The proliferation of 100B+ parameter Large Language Models (LLMs) with 100k+ context length support have resulted in increasing demands for on-chip memory to support large KV caches. Techniques such as StreamingLLM and SnapKV demonstrate how to control KV cache size while maintaining model accuracy. Yet, these techniques are not commonly used within industrial deployments using frameworks like vLLM or SGLang. The reason is twofold: on one hand, the static graphs and continuous batching methodology employed by these frameworks make it difficult to admit modifications to the standard multi-head attention algorithm, while on the other hand, the accuracy implications of such techniques on modern instruction-following and reasoning models are not well understood, obfuscating the need for implementing these techniques. In this paper, we explore these accuracy implications on Llama-3.1-8B-Instruct and DeepSeek-R1, and develop SnapStream, a KV cache compression method that can be deployed at scale. We demonstrate the efficacy of SnapStream in a 16-way tensor-parallel deployment of DeepSeek-671B on SambaNova SN40L accelerators running at 128k context length and up to 1832 tokens per second in a real production setting. SnapStream enables $4\times$ improved on-chip memory usage and introduces minimal accuracy degradation on LongBench-v2, AIME24 and LiveCodeBench. To the best of our knowledge, this is the first implementation of sparse KV attention techniques deployed in a production inference system with static graphs and continuous batching.

Long-context Large Language Models (LLMs) face significant memory bottlenecks during inference due to the linear growth of key-value (KV) cache with sequence length. While individual optimization techniques like KV cache quantization, chunked prefill, and model weight quantization have shown promise, their joint effects and optimal configurations for edge deployment remain underexplored. We introduce KV Pareto, a systems-level framework that systematically maps the trade-off frontier between total memory consumption and task accuracy across these three complementary optimization techniques. Our framework evaluates multiple LLM architectures (Qwen, Llama, Mistral) with varying KV quantization schemes (int2/4/8, mixed-precision), granularities (per-token, per-tensor, per-block), and 4-bit weight quantization via AWQ. Our framework identifies model-specific Pareto-optimal configurations that achieve 68-78% total memory reduction with minimal (1-3%) accuracy degradation on long-context tasks. We additionally verify the selected frontiers on additional benchmarks of Needle-in-a-Haystack, GSM8k and MMLU as well as extended context lengths of up to 128k to demonstrate the practical need of joint optimization for efficient LLM inference.

The rapid evolution of Large Language Models (LLMs) has driven the need for increasingly sophisticated inference pipelines and hardware platforms. Modern LLM serving extends beyond traditional prefill-decode workflows, incorporating multi-stage processes such as Retrieval Augmented Generation (RAG), key-value (KV) cache retrieval, dynamic model routing, and multi step reasoning. These stages exhibit diverse computational demands, requiring distributed systems that integrate GPUs, ASICs, CPUs, and memory-centric architectures. However, existing simulators lack the fidelity to model these heterogeneous, multi-engine workflows, limiting their ability to inform architectural decisions. To address this gap, we introduce HERMES, a Heterogeneous Multi-stage LLM inference Execution Simulator. HERMES models diverse request stages; including RAG, KV retrieval, reasoning, prefill, and decode across complex hardware hierarchies. HERMES supports heterogeneous clients executing multiple models concurrently unlike prior frameworks while incorporating advanced batching strategies and multi-level memory hierarchies. By integrating real hardware traces with analytical modeling, HERMES captures critical trade-offs such as memory bandwidth contention, inter-cluster communication latency, and batching efficiency in hybrid CPU-accelerator deployments. Through case studies, we explore the impact of reasoning stages on end-to-end latency, optimal batching strategies for hybrid pipelines, and the architectural implications of remote KV cache retrieval. HERMES empowers system designers to navigate the evolving landscape of LLM inference, providing actionable insights into optimizing hardware-software co-design for next-generation AI workloads.

Agentic reinforcement learning (RL) trains large language models to autonomously call tools during reasoning, with search as the most common application. These models excel at multi-step reasoning tasks, but their safety properties are not well understood. In this study, we show that RL-trained search models inherit refusal from instruction tuning and often deflect harmful requests by turning them into safe queries. However, this safety is fragile. Two simple attacks, one that forces the model to begin response with search (Search attack), another that encourages models to repeatedly search (Multi-search attack), trigger cascades of harmful searches and answers. The attacks succeed by triggering models to generate harmful, request-mirroring search queries before they can generate the inherited refusal tokens. This exposes a core weakness of current RL training: it rewards continued generation of effective queries without accounting for their harmfulness. As a result, RL search models have vulnerabilities that users can easily exploit, making it urgent to develop safety-aware agentic RL pipelines optimising for safe search. Instruction tuning (IT) is the standard method to align large language models (LLMs) with human preferences and teach them to refuse harmful requests (Schulman et al., 2017; Shao et al., 2024). However, IT only shapes static responses and is insufficient in agentic settings, where models must also decide when and how to call external tools, capabilities not explicitly learned during pre-training (Zhang et al., 2025). Agentic reinforcement learning (RL) for tool-use (Zhang et al., 2025) tackles this by fine-tuning models to interleave reasoning with tool use (Dong et al., 2025). In practice, search is the most common tool: agentic RL rewards effective, well-timed search queries and achieves strong gains on multi-hop reasoning tasks (Song et al., 2025a;b; Jin et al., 2025). Despite the progress, effect of agentic RL on safety of IT models remains unclear. While prior work reported safety degradation of retrieval-augmented agents (Y u et al., 2025), little is known about whether agentic RL for search preserves refusal of harmful requests. As agentic RL is now deployed in closed-source systems such as OpenAI's DeepSearch (OpenAI, 2025), this evaluation gap can create real deployment risks.

Large Language Model (LLM) inference has emerged as a fundamental paradigm. In real-world scenarios, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose ARES, an adaptive decoding rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with : A dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 74.77% and achieving up to 2.24 times higher goodput.

Modern large language models (LLMs) extend context lengths to millions of tokens, enabling coherent, personalized responses grounded in long conversational histories. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly becomes the bottleneck in resource-constrained environments. An active line of research for reducing memory bottleneck is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting the KV cache after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to failure cases 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%, maintains near-full KV accuracy under 4-6x compression, and reduces latency/memory by up to 2.4x/3.5x, enabling efficient multi-turn interaction under strict resource limits. Our code is available at https://github.com/apple/ml-epicache.

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) deployed on edge servers are increasingly used in latency-sensitive applications such as personalized assistants, recommendation, and content moderation. However, the non-stationary nature of user data necessitates frequent retraining, which introduces a fundamental tension between inference latency and model accuracy under constrained GPU resources. Existing retraining strategies either delay model updates, over-commit resources to retraining, or overlook iteration-level retraining granularity. In this paper, we identify that iteration-level scheduling is crucial for adapting retraining frequency to model drift without violating service-level objectives (SLOs). We propose MACE, a hybrid LLM system that colocates concurrent inference (prefill, decode) and fine-tuning, with intelligent memory management to maximize task performance while promising inference throughput. MACE leverages the insight that not all model updates equally affect output alignment and allocates GPU cycles accordingly to balance throughput, latency, and update freshness. Our trace-driven evaluation shows that MACE matches or exceeds continuous retraining while reducing inference latency by up to 63% and maintaining throughput under resource constraints. Compared to periodic retraining, MACE improves latency breakdown across prefill, decode, and finetune stages, and sustains GPU utilization above 85% in NVIDIA AGX Orin. These results demonstrate that iteration-level hybrid scheduling is a promising direction for deploying LLMs with continual learning capabilities on edge platforms.

Serving disaggregated large language models has been widely adopted in industrial practice for enhanced performance. However, too many tokens generated in decoding phase, i.e., occupying the resources for a long time, essentially hamper the cloud from achieving a higher throughput. Meanwhile, due to limited on-device resources, the time to first token (TTFT), i.e., the latency of prefill phase, increases dramatically with the growth on prompt length. In order to concur with such a bottleneck on resources, i.e., long occupation in cloud and limited on-device computing capacity, we propose to separate large language model between cloud and devices. That is, the cloud helps a portion of the content for each device, only in its prefill phase. Specifically, after receiving the first token from the cloud, decoupling with its own prefill, the device responds to the user immediately for a lower TTFT. Then, the following tokens from cloud are presented via a speed controller for smoothed TPOT (the time per output token), until the device catches up with the progress. On-device prefill is then amortized using received tokens while the resource usage in cloud is controlled. Moreover, during cloud prefill, the prompt can be refined, using those intermediate data already generated, to further speed up on-device inference. We implement such a scheme P/D-Device, and confirm its superiority over other alternatives. We further propose an algorithm to decide the best settings. Real-trace experiments show that TTFT decreases at least 60%, maximum TPOT is about tens of milliseconds, and cloud throughput increases by up to 15x.

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.