Daily Papers

Full-duplex speech-to-speech (S2S) models provide natural, low-latency conversational interaction and would benefit from the ability to use external tools and complete voice-agent tasks. We propose a frontend-backend architecture where a duplex speech-to-text frontend learns to emit a delegation token and forwards streaming ASR transcripts to a text-based backend LLM for tool calls. Tool-call results from the backend are injected back into the frontend through a lightweight prefill-and-repeat mechanism and then synthesized using streaming TTS to the user. Our approach largely preserves regular duplex turn-taking, interruption handling, and low-latency interaction as it requires minimal modifications to the frontend model. In a single-turn tool-call evaluation, our system achieves 92-97% tool-call recall, competitive tool-call prediction performance, and 81.2% accuracy in rejecting irrelevant calls. When equipped with a larger backend (e.g., Qwen3-235B-A22B), our system achieves competitive results on Full-Duplex-Bench-V3 compared to open and closed source models, and significantly outperforms GPT-realtime-mini and Qwen3-Omni-30B-A3B-Instruct on EVA-Bench. These results demonstrate that backend delegation is an effective and modular approach for combining natural duplex speech interaction with strong agentic tool-call capabilities.

Chunked prefill has become a widely adopted serving strategy for long-context large language models, but efficient attention computation in this regime remains challenging. Existing sparse attention methods are primarily designed for one-shot prefill and do not translate efficiently to chunked prefill: block-sparse kernels lose efficiency when the query length is limited by the chunk size, while fine-grained pattern search becomes costly when repeated over the accumulated KV cache at every chunk. QUOKA, a recent method that directly targets chunked prefill, avoids sparse-kernel overhead but relies on query-subsampled, token-level KV selection, which can miss query-specific KV entries and introduce explicit KV-copy overhead. To address these limitations, we propose CompactAttention, a chunked-prefill attention mechanism based on Block-Union KV Selection. CompactAttention treats 2D block-sparse masks as KV-selection signals rather than direct sparse-kernel execution plans, and converts them into GQA-aware per-group KV block tables through Q-block union and intra-group union. This construction produces the minimal block tables that preserve all KV blocks selected by the input masks under paged execution constraints, enabling selected KV blocks to be accessed in place without explicit KV compaction. On LLaMA-3.1-8B-Instruct, CompactAttention maintains accuracy close to dense attention on the RULER benchmark while delivering up to 2.72times attention speedup at 128K context length under chunked prefill.

Long-context inference in decoder-only language models is costly because long prompts are processed during Prefill, cached at every layer, and repeatedly attended to during autoregressive Decode. We introduce Shallow Prefill, dEEp Decode (SPEED), a phase-asymmetric KV-visibility policy that materializes non-anchor prompt-token KV states only in lower layers while keeping Decode-phase tokens full-depth. Unlike previous approaches that make upper-layer prompt KV states cheaper to store or construct, SPEED removes prefill tokens from the upper-layer Decode visibility set altogether. With a minimal BoS anchor, this simple change preserves broad benchmark quality while reducing long-context cost. In a controlled Llama-3.1-8B instruction-tuning study, SPEED using only 75\% of layers for prefill tokens reaches 51.2 average score on OLMES-style benchmarks, compared with 51.4 for the full-depth baseline, while improving TTFT by 33\%, TPOT by 22\%, and reducing active KV memory by 25.0\% at 128K context. Layer-wise diagnostics suggest that this cutoff retains the main prompt-selection and representation-stabilization regions of the full-depth model. These results show that long-context prompt tokens need not always persist as full-depth KV-cache objects when Decode-phase tokens remain full-depth.

Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source's KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B->32B, one source layer explains 56% of variance in the target's keys and 32% in values, rising to 79% and 65% with multiple source layers. Building on this, we design a closed-form ridge mapper that operates per head and proceeds in three steps. First, for each target layer we select the top-k most predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each. Surprisingly, across six pairs in three families, this linear mapper retains 73-98% of the receiver's standalone-prefill accuracy on four pairs, while two degrade sharply. A nonlinear MLP recovers up to +37 pp HellaSwag retention on the failures. The mapper runs 2.7-25x faster than re-prefill and remains stable across multi-turn handoff, making cross-model KV cache transfer practical.

Large language models (LLMs) are widely used to tackle complex tasks with autonomous workflows. Recently, reusable natural language skills have emerged as a popular paradigm to inject procedural knowledge into LLM applications. Since popular skills are often invoked repeatedly, placing their full text in every context significantly increases prefill cost and latency. While text compression techniques have the potential to solve this problem, most existing methods are designed to compress factual knowledge in documents instead of procedural knowledge, making them insufficient for skill compression. In this paper, we argue that an effective skill compression method should: 1) preserve logical dependencies among workflows and tool protocols, 2) enable lightweight, offline compression for frequently updated community skills, and 3) be adaptable to varying complexities across skills. To address this, we present SKIM (SKIll coMpression), an adaptive multi-resolution soft token compression framework for procedural skills. Depending on the complexity of each skill, SKIM creates different numbers of soft tokens that not only improve the efficiency of LLM inference, but also preserve the effectiveness of skill usage. Experiments indicate that SKIM compresses skills to 30 to 60 percent of their original token length while preserving task performance better than existing compression methods.We have released our code at https://github.com/bebr2/SKIM .

Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and computation redundancy in long-horizon tasks. Existing KV cache sharing methods reduce this repeated prefill, but they either require additional training or architectural constraints or retain substantial model computation. Moreover, direct cache reuse causes the current agent to rely on cache states generated by the previous agent's adapter, weakening the role-specific behavior encoded by its own LoRA. We present PReCache, a training-free KV cache sharing framework with two designs, namely PreLRShared and ReBaseShared, that share the base cache computed using the pretrained weights and precompute a compact agent-specific low-rank (LR) cache. To remove repeated prefill, PreLRShared precomputes each agent's LR cache when the shared context is first processed, allowing the current agent to use its own LR cache without reprocessing context processed by previous agents. To improve sharing accuracy, ReBaseShared reconstructs the shared base cache from adapter-free hidden states, reducing the remaining error caused by the previous agent's adapted representation. To minimize its reconstruction cost, we propose two inference schemes tailored to single-stream inference and concurrent serving, performing the same reconstruction after each agent's turn or alongside its execution, respectively. Across multiple models and agent benchmarks, PreLRShared achieves up to a 3.1x TTFT speedup and a 2.3x improvement in per-request throughput over inference without KV cache sharing. ReBaseShared best preserves accuracy overall among the evaluated cache-sharing methods, with an average drop of only 1.1 points relative to inference without cache sharing.

Prompt length is a major bottleneck in agentic large language model (LLM) workloads, where repeated inference steps and multi-call loops incur substantial prefill cost. Recent work on speculative prefill demonstrates that attention-based token importance estimation can enable training-free prompt compression, but this assumes the existence of a draft model that shares the same tokenizer as the target model. In practice, however, agentic pipelines frequently employ models without any smaller in-family draft model. In this work, we study cross-family speculative prefill, where a lightweight draft model from one model family is used to perform prompt compression for a target model from a different family. Using the same speculative prefill mechanism as prior work, we evaluate a range of cross-family draft-target combinations, including Qwen, LLaMA, and DeepSeek models. Across a broad diversity of tasks, we find that attention-based token importance estimation transfers reliably across different model families despite differences in model architectures and tokenizers between draft and target models. Cross-model prompt compression largely retains 90~100% of full-prompt baseline performance and, in some cases, slightly improves accuracy due to denoising effects, while delivering substantial reductions in time to first token (TTFT). These results suggest that speculative prefill depends mainly on task priors and semantic structure, thus serving as a generalizable prompt compression primitive. We discuss the implications of our findings for agentic systems, where repeated long-context inference and heterogeneous model stacks make cross-model prompt compression both necessary and practical.

Streaming video understanding is a critical capability for real-world applications, including embodied intelligence, autonomous driving, industrial monitoring, surveillance and early warning, and wearable assistants. However, processing continuous video streams with multimodal large language models (MLLMs) is computationally expensive. Existing efforts have explored reducing streaming overhead through visual token pruning, token merging, quantization, on-demand frame retrieval, and context offloading. However, most existing methods overlook the dimension of model depth. Repeatedly executing full-depth MLLM prefill over incoming frames is prohibitively expensive, incurring substantial computational overhead and causing the KV cache to grow at a rate directly proportional to the prefill depth. To address these challenges, we propose ShallowStream, a novel framework that leverages the shallow layers of an MLLM to simultaneously perform frame encoding and retrieval index building. During stream processing, ShallowStream maintains an always-on lightweight index using the KV cache of shallow layers. During query-time answering, we leverage the attention scores generated by the shallow layers to score context frames and employ a diversity-aware selection strategy to retrieve precise and comprehensive evidence. ShallowStream achieves performance on par with the strongest existing streaming methods, while reducing per-frame prefill latency and 10-second end-to-end latency by up to 52.1x and 11.9x, respectively. Our code is available at https://github.com/CURRENTF/ShallowStream.

Large language models rely on KV caching to reduce repeated prefill computation in long context and interactive applications. As lightweight adapters evolve, cached states reflect earlier versions, so stale reuse distorts current model outputs, while complete affected suffix recomputation restores fidelity at substantial cost. We seek minimal recomputation that recovers current adapter behavior. Existing systems track token, context, or stable adapter identity, but neither represent caches from earlier adapter versions nor distinguish update propagation from the recomputation required for behavioral recovery. To address these gaps, we introduce CacheReforge, which represents stale KV caches as layerwise mixed-version objects. It combines per-layer adapter anchors, calibrated sensitivity, accumulated drift, and executable restart boundaries to select direct reuse, bounded recomputation, or complete affected-suffix recovery. We distinguish dependency depth from the functional recomputation horizon and use cumulative tail influence to characterize when bounded recovery preserves current-model behavior. We evaluate CacheReforge on Qwen2.5-1.5B and Qwen2.5-7B with continual LoRA updates, including 16K HotpotQA and 2WikiMQA workloads. CacheReforge reduces mean KL divergence by 92.4% relative to stale reuse, while recomputing only 5.44% of layers and reducing cache-maintenance time by 93.2% relative to fresh full prefill. These results show that version-aware recovery preserves model fidelity and most KV caching gains.

Prompt-specialized multi-agent systems enable multiple agents to share a model while performing complementary roles to solve complex tasks. However, agent-specific prefixes change the KV cache generated for the same shared context, causing each agent to repeatedly prefill the growing context and construct a separate cache with high computation and memory overhead. Selective recomputation reduces this redundancy but still retains substantial model execution, while existing delta correction methods either support only recurring context relations or maintain memory-intensive online correction states for dynamically changing context. For first seen shared context, these methods also construct a reference cache outside the agent workflow, and an approximate correction at the first agent affects the outputs passed to subsequent agents. We present KVCMAS, an online KV cache correction framework that represents cross-agent cache deviations using compact low-rank states and seamlessly chains corrections along the agent workflow without an additional reference prefill. This design supports dynamically changing shared context while preserving an exact first-agent cache. Across multiple language and vision-language workloads, KVCMAS matches or improves the accuracy of prior KV cache sharing methods while achieving the lowest TTFT under highly concurrent serving. Under controlled serving traces, it provides a 2.0x TTFT speedup over inference without KV cache sharing and reduces peak GPU memory by up to 3.7x relative to a prior KV cache correction method. These results establish KVCMAS as an accurate and scalable KV cache sharing approach for prompt-specialized multi-agent serving.

End-to-end document parsers provide a unified interface, but serialize page layouts and regional contents into one autoregressive sequence. This formulation forces independent regions onto a decoding path whose length grows with the total content, whereas crop-based two-stage parsers expose region-level parallelism at the cost of repeated visual prefills and fragmented page context. To retain full-page context while removing dependencies, we propose PaDoc, a layout-grounded parser that treats the predicted layout as a branching structure over a shared page representation. Under a region-sufficiency assumption, we derive a prefix-conditioned factorization in which the layout stream and regional content branches advance concurrently, reducing the decoding depth to the longest layout-content path. We realize this factorization within a single MLLM: packed variable-length ancestor attention preserves the visibility under standard next-token training, while masked parallel decoding creates branches that the evaluated vLLM backend serves as concurrent requests with cache-resident shared-prefix reuse. On OmniDocBench Full, PaDoc attains an Overall layout F1 of 91.1 and, among end-to-end parsers, a top-tier Overall score of 94.24 together with the best Text Edit (0.038) and Formula CDM (95.59). On a 384-page subset and one A800 GPU, it is the fastest end-to-end parser at five concurrency levels, improving valid-page throughput by 67.4-118% and reducing P95 latency by 39.2-54.9% relative to a same-backbone Sequential SFT baseline. Code is available at https://github.com/Longin-Yu/Padoc

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

Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce memory traffic during generation. We propose "disaggregated quantization" (DQ), which specializes computation formats, weights and storage placement to both of these phases. On Qwen 3 and Gemma 3, removing activation quantization specifically on decode improves accuracy on decode-heavy tasks without increasing inference cost. Training separate compute-native prefill weights accelerates prompt processing relative to weight-only inference while matching or exceeding its accuracy at 2-3-bit decode on both decode-heavy and prefill-heavy tasks. With released Qwen3.8-27B GGUF decoders, training an NVFP4 prefiller improves 1-bit accuracy by 32.5 points on MMLU-Pro and 35.3 on MMMU-Pro without modifying the decode checkpoint. To accommodate the additional checkpoint on a single device, offloaded disaggregated prefill (ODP) streams its weights from SSD, amortizing loading over prompt length. On the same 27B model, ODP delivers a 1.78x time-to-first-token speedup over the weight-only baseline at 8K prompt length in llama.cpp. We evaluate accuracy under disaggregated serving in vLLM and further validate shared-weight format disaggregation through post-training quantization on models up to 2.8T parameters.

DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases. LLM applications often emphasize individual latency for each phase: time to first token (TTFT) for the prefill phase and time per output token (TPOT) of each request for the decoding phase. In the presence of stringent latency requirements, existing systems have to prioritize one latency over the other, or over-provision compute resources to meet both. DistServe assigns prefill and decoding computation to different GPUs, hence eliminating prefill-decoding interferences. Given the application's TTFT and TPOT requirements, DistServe co-optimizes the resource allocation and parallelism strategy tailored for each phase. DistServe also places the two phases according to the serving cluster's bandwidth to minimize the communication caused by disaggregation. As a result, DistServe significantly improves LLM serving performance in terms of the maximum rate that can be served within both TTFT and TPOT constraints on each GPU. Our evaluations show that on various popular LLMs, applications, and latency requirements, DistServe can serve 4.48x more requests or 10.2x tighter SLO, compared to state-of-the-art systems, while staying within latency constraints for > 90% of requests.

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.

Mixture-of-experts (MoE) models activate few experts per token, yet batched decoding can access nearly the entire expert pool, making expert-weight traffic a major bottleneck. Expert pruning reduces this traffic, but conventional approaches also prune compute-bound prefill, sacrificing model quality for little throughput benefit. We present SlimWise, a serving framework that tailors the expert pool to each inference phase. SlimWise performs prefill with the full model and decode with a pruned model that directly reuses the prefill-generated KV cache without conversion. Across two MoE backbones and three pruning criteria, this training-free KV cache handoff substantially narrows accuracy gaps relative to the full model in many settings. We also show that benchmark accuracy can conceal substantial pruning-induced changes in generation length. To address these distortions and residual accuracy loss, SlimWise introduces a low-cost distillation stage that trains the decoder to continue from full-model KV caches while updating only a small subset of parameters. Implemented in vLLM, SlimWise supports both prefill-decode (PD) disaggregation and PD-colocated serving. On Qwen3.6-35B-A3B, SlimWise improves decode throughput by up to 1.81x at 50% expert pruning with minimal accuracy loss.

While large language models (LLMs) excel at handling long-context sequences, they require substantial prefill computation and key-value (KV) cache, which can heavily burden computational efficiency and memory usage in both prefill and decoding stages. Recent works that compress KV caches with prefill acceleration reduce this cost but inadvertently tie the prefill compute reduction to the decoding KV budget. This coupling arises from overlooking the layer-dependent variation of critical context, often leading to accuracy degradation. To address this issue, we introduce FastKV, a KV cache compression framework designed to reduce latency in both prefill and decoding by leveraging the stabilization of token importance in later layers. FastKV performs full-context computation until a Token-Selective Propagation (TSP) layer, which forwards only the most informative tokens to subsequent layers. From these propagated tokens, FastKV independently selects salient KV entries for caching, thereby decoupling KV budget from the prefill compute reduction based on the TSP decision. This independent control of the TSP rate and KV retention rate enables flexible optimization of efficiency and accuracy. Experimental results show that FastKV achieves speedups of up to 1.82times in prefill and 2.87times in decoding compared to the full-context baseline, while matching the accuracy of the baselines that only accelerate the decoding stage. Our code is available at https://github.com/dongwonjo/FastKV.

With the emergence of wearable devices and other embedded systems, deploying large language models (LLMs) on edge platforms has become an urgent need. However, this is challenging because of their high computational and memory demands. Although recent low-bit quantization methods (e.g., BitNet, DeepSeek) compress weights to as low as 1.58~bits with minimal accuracy loss, edge deployment is still constrained by limited on-chip resources, power budgets, and the often-neglected long latency of the prefill stage. We present TeLLMe, the first table-lookup-based ternary LLM accelerator for low-power edge FPGAs that fully supports both prefill and autoregressive decoding using 1.58-bit weights and 8-bit activations. TeLLMe incorporates several novel techniques, including (1) a table-lookup-based ternary matrix multiplication (TLMM) engine utilizing grouped activations and online precomputation for low resource utilization and high throughput; (2) a fine-grained analytic URAM-based weight buffer management scheme for efficient loading and compute engine access; (3) a streaming dataflow architecture that fuses floating-point element-wise operations with linear computations to hide latency; (4) a reversed-reordered prefill stage attention with fused attention operations for high memory efficiency; and (5) a resource-efficient specialized decoding stage attention. Under a 5~W power budget, TeLLMe delivers up to 25~tokens/s decoding throughput and 0.45--0.96~s time-to-first-token (TTFT) for 64--128 token prompts, marking a significant energy-efficiency advancement in LLM inference on edge FPGAs.

Large Language Model (LLM) inference on large-scale systems is expected to dominate future cloud infrastructures. Efficient LLM inference in cloud environments with numerous AI accelerators is challenging, necessitating extensive optimizations for optimal performance. Current systems batch prefill and decoding to boost throughput but encounter latency issues, while others disaggregate these phases, leading to resource underutilization. We propose AcceLLM, a novel method addressing latency and load balancing, inspired by the cache data management. It strategically utilizes redundant data to enhance inference via load balancing and optimal hardware use. Simulated evaluations on Nvidia H100 GPU and Huawei Ascend 910B2 show AcceLLM surpasses state-of-the-art systems up to 30% in latency and efficiency, handling diverse workloads effectively.

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.

Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.

LLM inference for popular enterprise use cases, such as summarization, RAG, and code-generation, typically observes orders of magnitude longer prompt lengths than generation lengths. This characteristic leads to high cost of prefill and increased response latency. In this paper, we present SwiftKV, a novel model transformation and distillation procedure specifically designed to reduce the time and cost of processing prompt tokens while preserving high quality of generated tokens. SwiftKV combines three key mechanisms: i) SingleInputKV, which prefills later layers' KV cache using a much earlier layer's output, allowing prompt tokens to skip much of the model computation, ii) AcrossKV, which merges the KV caches of neighboring layers to reduce the memory footprint and support larger batch size for higher throughput, and iii) a knowledge-preserving distillation procedure that can adapt existing LLMs for SwiftKV with minimal accuracy impact and low compute and data requirement. For Llama-3.1-8B and 70B, SwiftKV reduces the compute requirement of prefill by 50% and the memory requirement of the KV cache by 62.5% while incurring minimum quality degradation across a wide range of tasks. In the end-to-end inference serving using an optimized vLLM implementation, SwiftKV realizes up to 2x higher aggregate throughput and 60% lower time per output token. It can achieve a staggering 560 TFlops/GPU of normalized inference throughput, which translates to 16K tokens/s for Llama-3.1-70B in 16-bit precision on 4x H100 GPUs.

LAPS identifies and disaggregates requests with different prompt lengths in LLM serving to reduce TTFT latency. While recent systems have decoupled the prefill and decode stages to improve throughput, they still rely on unified scheduling policies that fail to adapt to heterogeneous workload characteristics. We observe that prompt-length variations lead to distinct performance bottlenecks, motivating an adaptive scheduling strategy. LAPS disaggregates multi-turn long-prefill requests from short-prefill ones and introduces a length-aware smart batching mechanism for short-prefill workloads. It adopts a dual-queue design that supports temporal disaggregation on a single prefill instance or spatial disaggregation across multiple instances. For short-prefill batches, a batch waiting window and CUDA Graph-based clustering mitigate interference from heterogeneous computation, reducing batching delay and lowering average latency. In real multi-turn workloads, LAPS reduces prefill latency by over 30\% compared to vanilla SGLang under prefill-decode disaggregation, and further decreases SLO violations by 28\% in multi-instance deployments with vanilla data-parallel configuration. Compared to the SGLang router with load balancing, it further lowers SLO violations by 12\% in multi-GPU settings. Under high concurrency and mixed-request scenarios, LAPS improves request throughput by 35\% serving Qwen2.5-32B model for prefill instance, demonstrating its effectiveness in optimizing heterogeneous LLM serving workloads.

Large Language Models (LLMs) in agentic workflows combine multi-step reasoning, tool use, and collaboration across multiple specialized agents. Existing LLM serving engines optimize individual calls in isolation, while multi-agent frameworks focus on orchestration without system-level performance planning. As a result, repeated prompts, overlapping contexts, and concurrent executions create substantial redundancy and poor GPU utilization, especially in batch analytics scenarios. We introduce Halo, a system that brings batch query processing and optimization into agentic LLM workflows. Halo represents each workflow as a structured query plan DAG and constructs a consolidated graph for batched queries that exposes shared computation. Guided by a cost model that jointly considers prefill and decode costs, cache reuse, and GPU placement, Halo performs plan-level optimization to minimize redundant execution. Its runtime integrates adaptive batching, KV-cache sharing and migration, along with compute-communication overlap to maximize hardware efficiency. Evaluation across six benchmarks shows that Halo achieves up to 18.6x speedup for batch inference and 4.7x throughput improvement under online serving, scaling to workloads of tens of thousands of queries and complex graphs. These gains are achieved without compromising output quality. By unifying query optimization with LLM serving, Halo enables efficient agentic workflows in data analytics and decision-making applications.

Prefill-decode (PD) disaggregation has become the standard architecture for large-scale LLM serving, but in practice its deployment boundary is still determined by KVCache transfer. In conventional dense-attention models, prefill generates huge KVCache traffics that keep prefill and decode tightly coupled within a single high-bandwidth network domain, limiting heterogeneous deployment and resource elasticity. Recent hybrid-attention architectures substantially reduce KVCache size, making cross-cluster KVCache transport increasingly plausible. However, smaller KVCache alone does not make heterogeneous cross-datacenter PD serving practical: real workloads remain bursty, request lengths are highly skewed, prefix caches are unevenly distributed, and inter-cluster bandwidth fluctuates. A naive design that fully externalizes prefill can therefore still suffer from congestion, unstable queueing, and poor utilization. We present Prefill-as-a-Service (PrfaaS), a cross-datacenter serving architecture that selectively offloads long-context prefill to standalone, compute-dense prefill clusters and transfers the resulting KVCache over commodity Ethernet to local PD clusters for decode. Rather than treating reduced KVCache as sufficient, PrfaaS combines model-side KV efficiency with system-side selective offloading, bandwidth-aware scheduling, and cache-aware request placement. This design removes the requirement that heterogeneous accelerators share the same low-latency RDMA fabric, enabling independent scaling of prefill and decode capacity across loosely coupled clusters. In a case study using an internal 1T-parameter hybrid model, a PrfaaS-augmented heterogeneous deployment achieves 54% higher serving throughput and 64% lower P90 TTFT than a homogeneous PD baseline, with approximately 15% throughput gain at equal cost, while consuming only modest cross-datacenter bandwidth.

As large language models serve ever more requests, cumulative inference cost is growing relative to the one-time cost of training. In typical serving, prompt prefill runs in parallel and is compute-bound, whereas autoregressive decode is sequential and memory-traffic-bound. Conventional width or depth scaling raises both costs together, since every added layer is evaluated in both phases and enlarges the weights read at each decode step. We instead ask whether additional learned computation can be allocated to continuation prediction while preserving prompt-wide primary computation and a single KV cache. We realize this with the Decode-Branch Transformer. Its primary path alone processes the prompt and writes the KV cache; the decode branch is omitted during prefill and activated only from the final prompt position onward, adding continuation computation without writing state or affecting the primary path. The paths share attention, MLP, and output matrices, using separate token embeddings with lightweight coupling. Grouped decode reuses loaded weight tiles and the primary KV cache across both paths, so the added arithmetic does not proportionally increase dominant memory traffic or decode latency. Across matched-token comparisons, Decode-Branch achieves lower validation loss across architectures and data settings. In MoE models, the primary and branch expert fan-outs become independent knobs for trading prompt cost, decode cost, and predictive quality. We study two expert-allocation regimes, holding prefill or decode computation fixed, and expose a prefill-decode-quality trade-off enabled by phase-specific expert allocation.

Self-attention dominates the computational and memory cost of long-context LLM inference across both prefill and decode phases. To address this challenge, we introduce Sketch&Walk Attention, a training-free sparse attention method that determines sparsity with lightweight sketches and deterministic walk. Sketch&Walk applies Hadamard sketching to get inexpensive approximations of attention scores, then aggregates these estimates across layers via a walk mechanism that captures attention influence beyond direct interactions between tokens. The accumulated walk scores are used to select top-k attention blocks, enabling dynamic sparsity with a single training-free algorithm that applies uniformly to both the prefill and decode phases, together with custom sparse attention kernels. Across a wide range of models and tasks, Sketch&Walk maintains near-lossless accuracy at 20% attention density and can slightly outperform dense attention in some settings, while achieving up to 6x inference speedup.

LLM serving systems are provisioned for peak load to meet strict latency targets, leaving substantial GPU compute idle whenever traffic falls below peak. We present DeltaServe, a host-agnostic co-serving design that converts this idle inference capacity into LoRA fine-tuning throughput while preserving inference service-level objectives (SLOs). DeltaServe integrates with existing inference engines through a compact hook interface that requires only multi-LoRA batching support. It exploits the shared execution structure of inference prefill and LoRA fine-tuning forward passes, and uses an SLO-aware scheduler to admit and execute fine-tuning only when sufficient inference headroom is available. The scheduler is driven by a CUDA-graph-aware latency model calibrated offline and refined online. We integrate DeltaServe with vLLM, SGLang, and S-LoRA. On a production trace from Company X, DeltaServe on vLLM delivers 2.9x higher fine-tuning throughput than LLMStation at 100% inference SLO compliance, versus 85% for LLMStation. It also achieves 39% higher fine-tuning throughput than a baseline running vLLM+torchtune, using no additional hardware and maintaining full SLO compliance.

Recent multi-agent LLM systems increasingly combine heterogeneous models for specialized agent roles. However, text-based communication requires each receiver to prefill shared context already processed by the sender. Reusing the sender's key-value (KV) cache avoids this redundancy, but prefill-free transfer across model families must handle differences in tokenization, model depth, and KV representations. To address these issues, we propose HeteroFold, a prefill-free cross-family KV cache transfer method that keeps both the sender and receiver frozen. HeteroFold aligns model structures, maps the sender cache into the receiver space, and calibrates it to preserve receiver behavior. Across six transfer directions, HeteroFold achieves the best cache-transfer performance on all four long-context benchmarks and most short-context settings. It also matches text-based communication on the multi-agent benchmark. At 32K context length, Llama-3.1-8BrightarrowMinistral-3-14B transfer is 10.7times faster than Native Prefill and 1.18--1.47times faster than the state-of-the-art prefill-free baselines, Dense Latent and KV Ridge. These results show that HeteroFold enables efficient cross-family KV reuse without receiver prefill.

Large Language Model (LLM) inference in production must meet stringent service-level objectives for both time-to-first-token (TTFT) and time-between-token (TBT) while maximizing throughput under fixed compute, memory, and interconnect budgets. Modern serving systems adopt stall-free scheduling techniques such as chunked prefill, which splits long prompt processing along the token dimension and interleaves prefill with ongoing decode iterations. While effective at stabilizing TBT, chunked prefill incurs substantial overhead in Mixture-of-Experts (MoE) models: redundant expert weight loads increase memory traffic by up to 39% and inflate energy consumption. We propose layered prefill, a new scheduling paradigm that treats transformer layer groups as the primary scheduling unit. By vertically partitioning the model into contiguous layer groups and interleaving prefill and decode across the groups, layered prefill sustains stall-free decoding while eliminating chunk-induced MoE weight reloads. It reduces off-chip bandwidth demand, lowering TTFT by up to 70%, End-to-End latency by 41% and per-token energy by up to 22%. Evaluations show that layered prefill consistently improves the TTFT--TBT Pareto frontier over chunked prefill, reducing expert-load traffic and energy cost while maintaining stall-free decoding. Overall, shifting the scheduling axis from tokens to layers unlocks a new operating regime for high-efficiency, energy-aware LLM serving in co-located environments.

Large Language Models (LLMs) and Vision-Language Models (VLMs) have demonstrated remarkable capabilities. However, their deployment is hindered by significant computational costs. Existing structured pruning methods, while hardware-efficient, often suffer from significant accuracy degradation. In this paper, we argue that this failure stems from a stage-agnostic pruning approach that overlooks the asymmetric roles between the prefill and decode stages. By introducing a virtual gate mechanism, our importance analysis reveals that deep layers are critical for next-token prediction (decode) but largely redundant for context encoding (prefill). Leveraging this insight, we propose Prefill-Only Pruning (POP), a stage-aware inference strategy that safely omits deep layers during the computationally intensive prefill stage while retaining the full model for the sensitive decode stage. To enable the transition between stages, we introduce independent Key-Value (KV) projections to maintain cache integrity, and a boundary handling strategy to ensure the accuracy of the first generated token. Extensive experiments on Llama-3.1, Qwen3-VL, and Gemma-3 across diverse modalities demonstrate that POP achieves up to 1.37times speedup in prefill latency with minimal performance loss, effectively overcoming the accuracy-efficiency trade-off limitations of existing structured pruning methods.

Serving Large Language Models (LLMs) is a GPU-intensive task where traditional autoscalers fall short, particularly for modern Prefill-Decode (P/D) disaggregated architectures. This architectural shift, while powerful, introduces significant operational challenges, including inefficient use of heterogeneous hardware, network bottlenecks, and critical imbalances between prefill and decode stages. We introduce HeteroScale, a coordinated autoscaling framework that addresses the core challenges of P/D disaggregated serving. HeteroScale combines a topology-aware scheduler that adapts to heterogeneous hardware and network constraints with a novel metric-driven policy derived from the first large-scale empirical study of autoscaling signals in production. By leveraging a single, robust metric to jointly scale prefill and decode pools, HeteroScale maintains architectural balance while ensuring efficient, adaptive resource management. Deployed in a massive production environment on tens of thousands of GPUs, HeteroScale has proven its effectiveness, increasing average GPU utilization by a significant 26.6 percentage points and saving hundreds of thousands of GPU-hours daily, all while upholding stringent service level objectives.

Efficient key-value (KV) cache management is crucial for the practical deployment of large language models (LLMs), yet existing compression techniques often incur a trade-off between performance degradation and computational overhead. We propose a novel gating-based KV cache eviction method for frozen-weight LLMs that achieves high compression ratios with negligible computational cost. Our approach introduces lightweight sink-attention gating modules to identify and retain critical KV pairs, and integrates seamlessly into both the prefill and decoding stages. The proposed gate training algorithm relies on forward passes of an LLM, avoiding expensive backpropagation, while achieving strong task generalization through a task-agnostic reconstruction objective. Extensive experiments across the Qwen2.5-1M, Qwen3, and Gemma3 families show that our method maintains near-lossless performance while evicting up to 70% of the KV cache. The results are consistent across a wide range of tasks, including long-context understanding, code comprehension, and mathematical reasoning, demonstrating the generality of our approach.

Edge LLM deployment is shaped by more than model size and precision: inference backend, hardware platform, memory traffic, and power management all affect latency and efficiency. We present Hydra, a common-schema, phase-aware workload characterization framework for LLM inference on edge SoCs. Hydra instruments HuggingFace Transformers and llama.cpp with a shared per-prompt timing schema and fuses those records with hardware telemetry, enabling a multi-dimensional characterization of performance, system-resource utilization, and efficiency across prefill and decode phases. Using Hydra, we evaluate three consecutive edge System-on-Chip (SoC) generations (AGX Xavier, AGX Orin, and AGX Thor), 13 instruction-tuned LLMs from seven families, five execution formats, and consider input/output-length sensitivity. The resulting artifact contains roughly 107K per-prompt records and is publicly released with Hydra. Our analysis shows that aggregate latency alone hides key deployment effects: backend structure changes where latency is introduced, quantization reduces memory traffic and energy but does not predict power monotonically, and SoC generation changes how utilization and efficiency should be interpreted. By connecting phase-level timing with system-resource utilization and efficiency metrics, Hydra enables reproducible, phase-aware characterization of edge LLM inference. Hydra's source code and the collected per-prompt trace corpus are available open-source at: https://github.com/amirtaherin/hydra

Disaggregated inference has become an essential framework that separates the prefill (P) and decode (D) stages in large language model inference to improve throughput. However, the KV cache transfer faces significant delays between prefill and decode nodes. The block-wise calling method and discontinuous KV cache memory allocation increase the number of calls to the transmission kernel. Additionally, existing frameworks often fix the roles of P and D nodes, leading to computational imbalances. In this paper, we propose FlowKV, a novel disaggregated inference framework, which reduces the average transmission latency of KV cache by 96%, from 0.944s to 0.053s, almost eliminating the transfer time relative to the total request latency by optimizing the KV cache transfer. FlowKV introduces the Load-Aware Scheduler for balanced request scheduling and flexible PD node allocation. This design maximizes hardware resource utilization, achieving peak system throughput across various scenarios, including normal, computational imbalance, and extreme overload conditions. Experimental results demonstrate that FlowKV significantly accelerates inference by 15.2%-48.9% on LongBench dataset compared to the baseline and supports applications with heterogeneous GPUs.

Recent advances in long-text understanding have pushed the context length of large language models (LLMs) up to one million tokens. It boosts LLMs's accuracy and reasoning capacity but causes exorbitant computational costs and unsatisfactory Time to First Token (TTFT). KV cache reuse, which reuses the exact same KV cache of prefixes and templates or shares similar ones but with extra selective recomputation, offers a promising way to tackle this issue. However, prior studies overlook the cross-request KV reuse and the attention deviations introduced by new tokens during the decoding stage. In this paper, we present a KV cache management module that shares the KV cache across requests under multi-tenant scenarios without sacrificing model accuracy. Our system, KVShare, enables accurate and efficient LLM serving by 1) a Dual-Stage High Deviation algorithm (DHD) that conditionally selects a small portion of KV cache to be recomputed during both prefill and decode phases, and 2) a cache-aware scheduler that prioritizes requests based on their KV cache hit rates and orchestrates continuous batching to achieve enhanced system efficiency and faster TTFT. Multi-task experiments conducted on models such as Qwen2.5-7B,Llama3.1-8B and Yi1.5-9B demonstrate that KVShare reduces TTFT by up to 9.39x and increases 1.2x of the throughput compared to the full KV recompute. Moreover, KVShare achieves 20.38% boost in terms of accuracy compared to SOTA methods.

Agentic inference now dominates the LLM inference landscape, requiring LLMs to actively engage in multi-turn interactions with tool-calling capabilities. This introduces a more complex workload for the underlying inference system: serving stages such as prefill and decode exhibit substantially different behaviors and demand distinct compute and memory-bandwidth capabilities. As a result, a single homogeneous GPU system now struggles to support agentic inference, motivating an industry shift toward heterogeneous systems with disaggregated serving capabilities, such as the emerging Vera-Rubin platform with GPUs and Groq LPUs. However, the question of what the optimal hardware should look like for each component in a heterogeneous system remains underexplored. To this end, we propose a novel simulation framework for disaggregated serving, termed HeteroPanacea, that enables system-level simulation across three dimensions: 1) disaggregated quantization, 2) automated intra- and inter-device parallelization scheduling, and 3) PDAF (prefill-decode-attention-FFN) NPU architectural heterogeneity. By combining these three axes, we provide a cross-stack simulation framework for future heterogeneous agentic serving systems. We confirm the benefit of Prefill Decode disaggregation, simulating increased serving throughput by up to 75\% compared to traditional serving with current GPUs and demonstrate 4 way Prefill Decode Attention FFN disaggregation is the most consistent for increasing throughput across different models, assuming custom NPUs. We also investigate the relationship between model architecture and gain from disaggregation by running a set of ablation studies.

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

Large Language Model (LLM) inference consists of two distinct phases - prefill phase which processes the input prompt and decode phase which generates output tokens autoregressively. While the prefill phase effectively saturates GPU compute at small batch sizes, the decode phase results in low compute utilization as it generates one token at a time per request. The varying prefill and decode times also lead to imbalance across micro-batches when using pipeline parallelism, resulting in further inefficiency due to bubbles. We present SARATHI to address these challenges. SARATHI employs chunked-prefills, which splits a prefill request into equal sized chunks, and decode-maximal batching, which constructs a batch using a single prefill chunk and populates the remaining slots with decodes. During inference, the prefill chunk saturates GPU compute, while the decode requests 'piggyback' and cost up to an order of magnitude less compared to a decode-only batch. Chunked-prefills allows constructing multiple decode-maximal batches from a single prefill request, maximizing coverage of decodes that can piggyback. Furthermore, the uniform compute design of these batches ameliorates the imbalance between micro-batches, significantly reducing pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware. For the LLaMA-13B model on A6000 GPU, SARATHI improves decode throughput by up to 10x, and accelerates end-to-end throughput by up to 1.33x. For LLaMa-33B on A100 GPU, we achieve 1.25x higher end-to-end-throughput and up to 4.25x higher decode throughput. When used with pipeline parallelism on GPT-3, SARATHI reduces bubbles by 6.29x, resulting in an end-to-end throughput improvement of 1.91x.

Large language models (LLMs) excel across diverse tasks but face significant deployment challenges due to high inference costs. LLM inference comprises prefill (compute-bound) and decode (memory-bound) stages, with decode dominating latency particularly for long sequences. Current decoder-only models handle both stages uniformly, despite their distinct computational profiles. We propose OverFill, which decouples these stages to optimize accuracy-efficiency tradeoffs. OverFill begins with a full model for prefill, processing system and user inputs in parallel. It then switches to a dense pruned model, while generating tokens sequentially. Leveraging more compute during prefill, OverFill improves generation quality with minimal latency overhead. Our 3B-to-1B OverFill configuration outperforms 1B pruned models by 83.2%, while the 8B-to-3B configuration improves over 3B pruned models by 79.2% on average across standard benchmarks. OverFill matches the performance of same-sized models trained from scratch, while using significantly less training data. Our code is available at https://github.com/friendshipkim/overfill.

Disaggregated inference architectures physically separate prefill and decode phases onto distinct GPU pools, creating competing "agents" that share a fixed hardware budget. We provide, to our knowledge, the first formal game-theoretic analysis of this architecture, using NVIDIA Dynamo as a concrete case study. We model disaggregated serving as three coupled games: a two-player resource game between prefill and decode pools, a selfish caching game over the hierarchical KV cache, and a congestion game with positive externalities for request routing. We empirically validate the latter two; the P/D resource game is treated analytically (Section 9.2). We characterize how GPU saturation induces regime transitions that shift the game's payoff structure: below saturation, selfish behavior has bounded Price of Anarchy (PoA); at saturation, superlinear latency and cache externalities drive our empirical estimator PoA-hat (defined in Section 6.4) upward. Based on this analysis, we design an adaptive controller that detects saturation transitions in real time and adjusts routing parameters accordingly, shifting from cache-affinity exploitation to load-balanced congestion avoidance. We instantiate our framework on a 3-node NVIDIA B200 cluster running Dynamo with two models, Nemotron-4-340B (TP=8, full-node workers with cross-InfiniBand KV transfers) and Llama-3.1-70B (TP=4), and find the same three-regime PoA-hat structure with the same first post-knee grid point (C=128) on both models. Adaptive routing shifts each model to a better operating point. Our strongest result is on the 70B 1P/5D topology, where PoA-hat drops 3.1x (66.4 to 21.5) in the saturated phase at a 13% throughput cost. On the 70B 1P/2D, PoA-hat drops 2.2x and TTFT P99 drops 7.6x (see Section 8.5).

Large language models (LLMs) are increasingly deployed on customer devices. To support them, current devices are adopting SoCs (System on Chip) with NPUs (Neural Processing Unit) installed. Although high performance is expected, LLM inference on NPUs is slower than its CPU counterpart. The reason is that NPUs have poor performance on computations other than GEMM, like dequantization. Current works either disaggregate prefill on the NPUs and decoding on the CPUs, or put both on the NPUs but with an accuracy loss. To solve this issue, based on the insight that low-bit can enable target computation encoded within an acceptably sized table, we propose table lookup to subsume hardware operations otherwise unsupported. To realize this, we overcome the conflicting hardware behavior of prefill and decoding to design a unified table layout and tiling through (1) fused two-level table-based dequantization and (2) concurrency-hierarchy-guided tiling. Based on that, we implement the prefill phase by three-stage pipeline and map the table-lookup-based decoding to NPU's vector units. Results show 1.4x and 3.1x speedup for prefill and decoding respectively, and 84% energy savings compared to the baseline NPU methods. The code is available at https://github.com/microsoft/T-MAC/tree/main/t-man.

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.

Recent advances in large language models (LLMs) have extended context lengths, enabling assistants to sustain long histories for coherent, personalized responses. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly dominates under strict resource constraints. An active line of research for reducing this overhead is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting entries after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to degraded accuracy in multi-turn conversations. We introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and applies episode-specific KV cache eviction. We further design an adaptive layer-wise budget allocation strategy that measures each layer's sensitivity to eviction and distributes the memory budget across layers accordingly. Across three LongConvQA benchmarks, EpiCache improves accuracy by up to 40% over recent baselines, sustains near-full KV accuracy under 4-6x compression, and reduces latency and memory by up to 2.4x and 3.5x, thereby enabling efficient multi-turn interaction under strict resource constraints.

AI agent inference is driving an inference heavy datacenter future and exposes bottlenecks beyond compute - especially memory capacity, memory bandwidth and high-speed interconnect. We introduce two metrics - Operational Intensity (OI) and Capacity Footprint (CF) - that jointly explain regimes the classic roofline analysis misses, including the memory capacity wall. Across agentic workflows (chat, coding, web use, computer use) and base model choices (GQA/MLA, MoE, quantization), OI/CF can shift dramatically, with long context KV cache making decode highly memory bound. These observations motivate disaggregated serving and system level heterogeneity: specialized prefill and decode accelerators, broader scale up networking, and decoupled compute-memory enabled by optical I/O. We further hypothesize agent-hardware co design, multiple inference accelerators within one system, and high bandwidth, large capacity memory disaggregation as foundations for adaptation to evolving OI/CF. Together, these directions chart a path to sustain efficiency and capability for large scale agentic AI inference.

The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the standard attention mechanism. To address this challenge, we introduce BLASST, a drop-in sparse attention method that dynamically prunes the attention matrix without any pre-computation or proxy scores. Our method uses a fixed threshold and existing information from online softmax to identify negligible attention scores, skipping softmax computation, Value block loading, and the subsequent matrix multiplication. This fits seamlessly into existing FlashAttention kernel designs with negligible latency overhead. The approach is applicable to both prefill and decode stages across all attention variants (MHA, GQA, MQA, and MLA), providing a unified solution for accelerating long-context inference. We develop an automated calibration procedure that reveals a simple inverse relationship between optimal threshold and context length, enabling robust deployment across diverse scenarios. Maintaining high accuracy, we demonstrate a 1.62x speedup for prefill at 74.7% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs. Furthermore, we explore sparsity-aware training as a natural extension, showing that models can be trained to be inherently more robust to sparse attention patterns, pushing the accuracy-sparsity frontier even further.

We introduce Opt.Gear, a foundation model designed for efficient on-device deployment, real-tim inference, and strong task capability. It includes a dense model (1M, 270M, and 1B) with a context length of 64K. We designed a new hybrid architecture that combines a convolutional key-value gated mixer with local-global attention to reduce the KV-cache memory that tends to increase exponentially with long context. This architecture delivers up to X4.9 faster prefill and decoding speeds on the NPUs compared to models of a similar scale models. From a 2T tokens candidate corpus, Opt.Gear is trained on a curated 0.5T tokens subset without knowledge distillation. This is the most data-efficient of the existing foundation models. All models are released with open weights and deployment binaries for ONNX, Qualcomm NPU, and Apple ANE making Opt.Gear a practical base for edge applications that need fast, memory-efficient inference and strong task capabilities. Furthermore, to expand the ecosystem of on-device generative language models, we are introducing the Opt.Gear-1M that can be deployed on Micro-Controller Units (MCUs), a Tiny Language Model (TLM). Opt.Gear-1M is the first generative language model to achieve 20 TPS with W4A32 quantization on the ARM Cortex-M7 CPU of the STM32H747I-DISCO.

We present RAGMark, a modular benchmarking framework for advanced Retrieval-Augmented Generation (RAG) systems targeting small-scale multi-GPU environments. RAGMark evaluates diverse RAG components, including retrievers, vector databases, prompt-processing methods, and generator models, while collecting detailed per-stage metrics such as latency, GPU utilization, memory consumption, power usage, time to first token (TTFT), throughput, and answer quality. The framework is highly extensible, separating RAG stages, timing, and resource monitoring into modular components, and is designed to efficiently sweep large configuration spaces while minimizing repeated model and database initialization overhead. Using RAGMark, we characterize five RAG workloads on open-domain QA datasets across varying retrieval depths, model scales, reranking, compression methods, and vector database configurations. We show that while autoregressive generation dominates latency in naive pipelines, context-reduction techniques shift bottlenecks across compute, memory bandwidth, and preprocessing stages. Reranking and compression produce compounding benefits: reranking reduces compression workload itself, while both jointly reduce prefill and KV-cache traversal costs, lowering energy consumption by up to 66%. We further observe strong cross-stage interactions, where small upstream context reductions cascade through downstream latency, memory traffic, and energy consumption. The RAGMark source code is publicly available at: https://github.com/zferic/RAGMark.

The Large Language Model (LLM) is widely employed for tasks such as intelligent assistants, text summarization, translation, and multi-modality on mobile phones. However, the current methods for on-device LLM deployment maintain slow inference speed, which causes poor user experience. To facilitate high-efficiency LLM deployment on device GPUs, we propose four optimization techniques: (a) a symbolic expression-based approach to support dynamic shape model inference; (b) operator optimizations and execution priority setting to enhance inference speed and reduce phone lagging; (c) an FP4 quantization method termed M0E4 to reduce dequantization overhead; (d) a sub-tensor-based technique to eliminate the need for copying KV cache after LLM inference. Furthermore, we implement these methods in our mobile inference engine, Transformer-Lite, which is compatible with both Qualcomm and MTK processors. We evaluated Transformer-Lite's performance using LLMs with varied architectures and parameters ranging from 2B to 14B. Specifically, we achieved prefill and decoding speeds of 121 token/s and 14 token/s for ChatGLM2 6B, and 330 token/s and 30 token/s for smaller Gemma 2B, respectively. Compared with CPU-based FastLLM and GPU-based MLC-LLM, our engine attains over 10x speedup for the prefill speed and 2~3x speedup for the decoding speed.

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. It features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache. The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs). Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake's innovative architecture enables Kimi to handle 75% more requests.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision understanding, reasoning, and interaction. However, the inference computation and memory increase progressively with the generation of output tokens during decoding, directly affecting the efficacy of MLLMs. Existing methods attempt to reduce the vision context redundancy to achieve efficient MLLMs. Unfortunately, the efficiency benefits of the vision context reduction in the prefill stage gradually diminish during the decoding stage. To address this problem, we proposed a dynamic vision-language context sparsification framework Dynamic-LLaVA, which dynamically reduces the redundancy of vision context in the prefill stage and decreases the memory and computation overhead of the generated language context during decoding. Dynamic-LLaVA designs a tailored sparsification inference scheme for different inference modes, i.e., prefill, decoding with and without KV cache, to achieve efficient inference of MLLMs. In practice, Dynamic-LLaVA can reduce computation consumption by sim75\% in the prefill stage. Meanwhile, throughout the entire generation process of MLLMs, Dynamic-LLaVA reduces the sim50\% computation consumption under decoding without KV cache, while saving sim50\% GPU memory overhead when decoding with KV cache, due to the vision-language context sparsification. Extensive experiments also demonstrate that Dynamic-LLaVA achieves efficient inference for MLLMs with negligible understanding and generation ability degradation or even performance gains compared to the full-context inference baselines. Code is available at https://github.com/Osilly/dynamic_llava .

Key-Value (KV) cache has become a bottleneck of LLMs for long-context generation. Despite the numerous efforts in this area, the optimization for the decoding phase is generally ignored. However, we believe such optimization is crucial, especially for long-output generation tasks based on the following two observations: (i) Excessive compression during the prefill phase, which requires specific full context impairs the comprehension of the reasoning task; (ii) Deviation of heavy hitters occurs in the reasoning tasks with long outputs. Therefore, SCOPE, a simple yet efficient framework that separately performs KV cache optimization during the prefill and decoding phases, is introduced. Specifically, the KV cache during the prefill phase is preserved to maintain the essential information, while a novel strategy based on sliding is proposed to select essential heavy hitters for the decoding phase. Memory usage and memory transfer are further optimized using adaptive and discontinuous strategies. Extensive experiments on LongGenBench show the effectiveness and generalization of SCOPE and its compatibility as a plug-in to other prefill-only KV compression methods.

Transformer-based large language model (LLM) inference serving is now the backbone of many cloud services. LLM inference consists of a prefill phase and a decode phase. However, existing LLM deployment practices often overlook the distinct characteristics of these phases, leading to significant interference. To mitigate interference, our insight is to carefully schedule and group inference requests based on their characteristics. We realize this idea in TetriInfer through three pillars. First, it partitions prompts into fixed-size chunks so that the accelerator always runs close to its computationsaturated limit. Second, it disaggregates prefill and decode instances so each can run independently. Finally, it uses a smart two-level scheduling algorithm augmented with predicted resource usage to avoid decode scheduling hotspots. Results show that TetriInfer improves time-to-first-token (TTFT), job completion time (JCT), and inference efficiency in turns of performance per dollar by a large margin, e.g., it uses 38% less resources all the while lowering average TTFT and average JCT by 97% and 47%, respectively.

Vision Language Models (VLMs) have rapidly advanced in integrating visual and textual reasoning, powering applications across high-resolution image understanding, long-video analysis, and multi-turn conversation. However, their scalability remains limited by the growing number of visual tokens that dominate inference latency. We present SparseVILA, a new paradigm for efficient VLM inference that decouples visual sparsity across the prefilling and decoding stages. SparseVILA distributes sparsity across stages by pruning redundant visual tokens during prefill and retrieving only query-relevant tokens during decoding. This decoupled design matches leading prefill pruning methods while preserving multi-turn fidelity by retaining most of the visual cache so that query-aware tokens can be retrieved at each conversation round. Built on an AWQ-optimized inference pipeline, SparseVILA achieves up to 4.0 times faster prefilling, 2.5 times faster decoding, and an overall 2.6 times end-to-end speedup on long-context video tasks -- while improving accuracy on document-understanding and reasoning tasks. By decoupling query-agnostic pruning and query-aware retrieval, SparseVILA establishes a new direction for efficient multimodal inference, offering a training-free, architecture-agnostic framework for accelerating large VLMs without sacrificing capability.

Serving disaggregated large language models (LLMs) over tens of thousands of xPU devices (GPUs or NPUs) with reliable performance faces multiple challenges. 1) Ignoring the diversity (various prefixes and tidal requests), treating all the prompts in a mixed pool is inadequate. To facilitate the similarity per scenario and minimize the inner mismatch on P/D (prefill and decoding) processing, fine-grained organization is required, dynamically adjusting P/D ratios for better performance. 2) Due to inaccurate estimation on workload (queue status or maintained connections), the global scheduler easily incurs unnecessary timeouts in prefill. 3) Block-fixed device-to-device (D2D) KVCache transfer over cluster-level RDMA (remote direct memory access) fails to achieve desired D2D utilization as expected. To overcome previous problems, this paper proposes an end-to-end system P/D-Serve, complying with the paradigm of MLOps (machine learning operations), which models end-to-end (E2E) P/D performance and enables: 1) fine-grained P/D organization, mapping the service with RoCE (RDMA over converged ethernet) as needed, to facilitate similar processing and dynamic adjustments on P/D ratios; 2) on-demand forwarding upon rejections for idle prefill, decoupling the scheduler from regular inaccurate reports and local queues, to avoid timeouts in prefill; and 3) efficient KVCache transfer via optimized D2D access. P/D-Serve is implemented upon Ascend and MindSpore, has been deployed over tens of thousands of NPUs for more than eight months in commercial use, and further achieves 60\%, 42\% and 46\% improvements on E2E throughput, time-to-first-token (TTFT) SLO (service level objective) and D2D transfer time. As the E2E system with optimizations, P/D-Serve achieves 6.7x increase on throughput, compared with aggregated LLMs.

Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency.

Sparse long-context inference requires efficient token retrieval in both prefill and decode. Existing methods often use different retrieval strategies for the two stages, preventing one retrieval representation from being reused throughout inference. We propose Self-Indexing Attention, a training-free framework built on a shared transform-domain sign-magnitude representation. The key signs provide a reusable token-level index for grouped prefill selection and decode retrieval, while the same representation remains compatible with external KV-cache compression without separate indexer metadata. This 1-bit index enables efficient retrieval through bitwise operations widely supported by modern accelerators. At 5% attention density, Self-Indexing Attention remains close to dense attention on LongBench and RULER and achieves up to 6.1x prefill and 10.3x decode attention-operator speedups. Experiments with TurboQuant and DeepSeekV4-Flash further demonstrate compatibility with low-bit KV-cache compression and pretrained sparse-attention indexers.

Quantization is a proven effective method for compressing large language models. Although popular techniques like W8A8 and W4A16 effectively maintain model performance, they often fail to concurrently speed up the prefill and decoding stages of inference. W4A8 is a promising strategy to accelerate both of them while usually leads to a significant performance degradation. To address these issues, we present QQQ, a Quality Quattuor-bit Quantization method with 4-bit weights and 8-bit activations. QQQ employs adaptive smoothing and Hessian-based compensation, significantly enhancing the performance of quantized models without extensive training. Furthermore, we meticulously engineer W4A8 GEMM kernels to increase inference speed. Our specialized per-channel W4A8 GEMM and per-group W4A8 GEMM achieve impressive speed increases of 3.67times and 3.29 times over FP16 GEMM. Our extensive experiments show that QQQ achieves performance on par with existing state-of-the-art LLM quantization methods while significantly accelerating inference, achieving speed boosts up to 2.24 times, 2.10times, and 1.25times compared to FP16, W8A8, and W4A16, respectively.

Emerging Large Language Model (LLM) applications require long input prompts to perform complex downstream tasks like document analysis and code generation. For these long context length applications, the length of the input prompt poses a significant challenge in terms of inference efficiency since the inference costs increase linearly with sequence length. However, for many of these applications, much of the context in the prompt is fixed across different user inputs, thereby providing the opportunity to perform offline optimizations to process user inputs quickly, as they are received. In this work, we propose Squeezed Attention as a mechanism to accelerate LLM applications where a large portion of the input prompt is fixed. We first leverage K-means clustering offline to group the keys for the fixed context based on semantic similarity and represent each cluster with a single centroid value. During inference, we compare query tokens from the user input with the centroids to predict which of the keys from the fixed context are semantically relevant and need to be loaded during inference. We then compute exact attention using only these important keys from the fixed context, thereby reducing bandwidth and computational costs. We also extend our method to use a hierarchical centroid lookup to identify important keys, which can reduce the complexity of attention from linear to logarithmic with respect to the context length. We implement optimized Triton kernels for centroid comparison and sparse FlashAttention with important keys, achieving more than 4x speedups during both the prefill and generation phases for long-context inference. Furthermore, we have extensively evaluated our method on various long-context benchmarks including LongBench, where it achieves a 3x reduction in KV cache budget without accuracy loss and up to an 8x reduction with <0.5 point accuracy gap for various models.

This paper investigates an under-explored challenge in large language models (LLMs): the impact of KV cache compression methods on LLMs' fundamental capabilities. While existing methods achieve impressive compression ratios on long-context benchmarks, their effects on core model capabilities remain understudied. We present a comprehensive empirical study evaluating prominent KV cache compression methods across diverse tasks, spanning world knowledge, commonsense reasoning, arithmetic reasoning, code generation, safety, and long-context understanding and generation.Our analysis reveals that KV cache compression methods exhibit task-specific performance degradation. Arithmetic reasoning tasks prove particularly sensitive to aggressive compression, with different methods showing performance drops of 17.4%-43.3%. Notably, the DeepSeek R1 Distill model exhibits more robust compression tolerance compared to instruction-tuned models, showing only 9.67%-25.53% performance degradation. Based on our analysis of attention patterns and cross-task compression performance, we propose ShotKV, a novel compression approach that distinctly handles prefill and decoding phases while maintaining shot-level semantic coherence. Empirical results show that ShotKV achieves 9%-18% performance improvements on long-context generation tasks under aggressive compression ratios.

Multi-headed Attention's (MHA) quadratic compute and linearly growing KV-cache make long-context transformers expensive to train and serve. Prior works such as Grouped Query Attention (GQA) and Multi-Latent Attention (MLA) shrink the cache, speeding decode, but leave compute, which determines prefill and training speed, largely unchanged. We introduce Compressed Convolutional Attention (CCA), a novel attention method which down-projects queries, keys, and values and performs the entire attention operation inside the shared latent space. This simple design dramatically cuts parameters, KV-cache, and FLOPs all at once by the desired compression factor. Because CCA is orthogonal to head-sharing, we combine the two to form Compressed Convolutional Grouped Query Attention (CCGQA), which further tightens the compute-bandwidth Pareto frontier so that users can tune compression toward either FLOP or memory limits without sacrificing quality. Experiments show that CCGQA consistently outperforms both GQA and MLA at equal KV-cache compression on dense and MoE models. Additionally, we show that CCGQA outperforms all other attention methods on MoE models with half the KV-cache of GQA and MLA, achieving an 8x KV-cache compression with no drop in performance compared to standard MHA. CCA and CCGQA also dramatically reduce the FLOP cost of attention which leads to substantially faster training and prefill than existing methods. On H100 GPUs, our fused CCA/CCGQA kernel reduces prefill latency by about 1.7x at a sequence length of 16k relative to MHA, and accelerates backward by about 1.3x.

Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG. We propose Kascade, a training-free sparse attention method that leverages known observations such as 1) post-softmax attention is intrinsically sparse, and 2) the identity of high-weight keys is stable across nearby layers. Kascade computes exact Top-k indices in a small set of anchor layers, then reuses those indices in intermediate reuse layers. The anchor layers are selected algorithmically, via a dynamic-programming objective that maximizes cross-layer similarity over a development set, allowing easy deployment across models. The method incorporates efficient implementation constraints (e.g. tile-level operations), across both prefill and decode attention. The Top-k selection and reuse in Kascade is head-aware and we show in our experiments that this is critical for high accuracy. Kascade achieves up to 4.1x speedup in decode attention and 2.2x speedup in prefill attention over FlashAttention-3 baseline on H100 GPUs while closely matching dense attention accuracy on long-context benchmarks such as LongBench and AIME-24.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8timesA800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to 2.8times and 2.0times in the prefill and decode stages.

Long streaming video QA remains challenging due to growing visual tokens and limited reasoning length of large language models (LLMs). KV-caching stores the Key-Value (KV) of the historical tokens via LLM prefill and enables more efficient streaming QA. However, existing methods cache every one or two frames, causing redundant memory usage and losing fine-grained spatial details within frame or temporal contexts across frames. This paper proposes MuKV, a method that features a multi-grained KV cache compression module and a semi-hierarchical retrieval approach to improve both efficiency and accuracy for long streaming VideoQA. For the offline KV cache, MuKV extracts visual representations at patch-, frame-, and segment-levels. The multiple levels of granularity preserve both local cues and global temporal context, while maintaining efficiency with a dual signal token compression mechanism guided by self-attention and frequency. For online QA, MuKV designs a semi-hierarchical retrieval method to retrieve relevant KV caches for answer generation. Experiments on long-streaming VideoQA benchmarks show that MuKV significantly improves answer accuracy, without sacrificing memory and online QA efficiency. Moreover, our compression mechanism alone brings consistent benefits across answer accuracy, memory, and QA efficiency over baselines, showcasing highly effective contribution.

In this paper, we introduce Writing in the Margins (WiM), a new inference pattern for Large Language Models designed to optimize the handling of long input sequences in retrieval-oriented tasks. This approach leverages the chunked prefill of the key-value cache to perform segment-wise inference, which enables efficient processing of extensive contexts along with the generation and classification of intermediate information ("margins") that guide the model towards specific tasks. This method increases computational overhead marginally while significantly enhancing the performance of off-the-shelf models without the need for fine-tuning. Specifically, we observe that WiM provides an average enhancement of 7.5% in accuracy for reasoning skills (HotpotQA, MultiHop-RAG) and more than a 30.0% increase in the F1-score for aggregation tasks (CWE). Additionally, we show how the proposed pattern fits into an interactive retrieval design that provides end-users with ongoing updates about the progress of context processing, and pinpoints the integration of relevant information into the final response. We release our implementation of WiM using Hugging Face Transformers library at https://github.com/writer/writing-in-the-margins.

We present context parallelism for long-context large language model inference, which achieves near-linear scaling for long-context prefill latency with up to 128 H100 GPUs across 16 nodes. Particularly, our method achieves 1M context prefill with Llama3 405B model in 77s (93% parallelization efficiency, 63% FLOPS utilization) and 128K context prefill in 3.8s. We develop two lossless exact ring attention variants: pass-KV and pass-Q to cover a wide range of use cases with the state-of-the-art performance: full prefill, persistent KV prefill and decode. Benchmarks on H100 GPU hosts inter-connected with RDMA and TCP both show similar scalability for long-context prefill, demonstrating that our method scales well using common commercial data center with medium-to-low inter-host bandwidth.

The promising applications of large language models are often limited by the constrained GPU memory capacity available on edge devices. Mixture-of-Experts (MoE) models help address this issue by activating only a subset of the model's parameters during computation. This approach allows the unused parameters to be offloaded to host memory, thereby reducing the overall GPU memory demand. However, existing cache-based offloading solutions handle cache misses reactively, which significantly impacts system performance. In this paper, we introduce ProMoE, a novel proactive caching system that utilizes intermediate results to predict subsequent expert usage. By proactively fetching experts in advance, ProMoE eliminates passive cache misses, removes loading time from the critical path, and reduces the performance overhead associated with offloading. Our evaluations demonstrate that ProMoE achieves an average speedup of 2.20x (up to 3.21x) and 2.07x (up to 5.02x) in the prefill and decode stages, respectively, compared to existing offloading solutions.

We present LFM2, a family of Liquid Foundation Models designed for efficient on-device deployment and strong task capabilities. Using hardware-in-the-loop architecture search under edge latency and memory constraints, we obtain a compact hybrid backbone that combines gated short convolutions with a small number of grouped query attention blocks, delivering up to 2x faster prefill and decode on CPUs compared to similarly sized models. The LFM2 family covers 350M-8.3B parameters, including dense models (350M, 700M, 1.2B, 2.6B) and a mixture-of-experts variant (8.3B total, 1.5B active), all with 32K context length. LFM2's training pipeline includes a tempered, decoupled Top-K knowledge distillation objective that avoids support mismatch; curriculum learning with difficulty-ordered data; and a three-stage post-training recipe of supervised fine-tuning, length-normalized preference optimization, and model merging. Pre-trained on 10-12T tokens, LFM2 models achieve strong results across diverse benchmarks; for example, LFM2-2.6B reaches 79.56% on IFEval and 82.41% on GSM8K. We further build multimodal and retrieval variants: LFM2-VL for vision-language tasks, LFM2-Audio for speech, and LFM2-ColBERT for retrieval. LFM2-VL supports tunable accuracy-latency tradeoffs via token-efficient visual processing, while LFM2-Audio separates audio input and output pathways to enable real-time speech-to-speech interaction competitive with models 3x larger. LFM2-ColBERT provides a low-latency encoder for queries and documents, enabling high-performance retrieval across multiple languages. All models are released with open weights and deployment packages for ExecuTorch, llama.cpp, and vLLM, making LFM2 a practical base for edge applications that need fast, memory-efficient inference and strong task capabilities.

The performance of multi-turn, agentic LLM inference is increasingly dominated by KV-Cache storage I/O rather than computation. In prevalent disaggregated architectures, loading the massive KV-Cache from external storage creates a fundamental imbalance: storage NICs on prefill engines become bandwidth-saturated, while those on decoding engines remain idle. This asymmetry severely constrains overall system throughput. We present DualPath, an inference system that breaks this bottleneck by introducing dual-path KV-Cache loading. Beyond the traditional storage-to-prefill path, DualPath enables a novel storage-to-decode path, in which the KV-Cache is loaded into decoding engines and then efficiently transferred to prefill engines via RDMA over the compute network. DualPath combines this optimized data path -- which inherently avoids network congestion and avoids interference with latency-critical model execution communications -- with a global scheduler that dynamically balances load across prefill and decode engines. Our evaluation on three models with production agentic workloads demonstrates that DualPath improves offline inference throughput by up to 1.87times on our in-house inference system. It can also improve online serving throughput by an average factor of 1.96times without violating SLO.

Retrieval-augmented generation (RAG) improves large language model (LLM) answer quality by grounding generation in external evidence, but processing retrieved contexts makes the prefill stage a dominant serving cost. RAG cache fusion reduces this cost by reusing precomputed key-value (KV) caches for retrieved chunks and selectively recomputing tokens under the current prompt. Existing selectors, however, face a dilemma between quality and efficiency: fast query-agnostic or final-layer query-to-context selectors can miss request-relevant evidence, whereas full-view query-aware selectors require broad context and layer visibility before recomputation and therefore stall the layer-wise cache-fusion pipeline. We present QCFuse, a compressed-view query-aware selector for RAG cache fusion. QCFuse uses chunk-anchor query probing to condition user-query states on compact per-chunk anchors and critical-layer profiling to identify recomputation tokens without all-layer inspection. We implement QCFuse in SGLang and evaluate it on four open-weight LLMs across six datasets. QCFuse reaches full-prefill-level quality. At matched quality, QCFuse achieves an average prefill-time speedup of 1.7x over full prefill and 1.5x over ProphetKV, the strongest quality-preserving baseline.

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.

In modern large language models (LLMs), increasing sequence lengths is a crucial challenge for enhancing their comprehension and coherence in handling complex tasks such as multi-modal question answering. However, handling long context sequences with LLMs is prohibitively costly due to the conventional attention mechanism's quadratic time and space complexity, and the context window size is limited by the GPU memory. Although recent works have proposed linear and sparse attention mechanisms to address this issue, their real-world applicability is often limited by the need to re-train pre-trained models. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which simultaneously reduces the training and inference time complexity from O(T^2) to O(T log T) and the space complexity from O(T^2) to O(T). To this end, we devise a dynamic sparse attention mechanism that generates an attention mask through a novel tree-search-like algorithm for a given query on the fly. HiP is training-free as it only utilizes the pre-trained attention scores to spot the positions of the top-k most significant elements for each query. Moreover, it ensures that no token is overlooked, unlike the sliding window-based sub-quadratic attention methods, such as StreamingLLM. Extensive experiments on diverse real-world benchmarks demonstrate that HiP significantly reduces prompt (i.e., prefill) and decoding latency and memory usage while maintaining high generation performance with little or no degradation. As HiP allows pretrained LLMs to scale to millions of tokens on commodity GPUs with no additional engineering due to its easy plug-and-play deployment, we believe that our work will have a large practical impact, opening up the possibility to many long-context LLM applications previously infeasible.

Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4times fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers 1.8-3.8times faster prefill and 2.2-3.4times faster decode than the dense baseline MobileLLM-Pro.

Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.

Hybrid sequence models that combine efficient Transformer components with linear sequence modeling blocks are a promising alternative to pure Transformers, but most are still pretrained from scratch and therefore fail to reuse existing Transformer checkpoints. We study upcycling as a practical path to convert pretrained Transformer LLMs into hybrid architectures while preserving short-context quality and improving long-context capability. We call our solution HyLo (HYbrid LOng-context): a long-context upcycling recipe that combines architectural adaptation with efficient Transformer blocks, Multi-Head Latent Attention (MLA), and linear blocks (Mamba2 or Gated DeltaNet), together with staged long-context training and teacher-guided distillation for stable optimization. HyLo extends usable context length by up to 32times through efficient post-training and reduces KV-cache memory by more than 90%, enabling up to 2M-token prefill and decoding in our vLLM inference stack, while comparable Llama baselines run out of memory beyond 64K context. Across 1B- and 3B-scale settings (Llama- and Qwen-based variants), HyLo delivers consistently strong short- and long-context performance and significantly outperforms state-of-the-art upcycled hybrid baselines on long-context evaluations such as RULER. Notably, at similar scale, HyLo-Qwen-1.7B trained on only 10B tokens significantly outperforms JetNemotron (trained on 400B tokens) on GSM8K, Lm-Harness common sense reasoning and RULER-64K.

In retrieval-augmented generation and agentic LLM serving, prompts are assembled from independent segments into long contexts, making the prefill stage dominate per-request cost. Two directions have emerged to reduce this cost: position-independent caching (PIC) admits KV reuse for non-contiguous segments shared across requests, while hybrid-attention models cut computation by replacing most full-attention layers with linear attention. However, they cannot coexist: applying existing PIC methods to hybrid-attention models breaks down because per-token KV-cache reuse primitives do not transfer to the per-request recurrent state. We present Hypic, the first system to accelerate hybrid-attention LLM serving with position-independent caching. For linear-attention layers, we identify the segment-cumulative transition operator as the missing algebraic primitive and cache it alongside each segment's zero-start end-state, enabling near-exact and constant-time composition of independently cached segments. For the remaining full-attention layers, existing PIC methods also fail because linear layers do not expose the per-token hidden states needed for selective recomputation. We show that the largest deviations concentrate at segment beginnings and construct a small seam window that propagates hidden states through the hybrid-attention stack to repair cross-segment attention. Finally, Hypic introduces segment parallelism, which exploits PIC's segment-level self-containment to parallelize cache-miss prefill across instances, turning long cold requests into an accelerable workload. Evaluated across four hybrid-attention models and five workloads, Hypic reduces time-to-first-token by 3.25times on average and improves QPS by 1.66times over Prefix Cache, while preserving task quality with a 1.71-point gap from Full Recompute.

Smaller activation outliers do not necessarily imply better low-bit quantization: their alignment with the quantizer matters. We introduce PrismQuant, a quantizer-aware rotation framework that aligns the leading activation eigenspace with the constant group subspace of asymmetric grouped INT4. The affine offsets represent the energy in this subspace without widening the range within the group. We formulate rotation design as a Ky Fan trace maximization and derive a closed-form solution that is provably optimal for this alignment objective. Compact Householder transformations and their compact-WY representation enable gradient-free construction and efficient application at both foldable and online sites. A predictive range law further connects unaligned activation energy and group size to quantization-relevant variation. Experiments on Llama, Qwen, and Mistral span dense models up to 70B parameters and a 30B mixture-of-experts model. Under W4A4KV4, PrismQuant sets the state of the art on Llama-3.2-3B among the compared methods in both perplexity and accuracy. On Llama-3.1-70B, it attains 3.85 perplexity and 72.46% average zero-shot accuracy, only 0.22 percentage points below full precision. In the deployment study on Llama-3.1-8B, our optimized implementation achieves 1.51x prefill and 1.22x CUDA Graph decode speedups over matched FP16 baselines, with 56.34% lower decode peak memory and only 2.35% additional Graph decode latency over Hadamard. Code is available at https://github.com/ForeverBlue816/PrismQuant.

Structured pruning reduces LLM inference costs by removing model components to produce smaller dense models. However, existing methods can cause reasoning performance to collapse on long-form reasoning tasks such as AIME. We propose LayerTrim, a training-free structured pruning method that identifies redundant decoder layers and selectively trims their MLP features while preserving attention and decoder depth. LayerTrim is the first training-free structured pruning method designed to preserve long-form reasoning without gradients or recovery training. Across three reasoning language models and five benchmarks, LayerTrim outperforms existing structured pruning methods in most evaluated settings, improving average accuracy by up to 21.4 percentage points over RESP. On Qwen3-14B, LayerTrim achieves 1.24times faster prefill and 19.7\% lower peak GPU memory than the dense model, while better preserving reasoning ability. It also achieves the shortest average generation length among evaluated pruned models, while retaining the highest average accuracy. These results demonstrate that LayerTrim improves reasoning efficiency through both structural compression and reduced token generation.

The deployment of long-context Large Language Models (LLMs) poses significant challenges due to the intense computational cost of self-attention and the substantial memory overhead of the Key-Value Cache (KV Cache). In this paper, we introduce HieraSparse, a hierarchical KV Cache compression framework with acceleration kernels that leverage GPU sparse tensor cores to speed up semi-structured KV Cache attention for both the prefill and decode phases. With the hierarchical design, our method allows for a flexible quality-sparsity trade-off and successfully converts sparsity into efficiency. Compared to the state-of-the-art decode method that utilizes unstructured sparsity, HieraSparse achieves 1.2times KV compression ratio and 4.57times attention speedup at the same sparsity level. Furthermore, we extended the semi-structured KV Cache pruning to the prefill stage, which demonstrated up to 1.85times attention speedup at the highest sparsity. Lastly, we evaluate the generation quality of HieraSparse with a simple magnitude-based pruning method, and the results show that 1.37times prefill speedup and 1.77times decode speedup can be achieved without significant quality drop. The codebase can be found at https://github.com/psl-ntu/HieraSparse.

As large language models (LLMs) evolve to handle increasingly longer contexts, serving inference requests for context lengths in the range of millions of tokens presents unique challenges. While existing techniques are effective for training, they fail to address the unique challenges of inference, such as varying prefill and decode phases and their associated latency constraints - like Time to First Token (TTFT) and Time Between Tokens (TBT). Furthermore, there are no long context inference solutions that allow batching requests to increase the hardware utilization today. In this paper, we propose three key innovations for efficient interactive long context LLM inference, without resorting to any approximation: adaptive chunking to reduce prefill overheads in mixed batching, Sequence Pipeline Parallelism (SPP) to lower TTFT, and KV Cache Parallelism (KVP) to minimize TBT. These contributions are combined into a 3D parallelism strategy, enabling Mnemosyne to scale interactive inference to context lengths at least up to 10 million tokens with high throughput enabled with batching. To our knowledge, Mnemosyne is the first to be able to achieve support for 10 million long context inference efficiently, while satisfying production-grade SLOs on TBT (30ms) on contexts up to and including 10 million.

Recently, quantization has been widely used for the compression and acceleration of large language models~(LLMs). Due to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with the equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still remain steep and outspread. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach to enhance flatness of weights and activations. Our approach identifies optimal affine transformations tailored to each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead, we apply Kronecker decomposition to the transformation matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments show that FlatQuant sets up a new state-of-the-art quantization benchmark. For instance, it achieves less than 1% accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by 7.5%. For inference latency, FlatQuant reduces the slowdown induced by pre-quantization transformation from 0.26x of QuaRot to merely 0.07x, bringing up to 2.3x speedup for prefill and 1.7x speedup for decoding, respectively. Code is available at: https://github.com/ruikangliu/FlatQuant.

Long-context LLM inference is bottlenecked by the quadratic attention complexity and linear KV cache growth. Prior approaches mitigate this via post-hoc selection or eviction but overlook the root inefficiency: indiscriminate writing to memory. In this paper, we formalize KV cache management as a causal system of three primitives: KV Admission, Selection, and Eviction. We instantiate KV Admission via Write-Gated KV (WG-KV), a lightweight mechanism that learns to predict token utility before cache entry. By filtering out low-utility states early to maintain a compact global cache alongside a sliding local cache, WG-KV reduces memory usage by 46-68% and delivers 3.03-3.70x prefill and 1.85-2.56x decode speedups on Llama and Qwen models, while maintaining compatibility with FlashAttention and Paged-KV systems. These results demonstrate that learning what to write is a principled and practical recipe for efficient long-context inference. Code is available at https://github.com/EMCLab-Sinica/WG-KV.

Deploying Large Language Models (LLMs) on resource-constrained edge devices faces critical bottlenecks in memory bandwidth and power consumption. While ternary quantization (e.g., BitNet b1.58) significantly reduces model size, its direct deployment on general-purpose hardware is hindered by workload imbalance, bandwidth-bound decoding, and strict data dependencies. To address these challenges, we propose VitaLLM, a hardware-software co-designed accelerator tailored for efficient ternary LLM inference. We introduce a heterogeneous Dual-Core Compute Strategy that synergizes specialized TINT-Cores for massive ternary projections with a unified BoothFlex-Core for mixed-precision attention, ensuring high utilization across both compute-bound prefill and bandwidth-bound decode stages. Furthermore, we develop a Leading One Prediction (LOP) mechanism to prune redundant Key-Value (KV) cache fetches and a Dependency-Aware Scheduling framework to hide the latency of nonlinear operations. Implemented in TSMC 16nm technology, VitaLLM achieves a decoding throughput of 70.70 tokens/s within an ultra-compact area of 0.223 mm^2 and a power consumption of 65.97 mW. The design delivers a superior Figure of Merit (FOM) of 17.4 TOPS/mm^2/W, significantly outperforming state-of-the-art accelerators. Finally, we explore an extended bit-serial design (BoothFlex-BS) to demonstrate the architecture's adaptability for precision-agile inference.

We present Key-Value Means ("KVM"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong O(N) chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth. KVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and linear RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between O(N) and O(N^2). It can also be implemented in a hybrid solution in tandem with LRNN layers in place of traditional attention, to supplement the LRNN with improved sublinear memory growth context length usage and long context decoding. We release our code at https://github.com/recursal/KVM-paper and trained models at https://huggingface.co/collections/recursal/key-value-means under the Apache 2.0 license.

Multi-document retrieval-augmented generation (RAG) requires a language model to process multiple retrieved text chunks before answering a question. Precomputing each chunk's KV cache independently and concatenating the caches when the chunks are retrieved can accelerate this step. However, the assembled cache lacks cross-chunk attention information, reducing answer quality. Selective recomputation methods recover the missing cross-chunk context by rerunning the target LLM on selected tokens, incurring substantial online computation. We introduce CacheRepair, a lightweight network that learns the difference between independently computed KV caches and those produced by processing the chunks together. The network combines compressed KV features with token embeddings and uses attention that is bidirectional within each chunk and flows from earlier to later chunks. Each repair block receives the compressed cache features, and the predicted residual is added to every document token's cache. Each repair network is trained for a specific frozen target LLM on a generic retrieval corpus and reused across downstream datasets. Our analysis shows that repair reduces KV errors both near chunk boundaries and throughout chunk interiors. Evaluation across three target LLMs and four downstream datasets places CacheRepair on the measured answer-quality-latency Pareto frontier in eleven of twelve model-dataset combinations. Reported time to first token (TTFT) includes online cache transfer and repair. Across all twelve combinations, the largest repairers achieve 1.69-4.61times speedups in median TTFT over full prefill and improve mean F1 by 2.1-26.1 percentage points over direct cache reuse.

As telecommunication networks evolve toward autonomous 5G-Advanced and 6G operations, agentic artificial intelligence (AI) workflows, where large language models (LLMs) execute multi-step reasoning, invoke diagnostic tools, retrieve domain knowledge, and coordinate across agent teams, are increasingly embedded across the edge-cloud continuum. While the biological brain accomplishes complex cognition on an exceptionally modest metabolic power budget of approximately 20W contemporary LLMs are profoundly energy- and memory-intensive, making sustainable lifecycle orchestration a critical operational priority. However, existing AI lifecycle metrics evaluate only isolated, single-model inferences or overlook multi-agent execution graphs entirely. Consequently, network operators lack foundational models to determine whether distributed agent communication incurs meaningful energy costs and where across edge-cloud tiers agent teams should physically reside. To address this gap, we introduce agentic-eCAL, generalizing the Energy Cost of AI Lifecycle (eCAL) metric to directed multi-agent workflows by coupling a closed-form two-rate single-call energy model (compute-bound prefill and memory-bound decode) with 7-layer OSI data transport. Grounded in hundreds of GPU benchmark configurations on NVIDIA A100 and H100, 16 open-weight models and 8 orchestration topologies, we validate components of the metric and study workflow placement implications. Our findings demonstrate that inter-agent text transport incurs 0.25% of workflow energy across 5G RAN, metro, and optical links. Therefore in edge-cloud agent placement the dominant energy cost of distribution is often not the transmission of inter-agent text itself, but the additional inference and context processing induced by that communication.

Rotary Position Embedding (RoPE) encodes token positions by rotating each two-dimensional channel of the query and key vectors at a channel-specific frequency, making the attention logits invariant to a common shift of positions. However, this rotation is periodic, and it leads to position aliasing where relative positions separated by a full rotation period become hard to tell apart. To address this, we propose WavePrune, which restricts each channel to its first rotation period. We show that it removes the distractions in attention maps created by position aliasing and improves overall long-context performance. Specifically, WavePrune raises the HELMET score on four of five models we test without any extra tuning (e.g., 35.7 -> 40.0 on Qwen3-8B). When pretraining models from scratch, WavePrune also achieves lower validation loss at extrapolated lengths than pretraining without it. Because WavePrune restricts each channel to a sliding window, it induces a fine-grained sparsity that our hardware-aligned CUDA kernels exploit for 1.15x prefill and 1.24x decoding speedups over FlashAttention-2 at 32K context. Together, these results show that RoPE's periodic structure, widely regarded as essential, is largely redundant beyond the first rotation period.

Long-context LLM inference is bottlenecked by quadratic attention computation and growing KV-cache costs. Existing sparse attention and KV-compression methods typically decide which tokens or heads to preserve from runtime attention scores, observation windows, calibration prompts, or learned gates, making head diagnosis input-dependent and costly to deploy. We propose Autonomy-of-Heads (AoH), a data-free method that identifies retrieval and streaming heads from the spectral geometry of query-key projections. AoH defines the kernel attention operator M_h = W_K^{htop}W_Q^h and uses its effective-rank as a weight-space measure of head function: concentrated spectra indicate a small number of dominant query-key matching directions and are associated with retrieval heads, whereas diffuse spectra indicate the absence of a dominant global matching direction and are associated with streaming heads. We further derive an efficient d_head-dimensional computation that avoids constructing the full d_modeltimes d_model matrix. We conducted extensive experiments across models demonstrating that at 50\% sparsity, AoH retains 96.5\% of Full Attention performance on average while reducing prefill and decode latency by up to 41.4\% and 66.0\%, respectively, and KV-cache memory by 50.0\% at 256K tokens.

Context retrieval systems for LLM inference face a critical challenge: high retrieval latency creates a fundamental tension between waiting for complete context (poor time-to-first-token) and proceeding without it (reduced quality). Streaming context incrementally--overlapping retrieval with inference--can mitigate this latency, but doing so with concurrent requests introduces new challenges: requests contend for GPU compute and memory, and scheduling must adapt to dynamic context arrivals. We present Stream2LLM, a streaming-aware LLM serving system for concurrent prefill-decode disaggregated deployments. Stream2LLM introduces adaptive scheduling and preemption for two distinct retrieval patterns: append-mode (progressive context accumulation) and update-mode (iterative refinement with cache invalidation). It decouples scheduling decisions from resource acquisition, enabling flexible preemption strategies guided by hardware-specific cost models, and uses longest common prefix matching to minimize redundant computation when input changes dynamically. To evaluate Stream2LLM, we collect two large-scale, real-world streaming workloads based on web crawling and approximate nearest neighbor search. Our evaluation demonstrates that streaming architecture delivers up to 11x TTFT improvements, with cost-aware scheduling providing critical benefits under memory pressure, all while maintaining throughput parity with non-streaming baselines. Code: https://github.com/rajveerb/stream2llm/tree/mlsys_artifact

As large language models (LLMs) are increasingly deployed with highly heterogeneous workloads, chunked-prefill execution has emerged as a mainstream serving architecture. Balancing scheduling fairness and latency stability in such environments is critical; otherwise, severe head-of-line blocking and request starvation will degrade user experience. However, existing systems rely on rigid First-Come, First-Served (FCFS) policies and static token budgets, leading to fairness degradation and unpredictable latency jitter. To address these issues, we propose a fairness-aware and latency-controllable scheduling framework for chunked-prefill LLM engines. Specifically, we design a lightweight aging-based scheduling policy that dynamically calculates priorities using accumulated waiting time and remaining prefill work. Furthermore, we develop Latency-Prediction-Based Request Scheduling (LPRS) and Active Prefill Control (APC) to replace static budgets with target-time constraints and actively regulate prefill concurrency. We evaluated our scheduling framework on NVIDIA GPUs and Ascend accelerators using real-world workloads. Results show the aging policy reduces mean end-to-end latency by over 10\% compared to FCFS. Moreover, LPRS and APC significantly reduce P99 tail latency and suppress prefill fragmentation, confirming that the structural prefill control and the temporal latency constraints are fundamentally complementary. All codes have been released in Github.

Transformers operate as horizontal token-by-token scanners; at each generation step, attending to an ever-growing sequence of token-level states. This access pattern increases prefill latency and makes long-context decoding more memory-bound, as KV-cache reads and writes dominate inference time over arithmetic operations. We propose Parallel Hierarchical Operation for TOp-down Networks (PHOTON), a hierarchical autoregressive model that replaces horizontal scanning with vertical, multi-resolution context scanning. PHOTON maintains a hierarchy of latent streams: a bottom-up encoder compresses tokens into low-rate contextual states, while lightweight top-down decoders reconstruct fine-grained token representations in parallel. We further introduce recursive generation that updates only the coarsest latent stream and eliminates bottom-up re-encoding. Experimental results show that PHOTON is superior to competitive Transformer-based language models regarding the throughput-quality trade-off, providing advantages in long-context and multi-query tasks. In particular, this reduces decode-time KV-cache traffic, yielding up to 10^{3}times higher throughput per unit memory.

A common small-model deployment runs one shared backbone with several LoRA specialists that answer over the same context. Serving them naively re-prefills that shared context once per specialist. We study a narrow, practical question: for already-trained standard LoRA adapters -- not adapters retrained for cache compatibility -- how much task quality is preserved if the backbone's prefill KV cache is computed once and reused across specialists, and what does that buy in serving cost? On a Qwen3-1.7B backbone with two adapters (extractive QA on HotpotQA, arithmetic reasoning on GSM8K), we sweep the boundary at which the specialist takes over from the reused base cache and measure paired quality differences and serving cost. Full-prefix reuse had the lowest prefill cost and a small quality difference on held-out GSM8K (Delta = -4.6 EM at a 160-token budget; -3.0 at 320 tokens; -0.8 under a second training seed -- all favoring native, only the first excluding zero, and the magnitude not consistent). Partial recomputation provided no demonstrated advantage. Neither quality equivalence nor a general boundary-selection rule is established. We also report a closed-form ridge KV translator that did not beat direct reuse, and specialist-dependence contrasts whose intervals all include zero. The measured serving benefit is warm-cache time-to-first-token, which grows with context (~16x at 8K); two-branch peak memory was only 12% lower and, on inspection, the prefix was never physically shared across branches -- this implementation reuses KV values but copies their storage, so shared-cache memory savings are not achieved.

The Mixture of Experts (MoE) architecture has demonstrated significant advantages as it enables to increase the model capacity without a proportional increase in computation. However, the large MoE model size still introduces substantial memory demands, which usually requires expert offloading on resource-constrained platforms and incurs significant overhead. Hybrid CPU-GPU inference has been proposed to leverage CPU computation to reduce expert loading overhead but faces major challenges: on one hand, the expert activation patterns of MoE models are highly unstable, rendering the fixed mapping strategies in existing works inefficient; on the other hand, the hybrid CPU-GPU schedule for MoE is inherently complex due to the diverse expert sizes, structures, uneven workload distribution, etc. To address these challenges, in this paper, we propose HybriMoE, a hybrid CPU-GPU inference framework that improves resource utilization through a novel CPU-GPU scheduling and cache management system. HybriMoE introduces (i) a dynamic intra-layer scheduling strategy to balance workloads across CPU and GPU, (ii) an impact-driven inter-layer prefetching algorithm, and (iii) a score-based caching algorithm to mitigate expert activation instability. We implement HybriMoE on top of the kTransformers framework and evaluate it on three widely used MoE-based LLMs. Experimental results demonstrate that HybriMoE achieves an average speedup of 1.33times in the prefill stage and 1.70times in the decode stage compared to state-of-the-art hybrid MoE inference framework. Our code is available at: https://github.com/PKU-SEC-Lab/HybriMoE.

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.

The rapid growth of visual tokens in multimodal large language models (MLLMs) leads to excessive memory consumption and inference latency, especially when handling high-resolution images and videos. Token pruning is a technique used to mitigate this issue by removing redundancy, but existing methods often ignore relevance to the user query or suffer from the limitations of attention mechanisms, reducing their adaptability and effectiveness. To address these challenges, we propose Script, a plug-and-play pruning method that requires no retraining and generalizes across diverse MLLMs. Script comprises two modules: a graph-structured pruning module that removes visually redundant tokens, and a query-conditioned semantic pruning module that preserves query-relevant visual information. Together, they enhance performance on multimodal tasks. Experiments on fourteen benchmarks across image and video understanding tasks show that Script consistently achieves higher model efficiency and predictive accuracy compared to existing pruning methods. On LLaVA-NeXT-7B, it achieves up to 6.8x prefill speedup and 10x FLOP reduction, while retaining 96.88% of the original performance.

The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs that resolves both computation and memory bottlenecks through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform attention mechanism solely on those important tokens to accelerate the prefill phase. To mitigate the memory bottleneck in the decoding phase, we employ mixed-precision quantization to the KV cache, where high-bit quantization is used for caches of important tokens, while low-bit quantization is applied to those of less importance. Our experiments demonstrate that ZipVL can accelerate the prefill phase by 2.6times and reduce GPU memory usage by 50.0%, with a minimal accuracy reduction of only 0.2% on Video-MME benchmark over LongVA-7B model, effectively enhancing the generation efficiency of LVLMs.

Existing large language model (LLM) serving systems fall into two categories: 1) a unified system where prefill phase and decode phase are co-located on the same GPU, sharing the unified computational resource and storage, and 2) a disaggregated system where the two phases are disaggregated to different GPUs. The design of the disaggregated system addresses the latency interference and sophisticated scheduling issues in the unified system but leads to storage challenges including 1) replicated weights for both phases that prevent flexible deployment, 2) KV cache transfer overhead between the two phases, 3) storage imbalance that causes substantial wasted space of the GPU capacity, and 4) suboptimal resource adjustment arising from the difficulties in migrating KV cache. Such storage inefficiency delivers poor serving performance under high request rates. In this paper, we identify that the advantage of the disaggregated system lies in the disaggregated computation, i.e., partitioning the computational resource to enable the asynchronous computation of two phases. Thus, we propose a novel LLM serving system, semi-PD, characterized by disaggregated computation and unified storage. In semi-PD, we introduce a computation resource controller to achieve disaggregated computation at the streaming multi-processor (SM) level, and a unified memory manager to manage the asynchronous memory access from both phases. semi-PD has a low-overhead resource adjustment mechanism between the two phases, and a service-level objective (SLO) aware dynamic partitioning algorithm to optimize the SLO attainment. Compared to state-of-the-art systems, semi-PD maintains lower latency at higher request rates, reducing the average end-to-end latency per request by 1.27-2.58x on DeepSeek series models, and serves 1.55-1.72x more requests adhering to latency constraints on Llama series models.

Self-attention gives LLMs fine-grained, query-dependent access to context, but dense token interactions incur quadratic prefill cost and a key--value cache growing with context length. Research thus spans explicit-memory compression, sparse access, recurrent state construction, structured state dynamics, and heterogeneous mechanism composition. This survey analyzes these developments as model-internal contextual memory. We introduce a five-dimensional lens---Memory Representation, Memory Update, Access, Readout, and Integration---describing what is represented, how it changes, what is query-eligible, how it is read, and how readouts form outputs. This lens compares overlapping research lines without imposing one computational model. We reconstruct mechanism-level developments and architectural adoption using 59 release-level records from 14 major model lineages and 11 high-performing open-weight endpoints. First, explicit-memory and recurrent-state methods retain distinct interfaces but increasingly control overlapping memory functions. Second, heterogeneous architectures increasingly coordinate across network depth: layer-wise composition distributes complementary memory processing across representational stages, while cross-layer reuse carries selected memory and routing artifacts forward. Depth thus becomes a dimension along which contextual memory is constructed and managed. Third, these developments motivate a stateful multidimensional memory-routing hypothesis: persistent memory is organized across temporal scope, network depth, substrate type, and representation granularity, while coordinated Sparse Write and Sparse Read determine what is maintained and what contributes to each query. Overall, efficient sequence architecture design increasingly concerns the organization, lifecycle, and selective use of contextual memory rather than an isolated Attention operator.

We study secret elicitation: discovering knowledge that an AI possesses but does not explicitly verbalize. As a testbed, we train three families of large language models (LLMs) to possess specific knowledge that they apply downstream but deny knowing when asked directly. For example, in one setting, we train an LLM to generate replies that are consistent with knowing the user is female, while denying this knowledge when asked directly. We then design various black-box and white-box secret elicitation techniques and evaluate them based on whether they can help an LLM auditor successfully guess the secret knowledge. Many of our techniques improve on simple baselines. Our most effective techniques (performing best in 2/3 settings) are based on prefill attacks, a black-box technique where the LLM reveals secret knowledge when generating a completion from a predefined prefix. In our remaining setting, white-box techniques based on logit lens and sparse autoencoders (SAEs) are most effective. We release our models and code, establishing a public benchmark for evaluating secret elicitation methods.

KV cache has traditionally been stored in GPU memory to accelerate the decoding phase of large language model (LLM) inference. However, it is increasingly necessary to move KV caches outside GPU devices, to enable cache reuse across different queries and inference engines. Our real-world usage statistics confirm this trend: over time, the total KV cache stored by users has grown rapidly, far exceeding the capacity of GPU memory. Despite this need, there lacks an efficient solution for offloading and transferring KV caches. We present LMCACHE, the first and so far the most efficient open-source KV caching solution, which extracts and stores KV caches generated by modern LLM engines (vLLM and SGLang) out of the GPU memory and shares them across engines and queries. LMCACHE supports both cache offloading (prefix reuse across queries) and prefill-decode (PD) disaggregation (cross-engine/GPU cache transfer). LMCACHE's high performance and wide adoption stem from the following contributions: (1) highly optimized KV cache data movement powered by batched data movement operations, compute and I/O pipelining; (2) a modular KV cache connector component, decoupling LMCACHE from the rapid evolution of inference engines; (3) a first-class control API for flexible cache orchestration across GPU, CPU, storage, and network layers. Our evaluation shows that combining LMCACHE with vLLM achieves up to 15x improvement in throughput across workloads such as multi-round question answering and document analysis. Large-scale adoption of LMCACHE in enterprise settings provides us valuable insights, for example, fetching KV cache from remote storage has unsurprisingly benefits to prefill delay, and that context truncation, which is a widely applied technique in industry, can greatly reduce prefix cache hit ratio by half. The source code of LMCACHE is at: https://github.com/LMCache/LMCache.

Visual token reduction is an effective way to accelerate multimodal large language models (MLLMs), but performance deteriorates rapidly under extremely low token budgets. Existing work has explored both visual-token selection and training-based adaptation to reduced visual inputs. We take a step further by asking how a heavily compressed MLLM should learn from the states induced by its own generations. This setting naturally calls for on-policy self-distillation: a heavily compressed model is supervised on the states induced by its own generations, while its full-token counterpart serves as an information-rich teacher. Based on this insight, we propose LT-OPD, a training framework for extreme visual-token reduction. The student rolls out responses with only a small fraction of visual tokens, and a frozen full-token copy of the same MLLM provides distributional supervision along these student-generated trajectories. To stabilize on-policy learning when visual evidence is severely limited, we further introduce a budget-level curriculum that progressively decreases the token budget during training. Across nine benchmarks on Qwen3.5-4B, LT-OPD raises average retained performance under 5% visual-token retention from 68.6% to 82.3%, outperforming training-free, training-based, and reinforcement-learning baselines at the same budget. The gains transfer consistently to Qwen3.5-9B, GLM-4.6V-9B, and LLaVA-OV-1.5-4B. LT-OPD also reduces KV-cache usage by 85.2% and prefill FLOPs by 85.4% without additional inference overhead, demonstrating that on-policy learning can substantially recover capabilities lost to extreme visual-token reduction.

Large Multimodal Models (LMMs) extend Large Language Models (LLMs) by handling diverse inputs such as images, audio, and video, but at the cost of adding a multimodal encoding stage that increases both computational and memory overhead. This step negatively affects key Service Level Objectives (SLOs), such as time to first token (TTFT) and time per output token (TPOT). We introduce Encode-Prefill-Decode (EPD) Disaggregation, a novel framework that separates the encoding, prefill, and decode stages onto dedicated resources. Unlike current systems, which bundle encoding and prefill together, our approach decouples these steps, unlocking new opportunities and optimizations. These include a mechanism to cache multimedia tokens for efficient transfer, a novel way to parallelize the encoding load within a request, a module for optimal resource allocation for disaggregated serving, and a novel role-switching method to handle changing workload characteristics. Experimental evaluations with popular LMMs show substantial gains in memory efficiency (up to 15x lower peak memory utilization), batch sizes (up to 22x larger), 10x more images per request, and 2.2x larger KV caches. Furthermore, it leads to significant improvements in SLO attainment (up to 90-100% improvement) and TTFT (up to 71% reduction), compared to systems that do not disaggregate. The code is available at https://github.com/vbdi/epdserve.

State Space Models (SSMs) compress sequence history into a bounded recurrent state, making the resulting memory law a central architectural choice for long-context performance. Most modern SSMs rely on ODE-based dynamics that lead to exponential forgetting, limiting their ability to retain information over broad temporal ranges. We introduce FRAC, a selective SSM architecture derived from fractional dynamics that replaces this exponential decay with power-law long memory. To make fractional dynamics practical, FRAC approximates the heavy-tailed target kernel with a finite-state, log-spaced sum of exponential modes. This construction turns fractional memory into an efficient recurrent module with parallel training and prefill, while retaining bounded-state autoregressive decoding. Extensive experiments, including 1.3B-parameter language modeling, demonstrate that FRAC consistently improves long-context performance over state-of-the-art SSM baselines while staying competitive on short-context. These results show that fractional dynamics provide a practical and effective prior for long-context SSMs.

As large language models (LLMs) process increasingly longer prompts, computation and KV-cache memory costs have emerged as major bottlenecks in inference systems. Existing input-level prompt compression methods address this, but rank each sentence by a scalar relevance score, treating the document as an unstructured pool of words and sentences. Under tight budgets, this causes theme collapse, where the dominant theme(s) of a document consumes the budget, discarding less-frequent yet task-relevant themes. Preserving thematic coverage instead requires allocating the budget across recurring themes rather than scoring sentences in isolation. To this end, we propose SALT, a model-agnostic extractive framework that organizes per-sentence keywords into a trie ordered by sentence frequency (SF), a lightweight, reusable proxy for document thematic structure. This trie-based organization smooths memory allocation and prevents dominant themes from monopolizing the budget. Multi-anchor retrieval activates trie nodes labeled by query keywords at any depth, and the trie persists across dialogue turns, supporting multi-turn use without re-encoding the document. By preserving document themes, SALT reduces the prefill computation and memory cost of long-context prompts while remaining composable with KV-cache methods that target decoding-time latency and memory.

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

We introduce Long-VITA, a simple yet effective large multi-modal model for long-context visual-language understanding tasks. It is adept at concurrently processing and analyzing modalities of image, video, and text over 4K frames or 1M tokens while delivering advanced performances on short-context multi-modal tasks. We propose an effective multi-modal training schema that starts with large language models and proceeds through vision-language alignment, general knowledge learning, and two sequential stages of long-sequence fine-tuning. We further implement context-parallelism distributed inference and logits-masked language modeling head to scale Long-VITA to infinitely long inputs of images and texts during model inference. Regarding training data, Long-VITA is built on a mix of 17M samples from public datasets only and demonstrates the state-of-the-art performance on various multi-modal benchmarks, compared against recent cutting-edge models with internal data. Long-VITA is fully reproducible and supports both NPU and GPU platforms for training and testing. By leveraging our inference designs, Long-VITA models achieve a remarkable 2x prefill speedup and 4x context length extension in single node with 8 GPUs. We hope Long-VITA can serve as a competitive baseline and offer valuable insights for the open-source community in advancing long-context multi-modal understanding.

Coding agents are rapidly becoming a major application of agentic LLMs, but serving them efficiently remains challenging. Progress on this challenge requires understanding real workload patterns, yet the data needed for such analysis is largely absent. Existing public traces and benchmarks do not capture real, day-to-day coding-agent usage across multiple agents and model families for serving-system analysis. To help fill this gap, we collect and release a trace of roughly 4,300 coding-agent sessions, containing about 350,000 LLM steps and 430,000 tool calls from our own day-to-day use of Claude Code and Codex. Our analysis shows that coding-agent workloads feature long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates. These findings point to concrete opportunities for optimizing serving, including lower-overhead tool calling, append-length-aware prefill, semantic-aware tool-latency prediction, and improved KV-cache management around human-paced gaps. We release the dataset, trace collection pipeline, and analysis code at https://github.com/uw-syfi/TraceLab.git the project website is https://tracelab.cs.washington.edu.

Existing token compression methods for omnimodal large language models typically rely on one modality to determine what to retain in the other. We show that this assumption often breaks down: for the same query, audio and video relevance often peaks at different moments. This cross-modal salience mismatch makes unidirectional guidance prone to discarding answer-critical cues under aggressive compression. We propose OmniScope, a training-free token compression framework that uses the query as a shared semantic anchor while estimating relevance separately for audio and video. OmniScope allocates modality-specific token budgets, prunes visual tokens with an anchor-delta strategy that preserves both global context and temporal changes, and merges audio tokens within each second to reduce redundancy while maintaining temporal continuity. Across four audio-video benchmarks and two Qwen2.5-Omni model scales, OmniScope achieves the best average accuracy across all compression settings. At 25% overall token retention, it delivers up to 3.53x prefill speedup and more than 15% GPU memory reduction, with only a 0.35-point drop in average accuracy. These results suggest a simple design principle for OmniLLM inference: share the query across modalities, but not the salience estimates. The code is available at https://github.com/MAC-AutoML/OmniScope.

Omnimodal large language models (OmniLLMs) are rapidly extending multimodal reasoning to cover synchronized audio and video. However, the resulting audio-video token sequences are long, leading to high prefill latency and GPU memory usage at inference time. Existing token pruning methods, designed mainly for vision-only inputs, miss both the cross-modal links between audio and video and the user query that decides which content matters. To bridge this gap, we present Omni-Prune, a training-free, query-aware audio-visual token pruning framework that jointly removes redundancy from both modalities while keeping task-relevant cross-modal evidence. Specifically, Omni-Prune first splits the token sequence into adaptive time windows placed at audio saliency peaks, then scores audio and video tokens on a single scale that combines encoder attention with text-query relevance, and pairs related audio-video tokens so that they are kept together. Within each window, a final K-medoids step then selects a few representative tokens, adding diverse cues that score-based selection alone would miss. Extensive experiments demonstrate that Omni-Prune outperforms established baseline methods, delivering up to 3.25x prefill speedup and 1.3x memory reduction while retaining over 99% of full-model performance.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

Sparse attention reduces compute and memory bandwidth for long-context LLM inference. However, two key challenges remain: (1) KV cache capacity still grows with sequence length, and offloading to CPU memory introduces a PCIe transfer bottleneck; (2) the sparse selection step itself retains O(T^2) complexity and can dominate attention cost at long contexts. We propose SparDA, a decoupled sparse attention architecture that introduces a fourth per-layer projection, the Forecast, alongside Query, Key, and Value. The Forecast predicts the KV blocks needed by the next layer, enabling lookahead selection that overlaps CPU-to-GPU prefetch with current-layer execution. Because Forecast is decoupled from the attention query, our GQA implementation uses one Forecast head per GQA group, reducing selection overhead versus the original multi-head selector. SparDA adds <0.5% parameters and trains only the Forecast projections by matching the original selector's attention distribution. On two sparse-pretrained 8B models, SparDA matches or slightly improves accuracy and delivers up to 1.25times prefill speedup and 1.7times decode speedup over the sparse-attention offload baseline. By enabling larger feasible batch sizes on a single GPU, SparDA further reaches up to 5.3times higher decode throughput than the non-offload sparse baseline. Our source code is available at https://github.com/NVlabs/SparDA.

Large language models (LLMs) are reshaping recommender systems by enabling richer modeling of users, content, and context directly in natural language. At Netflix, we are exploring this direction through GenRec, an LLM-backed recommendation ranker built on top of an in-house foundational LLM. GenRec follows a two-phase framework: Phase 1 adapts an open-source LLM to Netflix data, developing deep understanding of the catalog and member behavior while balancing capabilities such as content understanding and instruction following. Phase 2 post-trains this foundation model with recommendation-ranking specific data, labels, and reward signals, aiming to align the ranker with business requirements and long-term member satisfaction. This paper focuses on Phase 2 and the transition from a traditional discriminative ranker with thousands of engineered features to an LLM-backed ranker driven by verbalized user histories and context. We describe our design for input verbalization and context engineering, post-training data construction, reward integration, model architecture, and a cost-constrained serving design based on a prefill-only inference approach. We report results from a large-scale A/B test comparing GenRec against the current production ranker model, where we show that a GenRec model trained with substantially fewer Phase-2 labeled training examples and input signals can achieve statistically significant gains in offline and online metrics. We discuss how LLM-backed recommenders could shift the recommendation paradigm: from feature engineering to context engineering, and from bespoke architectures to shared foundation backbones. We also outline practical lessons for serving such systems under real-world resource constraints.

The design of attention mechanisms in long-context large language models faces an ``impossible triangle'' among KV cache size, training and prefill FLOPs, and model quality. GQA achieves strong quality but incurs high KV cache overhead; alternative approaches such as MLA and MFA can compress the cache, but MLA suffers from quality loss while MFA substantially increases training and prefill compute. We propose Virtual Head Attention (VHA), which simultaneously improves all three vertices through two lightweight linear operations: Q Premix applies a near-identity transformation to queries within each KV group to recover virtual head diversity from halved physical query heads, and Linear Postmix fuses inter-head features via a low-rank I + AB^top residual structure. Since VHA introduces no nonlinear operations, Premix can be folded into the query projection and Postmix into the output projection at inference time, reducing the model to a standard GQA-2 architecture that directly reuses all existing GQA inference optimizations. With only two KV heads, the group size in large models readily reaches 64, fully exploiting Tensor Core parallelism during decoding. Across four scales with 4K-context 50B-token pretraining, VHA consistently surpasses the standard GQA baseline at every scale (+0.05\% to +0.81\%), while reducing KV cache by 4times vs.\ GQA-8 at dense scales and 2times vs.\ GQA-4 at the MoE scale, and lowering training and prefill FLOPs by 7.9\% (1.7B, S{=}4096). On an industrial 30B-A3B MoE model trained on 500B tokens at 8K context, VHA outperforms both MLA and GQA-4 by +0.70\,/\,+2.60\,\% on a 12-benchmark base-model evaluation suite; in a separate matched 10B-A1B MoE comparison after long-context mid-training, VHA attains an average RULER score of 84.4 over 8K--128K versus 73.0 for GQA, while using two rather than four KV heads in the full-attention layers.

Autoregressive (AR) models have reformulated image generation as next-token prediction, demonstrating remarkable potential and emerging as strong competitors to diffusion models. However, control-to-image generation, akin to ControlNet, remains largely unexplored within AR models. Although a natural approach, inspired by advancements in Large Language Models, is to tokenize control images into tokens and prefill them into the autoregressive model before decoding image tokens, it still falls short in generation quality compared to ControlNet and suffers from inefficiency. To this end, we introduce ControlAR, an efficient and effective framework for integrating spatial controls into autoregressive image generation models. Firstly, we explore control encoding for AR models and propose a lightweight control encoder to transform spatial inputs (e.g., canny edges or depth maps) into control tokens. Then ControlAR exploits the conditional decoding method to generate the next image token conditioned on the per-token fusion between control and image tokens, similar to positional encodings. Compared to prefilling tokens, using conditional decoding significantly strengthens the control capability of AR models but also maintains the model's efficiency. Furthermore, the proposed ControlAR surprisingly empowers AR models with arbitrary-resolution image generation via conditional decoding and specific controls. Extensive experiments can demonstrate the controllability of the proposed ControlAR for the autoregressive control-to-image generation across diverse inputs, including edges, depths, and segmentation masks. Furthermore, both quantitative and qualitative results indicate that ControlAR surpasses previous state-of-the-art controllable diffusion models, e.g., ControlNet++. Code, models, and demo will soon be available at https://github.com/hustvl/ControlAR.

This technical report presents QwenLong-CPRS, a context compression framework designed for explicit long-context optimization, addressing prohibitive computation overhead during the prefill stage and the "lost in the middle" performance degradation of large language models (LLMs) during long sequence processing. Implemented through a novel dynamic context optimization mechanism, QwenLong-CPRS enables multi-granularity context compression guided by natural language instructions, achieving both efficiency gains and improved performance. Evolved from the Qwen architecture series, QwenLong-CPRS introduces four key innovations: (1) Natural language-guided dynamic optimization, (2) Bidirectional reasoning layers for enhanced boundary awareness, (3) Token critic mechanisms with language modeling heads, and (4) Window-parallel inference. Comprehensive evaluations across five benchmarks (4K-2M word contexts) demonstrate QwenLong-CPRS's threefold effectiveness: (1) Consistent superiority over other context management methods like RAG and sparse attention in both accuracy and efficiency. (2) Architecture-agnostic integration with all flagship LLMs, including GPT-4o, Gemini2.0-pro, Claude3.7-sonnet, DeepSeek-v3, and Qwen2.5-max, achieves 21.59times context compression alongside 19.15-point average performance gains; (3) Deployed with Qwen2.5-32B-Instruct, QwenLong-CPRS surpasses leading proprietary LLMs by 4.85 and 10.88 points on Ruler-128K and InfiniteBench, establishing new SOTA performance.

Scaling language models unlocks impressive capabilities, but the accompanying computational and memory demands make both training and deployment expensive. Existing efficiency efforts typically target either parameter sharing or adaptive computation, leaving open the question of how to attain both simultaneously. We introduce Mixture-of-Recursions (MoR), a unified framework that combines the two axes of efficiency inside a single Recursive Transformer. MoR reuses a shared stack of layers across recursion steps to achieve parameter efficiency, while lightweight routers enable adaptive token-level thinking by dynamically assigning different recursion depths to individual tokens. This allows MoR to focus quadratic attention computation only among tokens still active at a given recursion depth, further improving memory access efficiency by selectively caching only their key-value pairs. Beyond these core mechanisms, we also propose a KV sharing variant that reuses KV pairs from the first recursion, specifically designed to decrease prefill latency and memory footprint. Across model scales ranging from 135M to 1.7B parameters, MoR forms a new Pareto frontier: at equal training FLOPs and smaller model sizes, it significantly lowers validation perplexity and improves few-shot accuracy, while delivering higher throughput compared with vanilla and existing recursive baselines. These gains demonstrate that MoR is an effective path towards large-model quality without incurring large-model cost.

Long-context agentic workflows have emerged as a defining use case for large language models, making attention efficiency critical for both inference speed and serving cost. Sparse attention addresses this challenge effectively, and DeepSeek Sparse Attention (DSA) is a representative production-grade solution: a lightweight lightning indexer selects the top-k most relevant tokens per query, reducing core attention from O(L^2) to O(Lk). However, the indexer itself retains O(L^2) complexity and must run independently at every layer, despite the fact that the resulting top-k selections are highly similar across consecutive layers. We present IndexCache, which exploits this cross-layer redundancy by partitioning layers into a small set of Full layers that run their own indexers and a majority of Shared layers that simply reuse the nearest Full layer's top-k indices. We propose two complementary approaches to determine and optimize this configuration. Training-free IndexCache applies a greedy search algorithm that selects which layers to retain indexers by directly minimizing language modeling loss on a calibration set, requiring no weight updates. Training-aware IndexCache introduces a multi-layer distillation loss that trains each retained indexer against the averaged attention distributions of all layers it serves, enabling even simple interleaved patterns to match full-indexer accuracy. Experimental results on a 30B DSA model show that IndexCache can remove 75% of indexer computations with negligible quality degradation, achieving up to 1.82times prefill speedup and 1.48times decode speedup compared to standard DSA. These positive results are further confirmed by our preliminary experiments on the production-scale GLM-5 model (Figure 1).

Autoregressive decoding with LLMs is primarily bottlenecked by GPU memory bandwidth, especially in edge-computing settings. While quantization is essential for mitigating this bottleneck, most existing methods treat inference as a uniform process and fail to account for the asymmetry between the compute-bound prefill stage and the memory-bound decoding stage. We propose GRINQH (GRaded INput-based Quantization Hierarchy), a weight-only post-training quantization framework that accelerates decoding by unifying quantization and sparsification. GRINQH leverages activation magnitudes as a proxy for computational importance to dynamically assign weight channels to different precision levels, enabling flexible average bit widths during decoding. Evaluated on Llama3 and Qwen3 models, GRINQH outperforms state-of-the-art fixed- and mixed-precision baselines at comparable 3- and 4-bit settings, even enabling effective 2-bit generation. We experimentally verify theoretical speedups by leveraging a hierarchical nested memory layout for multi-precision storage in a custom GPU kernel. Ultimately, GRINQH establishes a new state-of-the-art Pareto frontier for LLM generation, enabling a dynamic trade-off between generation quality and inference speed.

Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error. As pruning objectives, budgets, and model architectures diversify, manually navigating the expanding design space becomes increasingly difficult. This paper aims to build an AI4AI framework for visual-token pruning by addressing a natural question: Can large language models automatically design effective visual-token reduction algorithms? Although LLMs possess broad algorithmic knowledge and strong reasoning capabilities, translating such general knowledge into effective solutions for a specialized task remains nontrivial. We argue that the key lies in designing an appropriate search-state representation that connects the internal knowledge of LLMs with the structural requirements and constraints of visual-token pruning. Based on this insight, we propose AutoPrune, a training-free framework for LLM-driven visual-token pruning policy design. At its core, AutoPrune introduces a Token Pruning Domain-Specific Language (TPDSL) comprising 131 reusable atoms for budget control, token scoring, selection constraints, and token reassembly. A key property of TPDSL is that it represents each search state as a residual modification of a strong base policy. This residual formulation narrows the search space and directs the LLM's attention toward the policy components that are most consequential for performance. Experiments on 14 multimodal benchmarks and three MLLM backbones demonstrate the effectiveness, efficiency, and transferability of AutoPrune. Even when removing 94.4% of visual tokens, AutoPrune preserves more than 99% of full-token performance while reducing FLOPs by 9.9x and prefill latency by 6.4x.

Vision-language models face a fundamental scaling bottleneck: the number of visual tokens grows with both temporal duration and spatial resolution, making long-video understanding expensive for the vision encoder and the language model. Existing methods often compress visual tokens after dense encoding, creating a mismatch between the representation used during training and the compact interface required at deployment. We present CoVisco, a codec-native vision encoder with native token compression for unified image-video understanding. By combining codec-native input support with segmented attention, CoVisco can encode long visual inputs in a single forward pass without forming dense patch-to-patch interactions across all frames. Each temporal segment is equipped with learnable abstract tokens that learn a compact segment-level representation, while fine-grained patch tokens remain available throughout the encoder. Alternating intra-segment and abstract-communication layers preserve video-level context through the abstract-token channel. A lightweight selector further exposes either abstract tokens alone or abstract tokens augmented with a runtime-selected subset of patch tokens, yielding a compact visual interface that reduces the visual context and prefill burden of downstream MLLMs while retaining fine-grained evidence when needed. Pretrained with contrastive objectives on 565M image--text pairs and 6.4M videos, CoVisco shows competitive performance on video-oriented embedding and multimodal understanding benchmarks. In the evaluated four-segment, 64-frame setting, abstract-only inference uses only 400 visual tokens while achieving video-understanding performance close to, and on some benchmarks exceeding, OneVision-Encoder. Selected patch tokens further improve fine-grained video reasoning. Project URL: https://github.com/ernie-research/CoVisco.git