is increasingly suboptimal for latency-sensitive and interactive applications. This paper challenges this model by introducing CATS (Conductor-driven Asymmetric Transport Scheme), a framework that provides TCP with the semantic awareness necessary to prioritize critical content. By centralizing scheduling intelligence in a transport-native "Conductor", CATS significantly improves user-perceived performance by delivering essential data first. This architecture directly confronts a cascade of historical performance workarounds and their limitations, including the high overhead of parallel connections in HTTP/1.1, the transport-layer Head-of-Line blocking in HTTP/2, and the observed implementation heterogeneity of prioritization in HTTP/3 over QUIC. Built upon TCP BBR, our ns-3 implementation demonstrates this principle by reducing the First Contentful Paint by over 78% in a representative webpage download configured as a deliberate worst-case scenario, with no penalty to total page load time compared to the baseline.

This paper presents Block, a distributed scheduling framework designed to optimize load balancing and auto-provisioning across instances in large language model serving frameworks by leveraging contextual information from incoming requests. Unlike popular model serving systems that rely on monolithic and heuristic task schedulers, Block operates as a fully distributed, stateless, and predictive scheduling system to achieve low overhead, reliability, and scalability. It leverages the deterministic and predictable characteristics of LLM inferences, such as host configurations, response lengths, and hardware performance, to make scheduling decisions based on accurately predicted metrics. Evaluation on a 12 GPUs cluster shows that Block significantly outperforms heuristic schedulers, boosting serving capacity by up to 16.7\% and reducing P99 tail latency by up to 49.5\%. These performance gains remain consistent across diverse models, workloads and configurations. Code and data are open-sourced.

Synthetic data has become increasingly important for training large language models, especially when real data is scarce, expensive, or privacy-sensitive. Many such generation tasks require coordinated multi-agent workflows, where specialized agents collaborate to produce data that is higher quality, more diverse, and structurally richer. However, existing frameworks for multi-agent synthesis often depend on a centralized orchestrator, creating scalability bottlenecks, or are hardcoded for specific domains, limiting flexibility. We present Matrix, a decentralized framework that represents both control and data flow as serialized messages passed through distributed queues. This peer-to-peer design eliminates the central orchestrator. Each task progresses independently through lightweight agents, while compute-intensive operations, such as LLM inference or containerized environments, are handled by distributed services. Built on Ray, Matrix scales to tens of thousands of concurrent agentic workflows and provides a modular, configurable design that enables easy adaptation to a wide range of data generation workflows. We evaluate Matrix across diverse synthesis scenarios, such as multi-agent collaborative dialogue, web-based reasoning data extraction, and tool-use trajectory generation in customer service environments. In all cases, Matrix achieves 2--15times higher data generation throughput under identical hardware resources, without compromising output quality.

In LLM serving, reusing the KV cache of prompts across requests is critical for reducing TTFT and serving costs. Cache-affinity scheduling, which co-locates requests with the same prompt prefix to maximize KV cache reuse, often conflicts with load-balancing scheduling that distributes requests evenly across compute instances. Existing schedulers fail to reconcile this trade-off as they operate within a single mapping space, typically applying cache-affinity routing to a subset of requests and load-balanced routing to the rest, without a unified solution to achieve both goals. To address this limitation, we propose DualMap, a dual-mapping scheduling strategy for distributed LLM serving that achieves both cache affinity and load balancing. Its key idea is to map each request to two candidate instances via two independent hash functions based on the request prompt, then intelligently select the better candidate based on current system states. This design increases the likelihood that requests with shared prefixes are co-located, while evenly dispersing distinct prefixes across the cluster via ``the power of two choices''. To make DualMap robust under dynamic and skewed real-world workloads, we incorporate three techniques: 1) SLO-aware request routing, which prioritizes cache affinity but switches to load-aware scheduling when TTFT exceeds the SLO, enhancing load balance without sacrificing cache reuse; 2) hotspot-aware rebalancing, which dynamically migrates requests from overloaded to underloaded instances, mitigating hotspots and rebalancing the system; 3) lightweight dual-hash-ring scaling, which leverages a dual-hash-ring mapping to support fast and low-overhead instance scaling without costly global remapping. Experiments on real-world workloads show that DualMap improves effective request capacity by up to 2.25times under the same TTFT SLO constraints compared with SOTA work.

Multi-turn LLM agents interleave model calls with external tool invocations, shifting serving from stateless request processing to stateful program execution. Serving these workloads requires scheduling, KV-cache management, and routing policies that use program-level context, including turn dependencies, tool-induced gaps, and reusable KV state. Evaluating such policies directly on real systems is costly, since each design point may require dedicated accelerator time across arrival rates, model scales, serving-instance counts, and memory hierarchies. Simulation offers a scalable alternative, but existing LLM serving simulators target stateless request-level workloads and therefore omit the core dynamics of agent serving: multi-turn program execution, cross-turn cache locality, and KV-cache residency during tool gaps. We present AGENTSERVESIM, a hardware-aware simulator for multi-turn LLM agent serving. AGENTSERVESIM evaluates serving policies at program granularity through composable modules: a Program Orchestrator preserves program identity and turn order, a Tool Simulator materializes tool-induced gaps, a Session-Aware Router maintains program-to-instance affinity for cache-aware dispatch, and a KV Residency Model tracks policy-defined KV placement across HBM, host DRAM/CXL, and eviction. Across real serving deployments and hardware configurations, AGENTSERVESIM reproduces real-system behavior within 6% error across key performance metrics while running entirely on commodity CPUs. These results show that AGENTSERVESIM enables controlled, repeatable exploration of agent-serving policies without requiring exhaustive deployment on costly accelerators.

Large Mixture-of-Experts (MoE) model inference is challenging due to high resource demands and dynamic workloads. Existing solutions often deploy the entire model as a single monolithic unit, which applies a unified resource configuration to both attention and expert modules despite their different requirements, leading to limited scalability and resource inefficiency. In this paper, we propose Janus, a scalable MoE inference system that disaggregates attention and experts on separate GPU sub-clusters, enabling each module to be managed and scaled independently. Janus incorporates three key designs for efficient, disaggregated MoE inference. First, it proposes an adaptive two-phase communication scheme that exploits intra- and inter-node bandwidth hierarchies for low-latency data exchange. Second, motivated by the memory-bound nature of MoE modules, Janus introduces a lightweight scheduler and implements it as a GPU kernel to balance the number of activated experts across GPUs at minimal overhead, thereby reducing inference latency. Third, Janus performs fine-grained resource management to dynamically adjust expert placement and independently scale attention and MoE resources to improve overall efficiency. Evaluation shows Janus achieves up to 3.9 higher perGPU throughput than state-of-the-art systems while meeting per-token latency requirements.

The transition from sequential to parallel computing is essential for modern high-performance applications but is hindered by the steep learning curve of concurrent programming. This challenge is magnified for irregular data structures (such as sparse graphs, unbalanced trees, and non-uniform meshes) where static scheduling fails and data dependencies are unpredictable. Current Large Language Models (LLMs) often fail catastrophically on these tasks, generating code plagued by subtle race conditions, deadlocks, and sub-optimal scaling. We bridge this gap with ParEVO, a framework designed to synthesize high-performance parallel algorithms for irregular data. Our contributions include: (1) The Parlay-Instruct Corpus, a curated dataset of 13,820 tasks synthesized via a "Critic-Refine" pipeline that explicitly filters for empirically performant algorithms that effectively utilize Work-Span parallel primitives; (2) specialized DeepSeek, Qwen, and Gemini models fine-tuned to align probabilistic generation with the rigorous semantics of the ParlayLib library; and (3) an Evolutionary Coding Agent (ECA) that improves the "last mile" of correctness by iteratively repairing code using feedback from compilers, dynamic race detectors, and performance profilers. On the ParEval benchmark, ParEVO achieves an average 106x speedup (with a maximum of 1103x) across the suite, and a robust 13.6x speedup specifically on complex irregular graph problems, outperforming state-of-the-art commercial models. Furthermore, our evolutionary approach matches state-of-the-art expert human baselines, achieving up to a 4.1x speedup on specific highly-irregular kernels. Source code and datasets are available at https://github.com/WildAlg/ParEVO.

Mixture-of-Experts (MoE) models challenge serving infrastructures with dynamic, sparse expert utilization, causing instability on conventional systems designed for dense architectures. We propose EaaS, a novel serving system to enable efficient, scalable, and robust MoE deployment. Our system disaggregates MoE modules into independent, stateless services. This design enables fine-grained resource scaling and provides inherent fault tolerance by decoupling compute units. The architecture is powered by a high-performance, CPU-free peer-to-peer communication library that ensures minimal overhead and high throughput. Experiments confirm EaaS's scalability and efficiency, achieving performance comparable to monolithic systems while providing robust fault tolerance and strong scalability. EaaS incurs less than a 2% throughput reduction under simulated hardware failures that would otherwise halt monolithic architectures. It further saves up to 37.5% of computing resources through dynamic fine-grained adaptation to serving traffic, demonstrating strong resilience for large-scale MoE deployment in production.

As LLMs scale to multi-million-token KV histories, real-time autoregressive decoding under tight Token-to-Token Latency (TTL) constraints faces growing pressure. Two core bottlenecks dominate: accessing Feed-Forward Network (FFN) weights and reading long KV caches. While Tensor Parallelism (TP) helps mitigate the cost of FFN weight reads, it does not scale well for attention. When TP width exceeds the number of KV heads, it leads to inefficient KV duplication, limits parallelism, and constrains batch size. Simultaneously, DRAM reads for long KV histories scale linearly with batch size, further capping efficiency. We introduce Helix Parallelism, a hybrid execution strategy that applies KV parallelism during attention to shard KV caches across GPUs, then reuses the same GPUs for TP in dense LLMs or TPxExpert Parallel (EP) in MoEs during FFN computation. To preserve exact attention behavior, Helix includes a lightweight communication step. To minimize the exposed communication cost, we introduce Helix HOP-B. Helix HOP-B effectively minimizes communication overhead through batchwise overlap, preserving low TTL while improving GPU efficiency. Compared to conventional parallelism approaches, Helix reduces TTL by up to 1.5x at fixed batch sizes and supports up to 32x larger batches under the same latency budget for DeepSeek-R1, pushing forward the throughput-latency Pareto on Blackwell and making real-time inference with ultra-long-sequence practical.

Fine-tuning large language models (LLMs) with parameter-efficient techniques such as LoRA and QLoRA has enabled adaptation of foundation models on modest hardware. Yet the efficiency of such training on consumer-grade GPUs, especially under strict 8 GB VRAM limits, remains underexplored. We present a controlled profiling study of LoRA/QLoRA fine-tuning using the Qwen2.5-1.5B-Instruct model on a single NVIDIA RTX 4060. Across three representative configurations, we systematically vary batch size, sequence length, optimizer choice (AdamW vs. PagedAdamW), and precision (fp16 vs. bf16). We report throughput (tokens/s), time per 10k tokens, and VRAM footprint, alongside energy estimates derived from GPU board power limits. Our results show that paged optimizers improve throughput by up to 25% (628 tok/s vs. 500 tok/s baseline), while bf16 degrades efficiency relative to fp16. Despite 8 GB constraints, sequence lengths up to 2048 tokens were feasible using parameter-efficient strategies. To our knowledge, this is the first systematic case study of LLM fine- tuning efficiency on consumer GPUs, providing reproducible benchmarks and practical guidelines for resource-constrained researchers and practitioners.

KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls practitioners should be aware of when deploying KV cache compressed LLMs. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example of that, we highlight system prompt leakage as a case study, empirically showing the impact of compression on leakage and general instruction following. We show several factors that play a role in prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks.

The sparse Mixture-of-Experts (MoE) architecture is increasingly favored for scaling Large Language Models (LLMs) efficiently, but it depends on heterogeneous compute and memory resources. These factors jointly affect system Cost, Accuracy, and Performance (CAP), making trade-offs inevitable. Existing benchmarks often fail to capture these trade-offs accurately, complicating practical deployment decisions. To address this, we introduce MoE-CAP, a benchmark specifically designed for MoE systems. Our analysis reveals that achieving an optimal balance across CAP is difficult with current hardware; MoE systems typically optimize two of the three dimensions at the expense of the third-a dynamic we term the MoE-CAP trade-off. To visualize this, we propose the CAP Radar Diagram. We further introduce sparsity-aware performance metrics-Sparse Memory Bandwidth Utilization (S-MBU) and Sparse Model FLOPS Utilization (S-MFU)-to enable accurate performance benchmarking of MoE systems across diverse hardware platforms and deployment scenarios.

Serving systems for Large Language Models (LLMs) are often optimized to improve quality of service (QoS) and throughput. However, due to the lack of open-source LLM serving workloads, these systems are frequently evaluated under unrealistic workload assumptions. Consequently, performance may degrade when systems are deployed in real-world scenarios. This work presents BurstGPT, an LLM serving workload with 10.31 million traces from regional Azure OpenAI GPT services over 213 days. BurstGPT captures LLM serving characteristics from user, model and system perspectives: (1) User request concurrency: burstiness variations of requests in Azure OpenAI GPT services, revealing diversified concurrency patterns in different services and model types. (2) User conversation patterns: counts and intervals within conversations for service optimizations. (3) Model response lengths: auto-regressive serving processes of GPT models, showing statistical relations between requests and their responses. (4) System response failures: failures of conversation and API services, showing intensive resource needs and limited availability of LLM services in Azure. The details of the characteristics can serve multiple purposes in LLM serving optimizations, such as system evaluation and trace provisioning. In our demo evaluation with BurstGPT, frequent variations in BurstGPT reveal declines in efficiency, stability, or reliability in realistic LLM serving. We identify that the generalization of KV cache management, scheduling and disaggregation optimizations can be improved under realistic workload evaluations. BurstGPT is publicly available now at https://github.com/HPMLL/BurstGPT and is widely used to develop prototypes of LLM serving frameworks in the industry.

Heterogeneous LLM serving stacks split scheduling into two layers that optimize in isolation: model routers pick a model from quality and cost signals while ignoring instance load, and serving load balancers optimize queues while ignoring quality. We present RouteBalance, a serving-aware scheduling layer that fuses both into a single online assignment over concrete model instances, jointly trading off quality, latency, and cost. A batched in-process predictor stack and dead-reckoned instance state keep the joint decision cheap on the request hot path (approx32 ms at 12 req/s). On a 13-instance, 28-GPU heterogeneous cluster serving four model sizes, a single deployed RouteBalance stack traces the upper region of the three-way quality-cost-throughput frontier. Sweeping one weight vector reaches both the highest routing-decision quality (DeepEval 0.419, +0.013 over the strongest baseline, 95% CI [{+}0.005,{+}0.022]; the ordering holds when a second judge re-scores the actually served text) and, at its cost-priority corner, per-request cost that ties the cheapest baseline. With router engineering equalized against concurrent-scoring baseline variants we build, its balanced preset serves at 2.8 s and 30 req/s, leading 2.6 to 4.1times ahead of enhanced BEST-Route at high load. (Deploying those routers as published, one serial scoring call per request, makes them collapse 23times under load, a deployment-architecture effect we isolate separately, not the routing result.) A four-arm isolation shows the benefit follows from pricing latency at model-selection time; the learned predictors contribute calibration and SLO headroom rather than the headline frontier. Code: https://github.com/AKafakA/route-balance

Pipeline parallelism is one of the key components for large-scale distributed training, yet its efficiency suffers from pipeline bubbles which were deemed inevitable. In this work, we introduce a scheduling strategy that, to our knowledge, is the first to successfully achieve zero pipeline bubbles under synchronous training semantics. The key idea behind this improvement is to split the backward computation into two parts, one that computes gradient for the input and another that computes for the parameters. Based on this idea, we handcraft novel pipeline schedules that significantly outperform the baseline methods. We further develop an algorithm that automatically finds an optimal schedule based on specific model configuration and memory limit. Additionally, to truly achieve zero bubble, we introduce a novel technique to bypass synchronizations during the optimizer step. Experimental evaluations show that our method outperforms the 1F1B schedule up to 23% in throughput under a similar memory limit. This number can be further pushed to 31% when the memory constraint is relaxed. We believe our results mark a major step forward in harnessing the true potential of pipeline parallelism. We open sourced our implementation based on the popular Megatron-LM repository on https://github.com/sail-sg/zero-bubble-pipeline-parallelism.

In distributed computing, slower nodes (stragglers) usually become a bottleneck. Gradient Coding (GC), introduced by Tandon et al., is an efficient technique that uses principles of error-correcting codes to distribute gradient computation in the presence of stragglers. In this paper, we consider the distributed computation of a sequence of gradients {g(1),g(2),ldots,g(J)}, where processing of each gradient g(t) starts in round-t and finishes by round-(t+T). Here Tgeq 0 denotes a delay parameter. For the GC scheme, coding is only across computing nodes and this results in a solution where T=0. On the other hand, having T>0 allows for designing schemes which exploit the temporal dimension as well. In this work, we propose two schemes that demonstrate improved performance compared to GC. Our first scheme combines GC with selective repetition of previously unfinished tasks and achieves improved straggler mitigation. In our second scheme, which constitutes our main contribution, we apply GC to a subset of the tasks and repetition for the remainder of the tasks. We then multiplex these two classes of tasks across workers and rounds in an adaptive manner, based on past straggler patterns. Using theoretical analysis, we demonstrate that our second scheme achieves significant reduction in the computational load. In our experiments, we study a practical setting of concurrently training multiple neural networks over an AWS Lambda cluster involving 256 worker nodes, where our framework naturally applies. We demonstrate that the latter scheme can yield a 16\% improvement in runtime over the baseline GC scheme, in the presence of naturally occurring, non-simulated stragglers.

Scaling up Large Language Model(LLM) training involves fitting a tremendous amount of training parameters across a limited number of workers. However, methods like ZeRO-3 that drastically reduce GPU memory pressure often incur heavy communication to ensure global synchronization and consistency. Established efforts such as ZeRO++ use secondary partitions to avoid inter-node communications, given that intra-node GPU-GPU transfer generally has more bandwidth and lower latency than inter-node connections. However, as more capable infrastructure like Frontier, equipped with AMD GPUs, emerged with impressive computing capability, there is a need for investigations on the hardware topology and to develop targeted strategies to improve training efficiency. In this work, we propose a collection of communication and optimization strategies for ZeRO++ to reduce communication costs and improve memory utilization. In this paper, we propose a 3-level hierarchical partitioning specifically for the current Top-1 supercomputing cluster, Frontier, which aims at leveraging various bandwidths across layers of communications (GCD-GCD, GPU-GPU, and inter-node) to reduce communication overhead. For a 20B GPT model, we observe a 1.71x increase in TFLOPS per GPU when compared with ZeRO++ up to 384 GCDs and a scaling efficiency of 0.94 for up to 384 GCDs. To the best of our knowledge, our work is also the first effort to efficiently optimize LLM workloads on Frontier AMD GPUs.

Large Language Model (LLM) workloads have distinct prefill and decode phases with different compute and memory requirements which should ideally be accounted for when scheduling input queries across different LLM instances in a cluster. However existing scheduling algorithms treat LLM workloads as monolithic jobs without considering the distinct characteristics of the two phases in each workload. This leads to sub-optimal scheduling and increased response latency. In this work, we propose a heuristic-guided reinforcement learning-based intelligent router for data-driven and workload-aware scheduling. Our router leverages a trainable response-length predictor, and a novel formulation for estimating the impact of mixing different workloads to schedule queries across LLM instances and achieve over 11\% lower end-to-end latency than existing approaches.

Large language model (LLM) inference workload dominates a wide variety of modern AI applications, ranging from multi-turn conversation to document analysis. Balancing fairness and efficiency is critical for managing diverse client workloads with varying prefix patterns. Unfortunately, existing fair scheduling algorithms for LLM serving, such as Virtual Token Counter (VTC), fail to take prefix locality into consideration and thus suffer from poor performance. On the other hand, locality-aware scheduling algorithms in existing LLM serving frameworks tend to maximize the prefix cache hit rate without considering fair sharing among clients. This paper introduces the first locality-aware fair scheduling algorithm, Deficit Longest Prefix Match (DLPM), which can maintain a high degree of prefix locality with a fairness guarantee. We also introduce a novel algorithm, Double Deficit LPM (D^2LPM), extending DLPM for the distributed setup that can find a balance point among fairness, locality, and load-balancing. Our extensive evaluation demonstrates the superior performance of DLPM and D^2LPM in ensuring fairness while maintaining high throughput (up to 2.87times higher than VTC) and low per-client (up to 7.18times lower than state-of-the-art distributed LLM serving system) latency.

Large language model (LLM) serving infrastructures are undergoing a shift toward heterogeneity and disaggregation. Modern deployments increasingly integrate diverse accelerators and near-memory processing technologies, introducing significant hardware heterogeneity, while system software increasingly separates computation, memory, and model components across distributed resources to improve scalability and efficiency. As a result, LLM serving performance is no longer determined by hardware or software choices in isolation, but by their runtime interaction through scheduling, data movement, and interconnect behavior. However, understanding these interactions remains challenging, as existing simulators lack the ability to jointly model heterogeneous hardware and disaggregated serving techniques within a unified, runtime-driven framework. This paper presents LLMServingSim 2.0, a unified system-level simulator designed to make runtime-driven hardware-software interactions in heterogeneous and disaggregated LLM serving infrastructures explicit and analyzable. LLMServingSim 2.0 embeds serving decisions and hardware behavior into a single runtime loop, enabling interaction-aware modeling of batching, routing, offloading, memory, and power. The simulator supports extensible integration of emerging accelerators and memory systems through profile-based modeling, while capturing dynamic serving behavior and system-level effects. We validate LLMServingSim 2.0 against real deployments, showing that it reproduces key performance, memory, and power metrics with an average error of 0.95%, while maintaining simulation times of around 10 minutes even for complex configurations. These results demonstrate that LLMServingSim 2.0 provides a practical bridge between hardware innovation and serving-system design, enabling systematic exploration and co-design for next-generation LLM serving infrastructures.

Hardware prefetching is critical to fill the performance gap between CPU speeds and slower memory accesses. With multicore architectures becoming commonplace, traditional prefetchers are severely challenged. Independent core operation creates significant redundancy (up to 20% of prefetch requests are duplicates), causing unnecessary memory bus traffic and wasted bandwidth. Furthermore, cutting-edge prefetchers such as Pythia suffer from about a 10% performance loss when scaling from a single-core to a four-core system. To solve these problems, we propose CRL-Pythia, a coordinated reinforcement learning based prefetcher specifically designed for multicore systems. In this work, CRL-Pythia addresses these issues by enabling cross-core sharing of information and cooperative prefetching decisions, which greatly reduces redundant prefetch requests and improves learning convergence across cores. Our experiments demonstrate that CRL-Pythia outperforms single Pythia configurations in all cases, with approximately 12% IPC (instructions per cycle) improvement for bandwidth-constrained workloads, while imposing moderate hardware overhead. Our sensitivity analyses also verify its robustness and scalability, thereby making CRL-Pythia a practical and efficient solution to contemporary multicore systems.

We present JigsawRL, a cost-efficient framework that explores Pipeline Multiplexing as a new dimension of RL parallelism. JigsawRL decomposes each pipeline into a Sub-Stage Graph that exposes the intra-stage and inter-worker imbalance hidden by stage-level systems. On this abstraction, JigsawRL resolves multiplexing interference through dynamic resource allocation, eliminates fragmented utilization by migrating long-tail rollouts across workers, and formulates their coordination as a graph scheduling problem solved with a look-ahead heuristic. On 4-64 H100/A100 GPUs across different agentic RL pipelines and models, JigsawRL achieves up to 1.85x throughput over Verl on synchronous RL, 1.54x over StreamRL and AReaL on asynchronous RL, and supports heterogeneous pipelines with moderate latency trade-off.

Host-resident layer streaming enables full-parameter LLM fine-tuning beyond GPU memory, but data-parallel ranks compete for shared host resources. Replicated transfers amplify traffic, while strong scaling can expose host work as computation windows shrink. We present SlideDP, a synchronous data-parallel runtime for shared-host multi-GPU systems. It maintains one authoritative host state, decouples communication routes from state layout, and pipelines parameter delivery, gradient aggregation, and CPU updates across ranks and chunks. An analytical step-time model characterizes resource bottlenecks and pipeline exposure; runtime measurements guide communication, chunking, and activation policies under a GPU memory budget. In matched-batch sweeps, SlideDP achieves geometric-mean throughput ratios of 1.46-2.64times over SlideFormer, MegaTrain, and ZeRO-Offload. On four H100s, SlideDP approaches GPU-resident FSDP2 throughput for Qwen3-14B at a smaller batch size. With a larger batch, it processes over 1M tokens per step and exceeds FSDP2's measured peak throughput by 11.2%. Separately, it supports 256K-token sequences for the same model and fine-tunes Qwen2.5-72B on four RTX 4090 GPUs. Project page: https://github.com/RegiaYoung/SlideDP.

Efficient scheduling of LLM inference tasks is essential for achieving low latency and high throughput, particularly with the growing use of reasoning-capable LLMs. Traditional strategies like First-Come-First-Serve (FCFS) often suffer from Head-of-Line (HOL) blocking, where long-running tasks delay shorter ones queued behind them. In this paper, we introduce PARS, a prompt-aware LLM task scheduler that improves serving efficiency by approximating shortest-job-first (SJF) scheduling through pairwise ranking with margin ranking loss. PARS focuses on impactful scheduling decisions and is seamlessly integrated into the state-of-the-art LLM serving system vLLM. It effectively predicts response-length-based task ordering, reducing latency with minimal overhead. Extensive experiments across multiple LLMs and real-world inference datasets show that PARS significantly improves performance, including for reasoning workloads. Furthermore, our cross-model evaluations demonstrate that the design generalizes well, enabling effective scheduling even when predictors are trained on different LLMs.

Deploying a large language model (LLM) inference service remains costly because centralized serving depends on specialized GPU clusters and high-bandwidth interconnects in datacenters. An appealing alternative is to leverage collaborative decentralized GPU pools. However, heterogeneity in GPU and limited interconnected network bandwidth, along with potentially dynamic availability, make efficient scheduling the central challenge in this scenario. In this paper, we present Parallax, a decentralized LLM serving system that turns a pool of heterogeneous GPUs into an efficient inference platform via a two-phase scheduler. Parallax decomposes planning into (i) model allocation, which places layers of each replica across diverse GPUs to jointly optimize latency and throughput under memory and link-bandwidth constraints, and (ii) request-time GPU pipeline selection, which stitches layers from different replicas into end-to-end execution chains that balance load and adapt to current conditions. We implement Parallax and evaluate it on open-source LLMs deployed over real volunteer nodes. Parallax consistently reduces latency and increases throughput relative to decentralized baselines, demonstrating that principled scheduling can make volunteer compute a practical, affordable substrate for LLM inference. Github Repo at: https://github.com/GradientHQ/parallax.

The growing prevalence of Large Language Models (LLMs) and their substantial memory requirements have prompted renewed interest in CPU offloading as a method to compensate for limited GPU memory. In particular, when CPU memory is leveraged to temporarily store intermediate states of LLMs, CPU memory becomes a new bottleneck and soon reaches the capacity limitation of commodity CPUs. In this work, we investigate the effectiveness of Compute Express Link (CXL) add-in card (AIC) memory as an extension to CPU memory, enabling larger model sizes and longer context lengths during fine-tuning. Through extensive benchmarking, this study quantifies the performance overhead introduced by transferring data between CXL memory, CPU, and GPUs, focusing on how concurrency and data volume influence bandwidth utilization and latency. This study also compares CPUbased optimizer steps when model parameters, gradients, and optimizer states reside in local memory versus CXL memory, revealing that naive adoption of CXL often degrades performance during the optimizer phase. To overcome these challenges, this study proposes a CXL-aware allocation to strategically partition CPU offloading workloads across both local and CXL memory. This study further demonstrates that employing multiple AICs significantly reduces bandwidth contention, thus improving scalability. Experimental results show that these optimizations enable efficient long-context LLM fine-tuning, underscoring CXL as a promising avenue for unlocking the full potential of CPU offloading in long-context LLM fine-tuning.

Sparse Mixture of Experts (MoE) models, while outperforming dense Large Language Models (LLMs) in terms of performance, face significant deployment challenges during inference due to their high memory demands. Existing offloading techniques, which involve swapping activated and idle experts between the GPU and CPU, often suffer from rigid expert caching mechanisms. These mechanisms fail to adapt to dynamic routing, leading to inefficient cache utilization, or incur prohibitive costs for prediction training. To tackle these inference-specific challenges, we introduce ExpertFlow, a comprehensive system specifically designed to enhance inference efficiency by accommodating flexible routing and enabling efficient expert scheduling between CPU and GPU. This reduces overhead and boosts system performance. Central to our approach is a predictive routing path-based offloading mechanism that utilizes a lightweight predictor to accurately forecast routing paths before computation begins. This proactive strategy allows for real-time error correction in expert caching, significantly increasing cache hit ratios and reducing the frequency of expert transfers, thereby minimizing I/O overhead. Additionally, we implement a dynamic token scheduling strategy that optimizes MoE inference by rearranging input tokens across different batches. This method not only reduces the number of activated experts per batch but also improves computational efficiency. Our extensive experiments demonstrate that ExpertFlow achieves up to 93.72\% GPU memory savings and enhances inference speed by 2 to 10 times compared to baseline methods, highlighting its effectiveness and utility as a robust solution for resource-constrained inference scenarios.

Modern data parallel (DP) training favors collective communication over parameter servers (PS) for its simplicity and efficiency under balanced workloads. However, the balanced workload assumption no longer holds in large language model (LLM) post-training due to the high variance in sequence lengths. Under imbalanced workloads, collective communication creates synchronization barriers, leading to under-utilization of devices with smaller workloads. This change in training dynamics calls for a revisit of the PS paradigm for its robustness to such imbalance. We propose On-Demand Communication (ODC), which adapts PS into Fully Sharded Data Parallel (FSDP) by replacing collective all-gather and reduce-scatter with direct point-to-point communication. Compared to FSDP, ODC reduces the synchronization barrier from once per layer to once per minibatch and decouples the workload on each device so that faster workers are not stalled. It also enables simpler and more effective load balancing at the minibatch level. Across diverse LLM post-training tasks, ODC consistently improves device utilization and training throughput, achieving up to a 36\% speedup over standard FSDP. These results demonstrate that ODC is a superior fit for the prevalent imbalanced workloads in LLM post-training. Our implementation of ODC and integration with FSDP is open-sourced at https://github.com/sail-sg/odc.

Large reasoning models (LRMs) have exhibited the capacity of enhancing reasoning performance via internal test-time scaling. Building upon this, a promising direction is to further scale test-time compute to unlock even greater reasoning capabilities. However, as we push these scaling boundaries, systematically understanding the practical limits and achieving optimal resource allocation becomes a critical challenge. In this paper, we investigate the scaling Pareto of test-time scaling and introduce the Test-Time Scaling Performance Model (TTSPM). We theoretically analyze two fundamental paradigms for such extended scaling, parallel scaling and sequential scaling, from a probabilistic modeling perspective. Our primary contribution is the derivation of the saturation point on the scaling budget for both strategies, identifying thresholds beyond which additional computation yields diminishing returns. Remarkably, despite their distinct mechanisms, both paradigms converge to a unified mathematical structure in their upper bounds. We empirically validate our theoretical findings on challenging reasoning benchmarks, including AIME, MATH-500, and GPQA, demonstrating the practical utility of these bounds for test-time resource allocation. We hope that this work provides insights into the cost-benefit trade-offs of test-time scaling, guiding the development of more resource-efficient inference strategies for large reasoning models.

Distributed optimization methods such as DiLoCo have been shown to be effective in training very large models across multiple distributed workers, such as datacenters. These methods split updates into two parts: an inner optimization phase, where the workers independently execute multiple optimization steps on their own local data, and an outer optimization step, where the inner updates are synchronized. While such approaches require orders of magnitude less communication than standard data-parallel training, in settings where the workers are datacenters, even the limited communication requirements of these approaches can still cause significant slow downs due to the blocking necessary at each outer optimization step. In this paper, we investigate techniques to mitigate this issue by overlapping communication with computation in a manner that allows the outer optimization step to fully overlap with the inner optimization phase. We show that a particular variant, dubbed eager updates, provides competitive performance with standard DiLoCo in settings with low bandwidth between workers.

Agentic LLM applications interleave LLM generation requests with tool calls. These tool calls break the continuity of the workflow by creating pauses between LLM requests, bringing many challenges for the serving system, especially under multi-turn scenarios. Each pause potentially causes KV cache eviction and extra waiting time before entering the continuous batch for the following LLM request. Since these pauses happen for each call, this problem becomes increasingly severe as turn number grow for agentic programs. Previous works either fail to incorporate information from the tool call, evicting KV cache that leads to repetitive prefill or loading, or ignore the continuity of a multi-turn program, creating waiting time between turns that increases per-request latency. We present Continuum, a serving system to optimize job completion time for multi-turn agent workloads by combining tool-aware KV cache timeout with program-level scheduling. By predicting tool call durations in agentic workflows, Continuum selectively pins the KV cache in GPU memory with a time-to-live value based on total turn number. When combined with program-level first-come-first-serve, Continuum prevents scheduling bubbles, preserves multi-turn continuity, and optimizes for throughput for complex agentic workflows. By modeling the variability of tool call and agent program continuity, Continuum outperforms state-of-the-art baselines. Our evaluation on real-world agentic workloads (SWE-Bench and BFCL) with Llama-3.1 8B/70B models shows that Continuum significantly improves the average job completion times, and remains performant across different hardware setups and DRAM offloading schemes. Preview code is available at: https://github.com/Hanchenli/vllm-continuum