ai infrastructure engineers

Abstract: Coding agents are beginning to develop the production-grade AI infrastructure that trains and serves large models, yet existing benchmarks cover only single kernels or single-GPU tasks, and grade with tests and prompts that misjudge both correct and incorrect patches. We introduce InfraDev, a benchmark of 36 tasks built from real-world pull requests to Megatron-LM, vLLM, and verl, covering infrastructure development for training, inference, and reinforcement learning. Each task asks an agent to reimplement a real change that spans many files and must run correctly across many GPUs. To grade such changes, we jointly synthesize hidden probes, which enforce the requirements the original tests leave unchecked, and a prompt that states every interface the probes call without revealing the solution. We also design a shared execution environment that reuses dependencies and artifacts across tasks, cutting storage by 52-57% and build time by 13-14% on Megatron-LM and vLLM. Across six frontier models, the best resolves only 38.9% of the tasks; the probes reject 32% of the submissions that pass every official test, and with the original pull request descriptions as prompts, no model resolves a single task. Failure analysis shows that structural and state errors dominate and that agents' own tests rarely check memory, time, or traffic. Turning these findings into an agent skill raises the resolved rate of a frontier model on the Megatron-LM tasks by 10.5%.

Abstract: Reasoning language models generate thousands of tokens per answer, and generation speed is limited by memory bandwidth, so storing weights in two bits instead of bfloat16 can potentially make inference 8$\times$ cheaper. Two-bit methods are usually judged by perplexity and multiple-choice question-answering accuracy, both of which say little about whether a model can still reason through more difficult problems. We present ThinkBit, a strictly scalar two-bit framework built to preserve reasoning, in which every weight is an integer codeword times a bfloat16 scale shared by a group of weights. We first use a sensitivity metric to assign each layer a ternary, 2-bit, or 3-bit tier to allocate bitwidths optimally among layers. This is followed by a layerwise Gumbel-softmax-based scalar quantization algorithm that searches for optimal codewords block by block. Since a layerwise scheme does not capture how quantized layers interact with each other, we recalibrate scales, with the codewords frozen, using a knowledge-distillation-based loss against the teacher model. We then perform a short quantization-aware training (QAT) stage that lets the model mimic the base model's distribution by reassigning codewords. Finally, a paired reasoning test on training problems selects the final checkpoint between the recalibration stage and the QAT stage. On Qwen3-1.7B at matched size, our recalibrated checkpoint scores 46.7 on MATH-500 against 30.0 for EfficientQAT trained on the same data, and 73.0 after QAT on reasoning traces; at 0.52 GB it is smaller than every published 2-bit GGUF of the model yet more accurate on MATH-500. We show that this framework scales and up to 14B, a 3.74 GB model scores 94.8 on MATH-500, against 97.5 for the 29.5 GB bfloat16 model.

Abstract: This paper presents Memory- and Communication-Aware Framework for Long-Context LLM Inference (MCF) that optimizes long-context inference through communication-aware and stage-adaptive KV-cache layout transitions. At the core of MCF is {Memory-Aware Replication and Scheduling} (MARS), a dynamic orchestration mechanism for chunked prefill. MARS exploits early-stage memory headroom by selectively replicating highly reused KV blocks, improving locality and hiding KV propagation latency behind attention computation. As prefill progresses and memory pressure increases, MARS progressively drains redundant replicas and converges the KV cache into a sharded decode-ready layout, avoiding expensive layout conversion at the prefill--decode boundary. For decoding, MCF adopts a query-centric execution model in which KV shards remain stationary while lightweight query tensors are routed across workers. This changes communication from KV-cache-dependent transfer to query-dependent transfer, reducing communication growth with context length. MCF further integrates a hardware-software co-designed execution layer based on an FPGA compute-NIC datapath, enabling metadata-driven KV routing and collective operations directly in the network path. Experiments show that MCF improves TTFT, throughput, and scalability over state-of-the-art distributed inference systems, with larger benefits as context length and cluster size increase.

Abstract: Multi-turn LLM agents usually reuse long context prefixes, making key value (KV) caching crucial for reducing prefill latency. SSDs provide a low-cost extension to limited DRAM capacity but introduce two key challenges. Eagerly persisting all KV cache blocks incurs substantial write traffic and rapidly consumes finite SSD endurance. Meanwhile, the small number of SSDs provisioned in typical GPU servers limits aggregate reload bandwidth, making SSD-resident KV cache much slower to retrieve than that in DRAM and inflating prefill latency. We present \textsc{Saria} (Selective Admission and Reload for Inference Acceleration), a KV caching system that jointly manages DRAM and SSD by exploiting the session-level temporal locality of LLM agents to reduce SSD writes and accelerate cache reload. On the write path, \textsc{Saria} uses eviction and reuse history to selectively persist KV cache blocks that exhibit reuse opportunities beyond their DRAM residency, avoiding unnecessary writes for KV cache already hosted in DRAM. On the read path, \textsc{Saria} promotes selected SSD-hit KV cache blocks from GPU memory into DRAM after prefill, allowing subsequent reuse within the same session to reduce SSD read traffic while limiting DRAM contention. A transfer coordinator paces DRAM reads while SSD transfers are active, allowing both paths to make progress and efficiently share GPU ingress bandwidth. We implement \textsc{Saria} on LMCache MP and evaluate it using workloads from TraceLab and WEKA, covering both full-attention and hybrid-attention models. The results show that \textsc{Saria} reduces SSD writes by 57.1--58.9\% compared with popular full duplicate caching and the mean time to first token (TTFT) by 42.3--79.7\% over DRAM-only caching.

Abstract: Normalizing Flows (NFs) have shown strong results in generative modeling, recently reaching state-of-the-art image generation performance. However, their repetitive, sequential operations limit the computational and memory efficiency achievable with GPUs and TPUs. We propose M3D-FLOW, a novel monolithic 3D-stackable (M3D) DRAM-based processing-near-memory (PNM) architecture that eliminates memory bottlenecks using a Cu-Cu hybrid-bonded logic die integrated with an M3D DRAM die. The M3D DRAM provides dense bank integration and a tiered organization with variable access latencies that M3D-FLOW exploits for the efficient processing of different NF workloads. To maximize compute density while minimizing area, energy, and thermal overhead, M3D-FLOW employs hybrid binary--stochastic floating-point multipliers, performing mantissa multiplication in the stochastic domain and exponent operations in the binary domain. A pipelined execution strategy maps sequential NF stages across dynamically partitioned PE groups, and together with the stochastic hardware, enables M3D-FLOW to achieve at least $6.6\times$ higher performance and $3\times$ better energy efficiency than prior PNM accelerators, with only a 2.2% bits-per-dimension (BPD) increase over bfloat16 and 4.1% over FP32. An analysis of stochastic hardware variants further shows that shorter stochastic bitstreams reduce execution time by up to $3.3\times$ at the cost of higher approximation error, providing a concrete design guide for latency--accuracy trade-offs.

Abstract: Key-value (KV) cache management has become a critical bottleneck for scaling Large Language Models (LLMs) to long-context scenarios due to linear memory growth. While recent eviction methods have moved beyond simple attention-score heuristics to incorporate value representations, they typically face a fundamental trade-off: the underlying optimization for minimizing output perturbation is an NP-hard combinatorial problem, leading to mathematically suboptimal heuristics or implementations that are difficult to kernelize. In this work, we propose DropKV, a simple principled eviction framework based on Decoupled Residual-Output Perturbation. By decoupling the joint eviction decision into independent per-token scoring, DropKV circumvents combinatorial intractability and admits a provable constant-factor approximation guarantee under a cone condition that we empirically verify on three long-context LLMs. To ensure practical viability, we provide a high-performance implementation using fused Triton kernels that avoid materializing the full attention matrix, delivering a 19.8x scoring-kernel speedup that translates to up to a 9.9% matched-batch end-to-end prefill speedup. Extensive evaluations on the RULER and LongBench benchmarks demonstrate that DropKV consistently outperforms state-of-the-art baselines at equivalent cache budgets, achieving the lowest inference latency among eviction methods while also remaining faster than the dynamic cache baseline.

Abstract: The rapid adoption of Retrieval-Augmented Generation (RAG) and long-context LLMs has shifted the computational burden toward the prefill phase---the initial stage where the model processes all input tokens in parallel before generating a response---yet its energy impact remains poorly characterised. We investigate the contradiction between massive hardware-level power spikes and the sub-linear marginal cost of input tokens reported in system-level studies. Through high-resolution telemetry of Llama-3.1 and Qwen-2 models on NVIDIA A100 hardware, we identify the efficiency threshold where $O(n^2)$ attention complexity overcomes GPU parallelisation. Our results quantify how optimised memory orchestration (PagedAttention) delays this scaling bottleneck across three orders of context scaling, yielding up to an 8.9$\times$ energy efficiency advantage at 112,000 tokens compared to standard baselines. Furthermore, we characterise the energy disparity between input and output tokens, demonstrating that optimised serving (vLLM) causes this ratio to narrow (from $\approx$850$\times$ to 750$\times$) as context scales due to high decode efficiency, whereas unoptimised baselines show a much smaller, roughly constant gap ($\approx$150--230$\times$) because their prefill is itself expensive. This work establishes context-scaling laws that differentiate engine-level optimisations from inherent model constraints, providing an empirical framework for sustainable LLM deployment.

Abstract: Large language models impose substantial memory demands and arithmetic costs during inference. Dense-to-mixture-of-experts (MoE) conversion is a model optimization technique that aims to reduce these arithmetic costs. Dense-to-MoE conversion partitions pretrained feed-forward network (FFN) neurons into experts and selects a subset for each token. Leading methods fall into two categories: analytic methods produce a converted model quickly using heuristics, while learning methods produce a model much more slowly by optimizing via gradients. Both approaches, however, fix the same active expert count, or "utilization," at every layer. We introduce Spectral-MoE, which learns to allocate a utilization budget heterogeneously across layers. Its name draws a metaphor between separating a dense FFN into experts and separating a light beam into its spectrum. Spectral-MoE consists of two distinct approaches: a static approach that uses a single per-layer scalar which is shared across tokens, and a dynamic approach which emits a per-token schedule. We demonstrate that both allocators improve performance over uniform on both the analytic-based and learning-based SOTA dense-to-MoE methods. At 25% FFN utilization, dynamic allocation reduces WikiText-2 perplexity relative to uniform by 1.04, 2.96, and 2.71 points on the learned base for Llama-2-7B, Llama-3-8B, and Qwen2.5-7B, respectively. The corresponding reductions on the analytic base are 2.45, 6.26, and 2.17 points. On the learned base, dynamic allocation improves six-task mean accuracy over uniform by an average of 1.5, 4.2, and 3.6 percentage points across the three models at FFN utilizations of 12.5%, 25%, and 50%, respectively. These results establish per-layer expert utilization as a novel optimization axis in dense-to-MoE conversion.

Abstract: We introduce FORGE (Fast On-chip Rotor Generation for Experts), a procedural weight generation system for Sparse Mixture-of-Experts (MoE) inference. Rather than storing and loading every expert weight tensor from HBM, we show how a compact accelerator inside each Streaming Multiprocessor can regenerate them on-the-fly from $O(\log^2 d)$ rotor parameters per projection. Per replaced layer, gigabytes of stored expert weights are approximated with under a megabyte of seeds. This addresses the dual memory bottleneck of autoregressive Sparse MoE decode, where per-expert FFN operations are HBM-bandwidth-bound at small routed-token counts and the aggregate expert footprint forces multi-node Expert Parallelism (EP) with all-to-all traffic. We realize the approach as the Rotor Generation Unit (RGU), a fixed-function datapath that generates FFN weight tiles without competing with Tensor Core execution. With 30% of the MoE FFN layers replaced across five Sparse MoE models, we estimate 1.3-2.8x higher system throughput at matched GPU count and 1.10-1.15x lower per-user decode latency at short context. In exchange, downstream accuracy on three model families falls below that of the same model fine-tuned without rotor replacement, and the replacement ratio controls this tradeoff. FORGE also expands KV cache capacity by 1.2-2.6x. It is agnostic to the bit width at which weights are stored, so it can compose with quantization. The RGU is fully implemented/verified in SystemVerilog at ~0.9% SM area overhead (0.032 mm^2/SM, 7nm).

Abstract: Network simulation plays an important role in HPC/AI system research by enabling exploration of network architectures, communication strategies, and system configurations at scales that may be impractical to study experimentally. Existing approaches to network simulator calibration and validation lack a principled methodology that combines parameter alignment, constrained inference, and independent validation for collective-centric HPC/AI networks. We present CALIPER, a methodology for systematic calibration and validation of HPC/AI network simulators. CALIPER distinguishes directly observable system parameters from auxiliary parameters which are configured via calibration and constrains the latter to feasible domains consistent with the target system. We evaluate CALIPER's hierarchical, scale-aware calibration and its effects on parameter identifiability and generalization using collective communication benchmarks on a reference system with up to 64-nodes at the Jülich Supercomputing Centre. Validation results show that the calibrated parameter configurations identified by CALIPER predict system performance with less than 5% mean errors across scale-up and scale-out network domains. CALIPER demonstrates that hierarchical, constrained calibration can produce simulator configurations whose predictive accuracy persists beyond the observations used for fitting.

Abstract: The current state of the art for batch selection and noise addition in differentially private training is a combination of DP-BandMF (informally, DP-SGD where the noise added in two different steps can be correlated if the steps are within $b$ of each other) and privacy amplification by sampling via Monte Carlo accounting. First, whereas prior batch selection strategies for DP-BandMF were inspired by Poisson sampling, random allocation is known to outperform Poisson sampling for DP-SGD. We therefore propose and analyze new batch selection strategies for DP-BandMF inspired by random allocation, improving RMSE by up to $1.85\times$ over previous work. Second, Monte Carlo accounting typically requires drawing $Ω(1/δ)$ samples from a privacy loss distribution, making accounting costly. We increase sampling and scoring throughput by two orders of magnitude through a carefully optimized GPU implementation. Finally, previous work chose the DP-BandMF strategy matrix and the batch selection strategy independently. We give a simple framework for choosing these parameters jointly, reducing RMSE by up to $13.4\%$ over fixed strategy matrices.

Abstract: Sparse attention reduces the cost of long-context attention, but existing kernels typically process queries independently, repeatedly loading and dequantizing KV entries shared across queries. We observe substantial overlap in the KV entries selected by neighboring queries, creating opportunities for cross-query reuse. We present QUILT, a workload-aware sparse-attention execution mechanism that jointly processes neighboring queries and reuses shared KV entries to reduce redundant memory traffic and computation. QUILT introduces Shift-and-Compare Set Decomposition (SCSD), which transforms irregular set operations into regular data-parallel primitives suitable for modern accelerators, and pipelines SCSD with attention computation to hide its overhead. Cascaded sharing captures reuse hierarchically at multiple granularities. A tile-aware execution strategy balances sharing granularity with hardware tile utilization and selectively removes low-importance query-specific tails to eliminate underutilized tiles. We evaluate QUILT on LongBench using GLM-5.3 and DeepSeek-3.2 under both tensor and sequence parallelism. Compared with the state-of-the-art sparse-attention kernel, QUILT reduces average kernel latency by up to 55.1% and processed KV data by up to 55.9%, while reducing time-to-first-token (TTFT) latency by up to 36.8% with negligible accuracy degradation.

Abstract: A large language model can only use the text that fits in its context window, and it recomputes its internal key-value (KV) state for a prompt every time the prompt is sent. We test a memory layer, the public package galahad-kv, that saves the KV state of each block of about 16,000 tokens to encrypted local NVMe disk and loads it back later, byte-exact, without recomputing it. We ran it on 50,000,000 tokens of real public text, served through vLLM on one NVIDIA H100, with Gemma 4 12B and Gemma 4 31B. Every block we probed was loaded back from the encrypted store with no recompute (100 of 100, at depths from 0 to 50M tokens) on both models. Loading a block was 2.8x to 4.3x faster than recomputing it and used 8.8x to 12.3x less GPU energy, and GPU memory stayed flat over the whole 50M-token stream. Asked about facts planted millions of tokens earlier, the 12B model gave the right answer 82 times out of 100 and the 31B model 98 times out of 100. Neither model made up an answer. The limits are as follows. This is reuse of stored state, not a wider attention window: one block is loaded at a time, and how well a question is answered depends on the model. Writing the memory is a one-time cost, and the store takes terabytes of local NVMe disk. We describe the test protocol, which is built to resist common ways of gaming long-context benchmarks, and give a single-GPU reproduction that uses public software and a free licence for the package.

Abstract: Large language model (LLM) serving is increasingly dominated by agentic workloads, in which agents and their sub-agents accumulate context as KV cache across many requests, consuming substantial memory. High-bandwidth flash (HBF) is a promising solution, providing an order of magnitude greater capacity at HBM-class read bandwidth, but its finite write endurance is the key limiting factor. Our key insight is that KV cache should be placed across HBM and HBF by its lifetime. Placing shorter-lived data in HBM lets HBM absorb more of an agent run's writes and sends less of them to HBF. As the lifetime of KV cache in agentic serving is dictated by the harness, the program that orchestrates the agents, we analyze its behavior and identify three axes along which lifetime diverges, temporal, structural, and inter-worker. Guided by these observations, we present Lachesis, a lifetime-aware KV cache placement layer between the agent harness and the serving engine. At write time, it places each segment in HBM or HBF according to its lifetime, and frees its blocks once the segment is no longer read. In trace-driven simulation, Lachesis extends HBF lifetime by 1.19-3.13x over HBM-first placement, reaching 3.3-12.2 device-years. Even under continuous 24x7 operation at the full load a tight SLO admits, HBF outlasts its five-year warranty on the multi-agent trace.

Abstract: TRANSIT is a transparent scale-in framework to enable multi-node model training on fewer GPUs while maintaining training efficiency by transparently leveraging CPU DRAM as an extension of GPU memory during distributed training. It achieves this through a user-space interposition layer, requiring no modifications to the application, training framework, cluster scheduler, device driver, or operating system. Furthermore, TRANSIT achieves higher efficiency by leveraging a zero-copy data path for CPU-GPU transfers. We evaluate TRANSIT on dense and MoE models across scales up to 64 NVIDIA H100 GPUs and multiple parallelism configurations over a RoCE network. Our evaluation shows that TRANSIT can: (a) outperform state-of-the-art framework-managed offloading techniques, achieving up to 68%, 59%, and 42% higher per-GPU throughput than TorchTitan, ZeRO-Offload, and ZeRO-Infinity, respectively, (b) enables training with 50% fewer GPUs while maintaining over 90% of baseline per-GPU throughput, (c) lower per-node network traffic by up to 33%, and (d) improve per-GPU throughput by up to 35% in communication-bound settings.

Abstract: Solid State Drives (SSDs) are becoming the dominant medium for performance-critical storage. Zoned Namespace (ZNS) SSDs are getting more and more attractive because sequential writes and explicit zone resets reduce address-mapping, over-provisioning, and internal garbage-collection costs. However, reclaiming space requires resetting an entire zone, so deleting one file does not free its space while other files there must be kept. This erase-before-write constraint becomes a bottleneck for write-intensive workloads, particularly AI checkpointing during model training, where frequent saves of model and optimizer state compete for space with long-retained checkpoints. We observe that retention policies inherently encode when data can be retired, avoiding the need for a learned lifetime predictor. We present RetainZ, a storage backend that translates these policies into explicit reclaim epochs. Its allocator isolates long-retained data and groups objects with nearby reclaim epochs, balancing delayed space reclamation against unused space at zone ends. Evaluated on FEMU with multi-tenant streams and checkpoint manifests up to Pythia-1B, RetainZ reduces space occupied by deleted data and unused zone tails by 41.7% and mean per-input checkpoint p99 latency by 47.5% at 84% target load compared with the allocator of ZenFS, an established lifetime-aware ZNS backend, and lowers host write amplification from 1.165 to 1.003.

Abstract: Coding agents powered by large language models (LLMs) repeatedly alternate between model inference and tool calls, creating long-lived sessions with reusable key-value (KV) states and asynchronous request resumptions. Logical readiness, however, does not ensure efficient admission in a shared serving system. Through direct trace analysis and trace-driven replay, we identify two mismatches: reusable KV states reside across storage tiers and incur unequal computational costs for state preparation, while requests with heterogeneous context lengths can decode inefficiently together. Our central insight is that exploiting admission flexibility can improve serving performance while preserving request progress. We present Coda, a coding-agent serving system that realizes this insight through a readiness-informed admission layer incorporating two mechanisms. Tiered-Aging state admission exploits bounded flexibility in admission order to improve admission efficiency and preserve request progress. Compatibility-Aware execution admission exploits flexibility in attention grouping within each model iteration to reduce mixed-context interference and improve shared decoding efficiency. For multi-worker configurations, Coda introduces a separate routing layer that considers KV-state residency, worker load, and request-to-worker context-length compatibility to guide efficient request placement. We evaluate Coda across different models and workloads against state-of-the-art coding-agent serving systems and vLLM. Across the single-worker and multi-worker experiments, Coda improves output-token and SLO-compliant throughput by 20.3% and 70.5% on average, with peak gains of 29.3% and 140.2%.

Abstract: Most KV-cache compression methods classify attention heads once, either offline or during prefill, and keep this classification fixed throughout generation. Across three models (1.5B-8B) and three regimes (needle retrieval, long chain-of-thought, and multi-turn recall), we measure head behavior on four model-regime combinations and find that most heads change their reading behavior at least once during generation. We introduce WakeKV, a reactive residency policy that moves cooling heads to a recoverable CPU reservoir rather than freezing or permanently evicting their state. At matched memory or budget, WakeKV consistently improves miss rate over frozen classification and destructive eviction, evaluated across five model-regime combinations and over three cited baselines (SnapKV, uniform R-KV, and ReasonAlloc) across four eligible combinations. A FlexiCache/vLLM implementation on Mistral-7B confirms the benefit on real hardware, improving throughput while retaining LongBench quality.

Abstract: Activation sparsity speeds up large language model (LLM) inference by setting unimportant activations to zero so that the corresponding computations can be skipped. Existing training-free methods, however, make different trade-offs: threshold-based methods such as TEAL adapt the sparsity level to each token but do not tightly control the realised sparsity, while TopK-based methods such as WINA enforce a fixed sparsity level but use the same sparsity budget for every token. Both also apply the same budget across transformer blocks, despite large differences in block sensitivity. We introduce TopK-Guided, a training-free method that addresses both limitations by combining bounded token-level sparsity adaptation with sensitivity-aware block-level budget allocation. Across Llama-2 and Llama-3 models, TopK-Guided consistently improves perplexity and downstream accuracy over TEAL and WINA while preserving essentially the same sparsitydependent projection compute as WINA, with the largest gains at high sparsity. Ablations show that both components provide complementary improvements.

Abstract: 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.

Abstract: Concurrent autoregressive inference creates a fundamental interference problem: prefilling a newly arrived long prompt can delay tokens for requests that are already decoding. Fixed prefill chunks reduce this interference, but the best chunk size depends on the model, hardware, load, and latency objective. We introduce Decode-Latency Feedback Prefill (DLFP), a model-free controller that changes only prefill work that overlaps active decodes. After a guarded scheduling cycle, DLFP uses the observed interval as proportional feedback to resize the next prefill chunk; isolated prefills remain unrestricted. We implement DLFP in vLLM and evaluate it with open-loop Poisson arrivals, exact token accounting, raw request traces, and NVIDIA telemetry. On Qwen3-0.6B in BF16 on one A100 80 GB GPU, three paired 100-request trials reduce P99 inter-token latency by 24.8%, 30.1%, and 28.2% (mean 27.7%, paired 95% confidence interval 21.0% to 34.3%) with exact output agreement, no failures, and unchanged SLO compliance. The benefit is not free: mean P99 time to first token increases 34.8% while remaining inside the declared SLO. Crucially, the mechanism does not generalize to Qwen3-8B, Qwen3-32B, or a two-GPU tensor-parallel configuration. We trace the failure to an asynchronous scheduler-call interval that is only a proxy for completed GPU iteration time. This negative result defines the boundary of the contribution and motivates a completion-timed controller for concurrent CPU and on-device inference. We do not claim mobile-device performance; the present work is a reproducible proof-of-concept and generalization study.

Abstract: As the context size of text processed with an LLM grows, the size of KV caches can outstrip the memory allocated for the original model weights. This impacts LLM throughput negatively, since decoding is memory-bound and decode cost grows with cache size. Recent work alleviates this bottleneck by discarding the least relevant tokens. Eviction introduces a tension, since a one-off decision to discard content may prove detrimental later. Instead, we focus on alternative choices that can lead to cache compression without evicting tokens. We achieve this by learning a selector that is able to produce, based on context, a per-layer cache configuration towards an overall compression budget. The selector operates along three axes: sharing one cache across layers (depth), caching at fewer bits (precision), or truncating the low-rank latent cache representations (rank). We call the resulting method KV-Kaizen, for the many small per-layer choices it compounds. We observe that these interventions taken independently and uniformly over all layers limit achievable compression because they degrade accuracy. Crucially, composing them locally and adaptively to the context can instead preserve accuracy while achieving large memory savings. At inference, the selector runs once, before pre-fill. In evaluations on instruction following and reasoning tasks, our selectors reach the Pareto frontier of accuracy against cache size, against learning-free and post-hoc baselines. On long-context tasks, KV-Kaizen improves on eviction and can be composed with it, reaching a 32x smaller decode-time cache on a 14B model while preserving accuracy. A 4x cache size reduction incurs no accuracy degradation from 7B parameters up, and a compressed model is more accurate than a smaller uncompressed one with the same cache size. Together, these findings support pre-training large models and compressing them only afterwards.

Abstract: Agentic sessions driven by Large language models (LLMs) often alternate between model inference and tool use, accumulating long histories across successive rounds. Serving these sessions efficiently requires reducing attention computation and retaining history key-value (KV) caches to avoid recomputation. Recently, frontier open-source LLMs adopt sparse attention to reduce computation by selecting only part of the history, while SSDs provide a cheaper alternative to CPU DRAM for storing KV caches. However, sparse KV selection depends on the ad hoc intermediate values during model inference, so it forces SSD reads to lie on the inference critical path. These reads are further slowed by fragmented accesses and read-write interference in SSDs. To address these challenges, we present Janus, an agentic serving framework for sparse attention LLMs with SSD-centric KV storage. Janus focuses on append prefill, which processes each round's newly added inputs and accounts for most history KV loading. To move SSD reads out of the critical path, Janus runs the model's own KV selection module on earlier intermediate values, predicting KV demand without additional training. The predicted reads overlap with model computation, and any prediction misses are fetched before attention executes to preserve model outputs. To improve SSD efficiency, Janus coalesces adjacent reads, packs scattered KV pages into sequential writes on the CPU, and limits background writes while reads are active. Across three models and three agentic traces, Janus outperforms existing works by up to 1.57-3.69 times (1.22-1.85 times on average) in terms of the time to first token latency, while maintaining decode efficiency.

Abstract: In agentic LLM services, a session calls an external tool, waits for it, and re-arrives to continue inference. Statically compiled NPU serving can fix the batch bucket set, the maximum batch size, and the number of KV cache slots at compile time. We define such an environment as a compile-time-static serving substrate and analyze the execution-time cost that tool waiting and re-arrival incur in it. On a single LLM instance, we run synthetic workloads following a measured tool waiting time distribution and compare, on the same inputs, a baseline configuration with settings {1, 2, 4, 8}, 8, and 8 against configurations that change some of them. Because re-arrival times differ across configurations, we build a simulator that replays request processing in time order, select the candidate with the lowest predicted cost among 2,077 configurations, and validate it on new inputs. We identify three mechanisms: discrete batch alignment, KV cache survival, and prefill interference. At a concurrency of 6, absent from the bucket set, tool waiting lowered the padding ratio (0.235 to 0.120) yet increased decode execution time 1.51-fold, so padding alone did not indicate cost. On new inputs at a concurrency of 8, where the baseline reused KV in 9 of 24 re-arrivals, enlarging the maximum batch size alone cut execution cost by 8.25%, and the selected configuration, which also adjusted the bucket set, by 9.72%. Where 17 of 18 re-arrivals were already reused, the effect was 0.59%. Compile-time configurations should thus be selected by diagnosing KV reuse loss and the resulting change in execution.

Abstract: 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.

Abstract: Large language model (LLM) serving is increasingly constrained by the GPU memory consumed by key-value (KV) caches. Existing compression, eviction, and offloading techniques alleviate this pressure, but serving runtimes typically treat only the configured target KV representation as execution-ready. Under memory pressure, this target-only contract can turn KV shortage into request stalls and preemptions. We present ElasticKV, a mixed-fidelity KV runtime built on the observation that target fidelity need not gate execution. ElasticKV introduces a compact intermediate KV state, making fidelity a runtime-managed execution property. To realize this state in a paged serving runtime, ElasticKV combines (i) a pair-structured layout that turns fidelity reduction into reusable GPU capacity, (ii) a dual-mode attention backend that directly consumes the compact state while preserving the native target-only path, and (iii) pressure-aware fidelity management that adapts KV fidelity to memory pressure. Our extensive evaluation across diverse workloads, model families and scales, and GPU platforms demonstrates the effectiveness and generality of ElasticKV. Under high concurrency, ElasticKV achieves 3.8-4.0$\times$ lower time-to-first-token (TTFT) and 9.1$\times$ lower P90 TTFT than vLLM while preserving generation quality.

Abstract: LLM-based agents execute multi-turn workflows with interleaved model inference and tool calls, making efficient serving increasingly important. However, evaluating serving optimizations is challenging because identical tasks can produce different execution trajectories. Changes in generated tokens can alter subsequent prompts, tool calls, and reasoning turns, making it difficult to distinguish system improvements from workload variation. Greedy decoding does not guarantee identical outputs, while replaying only sequence lengths loses token information that affects prefix caching and mixture-of-experts (MoE) routing. To address these problems, we propose AgentReplay, a configurable trace record-and-replay framework for agent serving. AgentReplay records input/output tokens, expert selections, request dependencies, and tool durations. During replay, it forces the recorded output tokens while performing normal autoregressive computation, with optional controls for MoE expert selection and tool delays. This allows different systems to execute the same recorded workload while retaining their own batching, scheduling, and parallelization decisions. We further separate trajectory generation from performance evaluation, enabling compatible smaller models to replay long-horizon traces collected with more capable models. Our experiments show that token-wise replay in AgentReplay effectively eliminates workload variation that greedy decoding and length-wise replay cannot avoid, enabling fairer performance comparisons across serving configurations.

Abstract: Inference in LLMs is conventionally a fixed-depth, fixed-order forward pass through every layer, regardless of how difficult the input is. The human brain does not work this way: using the thalamus as a central hub, it routes information flexibly to all regions of the cortex according to demand. Li et al. (2026) recently showed, with a system they call program-of-layers (PoLar), that transformers can be given an analogous flexibility if their layers are treated as a library of functions rather than a fixed sequence. Performance improves over the standard forward pass when each input is dynamically routed through an adaptive sequence of skipped or repeated contiguous layer blocks. We reconstructed PoLar's diagnostic MCTS in more detail than the original paper and applied it across 5 models. We reproduced several of PoLar's findings: skipping outperformed the standard pass, repeating outperformed skipping, and combining both outperformed either alone. Shorter programs sufficed for easier questions, while harder questions required more layer repeats. However, we failed to replicate the main claim regarding their learned router for single-shot inference: its top-ranked prediction consistently collapsed back to the standard pass, even though its top-k predicted programs, taken together, did show a real accuracy gain. Beyond reproduction, we find that a small number of generic programs are enough to solve most of the questions. We also provide a much deeper analysis of these programs' structure and robustness: for example, we found that programs that correct errors are highly brittle: undoing even a single edit inside a program typically breaks the correction. Connecting this to the brain's routing mechanisms, PoLar mirrors principles of thalamo-cortical coordination between cortical-area-like transformer layers. We publicly release the code at https://datexis.github.io/RE-PoLar/

Abstract: Large-scale agentic training and evaluation with large language models (LLMs) rely on isolated, stateful execution environments in which models inspect repositories, invoke tools, execute commands, and interact with task-specific services. These workloads create sandboxes in large bursts, span heterogeneous functionality and isolation requirements, retain state across long interactions, and draw from large image corpora with limited reuse. Supporting them therefore requires an elastic execution platform rather than a single sandbox runtime. This report presents DeepSeek Elastic Compute (DSec), a production sandbox platform that exposes FnCall, container, microVM, and full-VM sandbox backends through a unified SDK. DSec coordinates placement and lifecycle management across the cluster, composes environments from independently versioned layers, combines memory sharing, reclamation, and CPU scheduling for high-density execution, and loads image data on demand from Fire-Flyer File System (3FS), a cluster-wide distributed filesystem. DSec is co-designed with the reinforcement learning (RL) framework, decouples stateful rollout execution from preemptible GPU training, coordinates sandbox lifecycle with training to preserve rollout state while reclaiming idle resources, and mitigates agent misbehavior such as reward hacking. A single production-scale unit of DSec spans around 160 nodes, serving about 3 million sandboxes per day; in production, it supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. Our evaluation and deployment experience show that these mechanisms reduce environment setup and image-distribution overhead, improve memory efficiency, and preserve latency-sensitive performance under high-density overcommit.

Abstract: The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.

Abstract: Distributed model training at scale is frequently interrupted by transient network failures, conventionally forcing cluster managers to abort all processes and roll back to the latest checkpoint. While periodic checkpointing provides durability, frequent snapshotting introduces severe storage backpressure: our measurements on a 1.216B-parameter decoder reveal that per-update asynchronous checkpointing incurs up to a +656.7% latency overhead, consumes 32.5 GiB of host memory, and generates 3.39 TB/hour of storage traffic. To eliminate this overhead, we present AccelPact, a parallel runtime enabling zero-I/O in-memory fault recovery for sharded distributed training. When communication fails at a committed optimizer step, device memory remains quiescent and uncorrupted, yet continuation fails because frameworks like PyTorch FSDP cache internal communication handles across module wrappers and parameter hierarchies. AccelPact resolves this dependency invalidation via a non-invasive reference-rebinding mechanism coordinated by an out-of-band Gloo consensus protocol. On 16 NVIDIA RTX 5880 GPUs training full-parameter Mistral-7B, AccelPact eliminates checkpoint replay, yielding a 1.197x whole-run goodput improvement over cold restart and 1.194x over NVRx checkpoint restoration at checkpoint age 5, rising to 1.698x at age 18. Across ten successive fault injections, all 16 ranks maintain bit-identical parameter state with zero numerical drift. Across 4-to-16 GPU cluster topologies, reference rebinding executes in constant time (0.493-0.518 ms). Operating directly on native communicator instances, AccelPact requires zero application-code modifications and avoids compiler graph breaks under torch.compile, providing an efficient foundation for resilient deep learning.

Abstract: Long-context LLM inference is bottlenecked by attention, whose repeated KV-cache reads make decoding memory-bound. Self-speculative decoding alleviates this by drafting tokens with sparse attention and verifying them with full attention, but existing batched methods remain synchronized: all requests in a batch share a single draft-verify schedule, even though the optimal draft length varies widely across requests and changes dynamically within each request. We propose ASPIRE, a non-synchronized batched self-speculative decoding framework built on three components. First, a unified mixed forward allows drafting and verifying requests to coexist in the same batched forward pass, removing the need for global draft-verify phases. Second, a lightweight online speculation scheduler uses per-request acceptance-rate estimates and a batch-aware cost model to let each request independently choose when to verify. Third, an intra-draft refresh layer performs full attention at a single designated layer during drafting, updating the sparse context at every draft step to reduce staleness during drafting. Across three models and five reasoning and long-context benchmarks, ASPIRE achieves $1.70$-$4.58\times$ speedup in decoding throughput over autoregressive baselines and improves average speedup by approximately $27\%$ over the strongest prior self-speculative baselines.

Abstract: GPU high bandwidth memory is scarce and expensive, and KV caches consume much of it as chats, agent loops, and document question answering accumulate state. Systems such as Mooncake, LMCache, FlexGen, InfiniGen, and AttentionStore extend GPU memory with CPU DRAM and SSD. The harder question is which blocks belong in each tier, when to move or evict them, and whether prefetching helps. We study these choices in a discrete event simulator spanning GPU HBM, CPU DRAM, and SSD, calibrated against a random forest execution time predictor. We compare recency, reuse frequency, predicted reuse, and an EWMA predictor with prefetch lookahead across chat, agent, and document question answering workloads. Tiering supports 73.02 times more concurrent sessions per GPU and lowers cost per session by 62.04 times. These gains come from tier capacities of 1 plus 8 plus 64, not placement policy. Decode is compute bound at batch size one in our setup, so placement barely affects throughput. It mainly changes PCIe migration traffic and time to first token. Recency produces 2.30 times less migration traffic than reuse frequency for chat. Reuse frequency performs best for agents and document question answering. The existing predicted reuse policy is byte identical to recency, making its agent recommendation effectively recency. A genuine EWMA predictor changes behavior but still ranks behind reuse frequency on the workloads prediction was expected to help. Prefetching does not justify its bandwidth cost. Across the policy and cache size grid, even an oracle with knowledge of future requests never beats no prefetch on migration traffic. Workload specific placement can reduce data movement, but the predicted reuse and prefetch recommendations are not supported as implemented.

Abstract: Large language models (LLMs) demonstrate impressive capabilities, but their deployment presents significant efficiency challenges. Autoregressive decoding imposes substantial inference latency and under-utilizes hardware accelerators in low batch size regimes. Discrete diffusion models can generate in parallel but struggle to match autoregressive quality without many diffusion denoising steps. Long-context reasoning creates memory bottlenecks that strain even state-of-the-art accelerators. My thesis is that language models can direct their own inference execution strategy by annotating semantic dependence -- which tokens depend on which others -- in their generation. I call such models self-orchestrating language models. For each system, I design a runtime that acts on these annotations to parallelize autoregressive decoding, evict intermediate context, or derive denoising orders, achieving Pareto-optimal quality-efficiency trade-offs. I demonstrate this approach through three self-orchestrating systems. First, PASTA uses semantic dependence to parallelize autoregressive decoding, training the model to annotate which output chunks can generate independently. Second, TIP uses semantic dependence to evict intermediate reasoning steps from the KV cache, reducing memory consumption while preserving accuracy. Third, Planned Diffusion uses semantic dependence to derive a denoising order for discrete diffusion, autoregressively generating a plan that specifies which chunks to denoise in parallel.

Abstract: Single-GPU long-context inference with Mixture-of-Experts (MoE) models requires spilling the key-value cache (KVCache) to CPU memory. The spilled KV serves two complementary purposes---transferring to the GPU for attention computation, or computing in-place on the CPU---which demand opposing physical states. The optimal split between them varies with workload, yet existing KVCache abstractions offer only storage semantics over a monolithic object of a single physical state, and cannot express dynamic load balancing. We propose InplaceKVCache, the first KVCache abstraction whose format fixes each byte's physical residency at write time, so that the CPU--GPU load balance can be adjusted without moving data after placement. It realizes this as a four-region layout along two dimensions---device affinity and access pattern---turning load balancing into pure scheduling. Built on this abstraction, WriteScope splits CPU--GPU shares along the sequence dimension, and a portable roofline performance model determines the optimal CPU share as sequence length evolves, with online feedback tracking CPU cost drift. On three MoE models (DeepSeek-V2-Lite, Qwen3-30B-A3B, Mixtral-8$\times$7B) with a 32~GB VRAM budget, WriteScope supports end-to-end inference at the 1M-token aggregate scale. In the long-context regime ($\ge$8K), it achieves geometric-mean speedups of $1.5\times$--$2.5\times$ on A100 and $1.4\times$--$1.7\times$ on V100 over four reproduced baselines, while vLLM, SGLang, and KTransformers fail even with a doubled KV budget. A DeepSeek-V4-Flash case study validates composition with native sparse attention.

Abstract: Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context units (tokens or blocks) for each query. Existing trainable methods usually use a lightweight selector to score context units, followed by hard Top-K selection that blocks gradients from the language modeling loss. Consequently, these methods commonly distill layer-wise dense attention distributions. Although this encourages the selector to rank context units by dense attention weights in the original model, the ranking is not directly aligned with their impact on predictions under a fixed attention budget (i.e., the number of attended context units per query), potentially wasting the limited budget on less useful units. To address this misalignment, we propose Simple Attention Sparsification (SAS), a gated sparse attention mechanism that optimizes context ranking end-to-end with the language modeling loss. The key idea is to inject the selector's continuous scores into attention logits during training, allowing the loss to update the selector through standard backpropagation. We identify several choices crucial for this simple design to work well in practice: placing the gate inside the attention softmax in log form, using normalized softmax gates to calibrate historical context against the always-retained current block, and preserving continuous selector scores so the model learns relative priorities rather than only hard selections. To support long-sequence training, we implement a memory-efficient Triton kernel that integrates SAS into FlashAttention-style computation. Across reasoning, long-context understanding, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across attention budgets, with especially large gains under tight budgets, demonstrating more effective context ranking for downstream tasks.

Abstract: A single SM utilization percentage can make an LLM inference workload look compute-saturated while hiding how much useful work is being done. The problem is not that the counter is wrong, but that it collapses several different mechanisms into one number. This is most severe during decode, where each request contributes only one new token and dense projection GEMMs become small-row matrix multiplications. On Hopper, the bfloat16 GMMA path executes these operations in fixed 64-row matrix fragments, so small-batch decode can fill only a small fraction of each fragment with real token rows. In this paper, we profile vLLM with FlashAttention-3 and cuBLASLt on an H100 NVL across cold prefill, warm prefill, and decode, sweeping sequence length and batch size. We replace the usual single utilization number with eight counter-validated views derived from raw Nsight Compute reports, each pinned to an NCU counter or explicit formula. Together, these views map utilization gaps to concrete mechanisms - fragment fill, occupancy limits, stall signatures, wave quantization, and kernel selection - across four production models and six per-layer kernel roles.

Abstract: AI is beginning to make substantive contributions to LLM inference optimization. Existing AI optimizations are predominantly profiling-based. Profiling-bound feedback confines the search to the capabilities and performance of an existing software stack, preventing a fundamentally better architecture of LLM inference systems from being identified. To enable the AI-driven LLM inference system architecting loop, we argue that a general workload representation, a verifiable mutation space, and an implementation-independent evaluator are required. We present the RoofLang domain-specific language (DSL) that provides these features. In our evaluation, RoofLang reveals that DeepSeek V4-series models could achieve 3.5-39.5$\times$ higher peak decode throughput than other representative models. This gap is disproportionate to their total parameter counts and arises largely from compact KV-cache designs that support larger batches and reduce memory traffic. A persistent optimizer agent further discovered several new architectures that improved both throughput and interactivity of DeepSeek V4 Pro on NVIDIA B300 by 6.23-50.1%.

Abstract: As Omni-modal large language models (Omni-LLMs) take in audio, video and text together, their KV cache memory cost grows. KV cache quantization is the de facto approach in text-only LLMs, but its application to Omni-LLMs remains unexplored. In this paper, we analyze how TurboQuant, a representative rotation-based KV cache quantization method, behaves on multimodal caches and identify two critical issues: temporal key drift and heterogeneous value geometry. To address these, we propose OmniKVQuant, a training-free framework that (i) sets the key quantization range over each short window of the input stream; and (ii) rotates values separately per modality. On Qwen2.5-Omni and Qwen3-Omni, OmniKVQuant enables 2-bit KV caches while substantially preserving performance across seven audio-visual benchmarks. We further provide a fused Triton decode kernel that unpacks the 2-bit cache during attention, so no dense FP16 cache is ever built. Code: https://github.com/kaistmm/OmniKVQuant

Abstract: Distributed HPC and LLM workloads increasingly require efficient communication for scalability, yet growing data movement has become a major performance bottleneck. Communication compression can reduce this overhead and complement execution-level optimizations, but its benefits remain difficult to assess because existing benchmarks lack support for diverse backends, realistic datasets, application-specific accuracy metrics, and overlap-induced resource contention. We present CC-Bench, a lightweight, extensible, and application-oriented benchmark suite for evaluating communication compression under realistic execution conditions. CC-Bench uses declarative application-environment modeling to decouple profiling logic from communication libraries, datasets, and fidelity metrics, enabling portable cross-library evaluation. It further combines function-level interception and hardware counter monitoring to characterize per-phase latency, hardware utilization, numerical fidelity, and computation interference. With representative datasets from HPC and LLM workloads, CC-Bench evaluates three compression-enabled communication libraries on CPU and GPU clusters, revealing accuracy-performance trade-offs and bottlenecks to guide practical deployment and optimization.

Abstract: Key-value (KV) cache eviction is essential for scaling long-context inference in Large Language Models. However, existing policies predominantly rely on empirical heuristics, lacking a rigorous characterization of token utility under the inherently nonlinear softmax attention mechanism. In this work, we rethink KV cache eviction through the lens of local information geometry, modeling the attention process as a nonlinear Gaussian communication channel. By performing a first-order Taylor expansion of the attention mapping, we derive the Jacobian Information Capacity, a novel objective that explicitly captures query relevance, softmax sensitivity, and structural diversity. Guided by this theory, we introduce Jacap, a capacity-aware eviction method that utilizes softmax-aware importance weighting and statistical leverage scores for subset selection. Extensive experiments across diverse architectures and benchmarks demonstrate that \textsc{Jacap} delivers superior performance in most scenarios, particularly in high-compression regimes.