Dong Liu, Yanxuan Yu, Eric Jiang et al. · cs.LG, cs.DC
Inference speed 96%Serving systems 48%
Abstract
Disaggregated \textsc{LLM} serving separates prefill and decode across node pools, placing a network fabric between \textsc{KV}-cache computation and consumption. This breaks a key assumption of classical cache policies: a miss no longer incurs only local recomputation, but also network transfer to the decode node. Likewise, admitting a block to the global cache requires an additional transfer, so poor retention decisions waste both memory and bandwidth. We present \textbf{KVLearn}, a learning-based framework for cost-aware \textsc{KV}-cache retention in disaggregated serving. KVLearn includes: (i) a lightweight Prefix Reuse Predictor (\textsc{PRP}) that estimates reuse probability from structural and temporal prefix features without modifying model weights; (ii) a Cost-Aware Retention Score (\textsc{CARS}) that converts reuse probability into keep/admit decisions using recomputation, transfer, and storage costs; and (iii) an Adaptive Threshold Controller (\textsc{ATC}) that adjusts admission thresholds online from observed hit rates and memory pressure. We integrate KVLearn into a globally disaggregated serving topology and evaluate it on text and multimodal workloads, including large image/video-derived \textsc{KV} blocks with heterogeneous reuse patterns. KVLearn reduces end-to-end time-to-first-token by up to 56% versus No-Cache, 38% versus LRU-Pool, and 33% versus Mooncake-style baselines, while cutting inter-node \textsc{KV} transfer volume by up to 53% versus LRU-Pool. On MM-Session, throughput remains within ${\sim}5%$ of oracle. Code is available at https://github.com/FastLM/KVLearn.
Speculative decoding accelerates LLM inference, but deploying it on CPUs introduces severe systems bottlenecks: context-swapping latency between draft and target graphs, and memory allocation overhead during KV cache rollbacks. We introduce SubTernary, an execution framework that eliminates these costs through Unified Graph Caching and Transactional KV-Cache Isolation. By compiling both models into a single memory arena with a persistent thread pool and routing speculative tokens to a static scratchpad for zero-copy pointer resets, SubTernary avoids dynamic reallocation and thread synchronization entirely. We integrate a 256-bit AVX2 ternary SIMD kernel as a low-bit case study. Ablations reveal our systems runtime drives 35.0% of the throughput gains, while the ternary kernel adds 6.0%. SubTernary achieves a $1.35\times$ speedup over llama.cpp speculative decoding and $3.99\times$ over autoregressive decoding, maintaining $O(1)$ rollback latency.
Nicol\'as Mauricio Cuadrado \'Avila, Ivo Gollini Navarrete, Martin Tak\'a\v{c} et al. · cs.LG
Compression 96%Inference speed 57%
Abstract
Non-uniform allocation can improve post-training pruning of large language models, yet quality can still degrade sharply at high sparsity. To control degradation as sparsity increases, we introduce LADDER (Layer-Wise Allocation via Damage-Driven Evaluation and Refinement), which allocates sparsity incrementally. At each step, it evaluates pruning an additional fraction of weights from each eligible layer and commits the layer update that causes the least end-to-end degradation per removed parameter. This process produces non-uniform sparsity profiles and exactly nested masks across multiple budgets. We evaluate LADDER on five OPT, Llama, and Qwen models across 10-90% sparsity, using SparseGPT as its candidate-generation backend and comparing against uniform and adaptive allocation baselines. Results reveal a clear operating-regime trade-off: LADDER is comparable or slightly weaker at benign sparsity but becomes substantially more robust from 70% onward. This high-sparsity advantage largely persists after posterior distillation across four of five models and translates into improved retention of aggregate zero-shot accuracy.
Time series anomalies take diverse forms, and individual detectors encode different anomaly scoring principles that rarely cover all anomaly types. Recent unsupervised time-series anomaly detectors provide complementary scoring signals, yet deploying them as an ensemble preserves the inference and maintenance cost of every individual detector. We study whether heterogeneous teacher detectors can be distilled into a single lightweight detector in the unsupervised setting, where no anomaly labels are available to define a shared target. AnoMix addresses this by extracting weak pairwise supervision from the training-set anomaly scores of frozen teachers. It normalizes each teacher score sequence independently and trains the student with CoRank, a multi-teacher ranking-distillation objective that converts teacher score gaps into pairwise supervision through teacher consensus. Across seven benchmark datasets, AnoMix successfully integrates complementary teacher scoring principles into a single deployable model, improving over individual detectors and static score ensembles while using a compact student with under 80k parameters on the benchmark settings. A zero-shot variant further shows that the distilled ranking signal transfers to unseen datasets and can outperform large pretrained time-series foundation models for anomaly detection.
Yifan Yang, Hanling Yi, Dongquan Yang et al. · cs.LG, cs.CL
Compression 95%Inference speed 84%
Abstract
Mixture-of-Experts (MoE) architectures scale model capacity efficiently, yet their substantial memory footprint poses a major barrier to deployment. Expert pruning reduces this footprint, but most methods focus on which experts to remove and prune every layer equally. We propose SPARE (Sparsity Allocation via Routing Entropy), a training-free, selector-agnostic method that decides how many experts to remove from each layer using routing entropy, the entropy of tokens' normalized Top-K gate weights on calibration data. When a token's gate weight is spread evenly, removing one active expert discards little of it, so SPARE prunes layers with more diffuse routing more aggressively. A perturbation bound under random pruning formalizes this intuition, and controlled interventions with a fixed selector link concentrated routing to larger output changes. With the same selector and budget, SPARE improves average accuracy over uniform allocation on Mixtral, Qwen3, and DeepSeek models, especially at high pruning ratios. At 50% expert pruning, it raises the eight-task average of Qwen3-30B-A3B from 57.2% to 63.8%, and the math, code, and science average of DeepSeek-R1 from 80.1 to 82.8.
Long contexts increase the memory cost of the key-value (KV) cache in large language model (LLM) inference. Sparse attention reduces attention computation and memory traffic by selecting historical tokens, but the retained KV cache must remain available for later decoding steps. We present UlaCache, a KV cache management system built around a unified logical address space spanning GPU KV pools and optional CPU memory. Stable logical references let the inference engine allocate and select KV independently of physical placement. Common allocation and sparse-access interfaces support data-parallel (DP) and tensor-parallel (TP) attention. For distributed storage, local GPU buffers retain selected KV across decoding steps, so only missing entries require transfers. When CPU storage is enabled, Parallel Bucketed LRU (PB-LRU) manages a GPU pool shared across requests. It updates recency metadata and selects eviction victims in parallel. UlaCache transfers KV and then updates the address mappings. With speculative decoding enabled under matched CPU and GPU memory budgets, UlaCache achieves peak decoding throughput speedups of 1.16 $\times$ (DP) and 1.70 $\times$ (TP) on H200, and 1.73 $\times$ (DP) and 2.51 $\times$ (TP) on B300, compared with the state-of-the-art baseline.
Wenhai Wan, Teng Zhang, Shao-Yuan Li et al. · cs.LG
Compression 93%
Abstract
Real-world datasets often follow a long-tailed distribution, making generalization to tail classes difficult. We revisit this problem through the lens of shortcut learning, where models prefer the easiest predictive cues (e.g., background or textures) over object-centric semantics, especially under scarce and biased supervision. We find that this tendency is amplified for tail classes: limited examples often share similar contexts, making non-semantic signals highly correlated and thus tempting shortcuts, whereas head classes with diverse appearances and environments encourage more stable object-focused representations. Motivated by this observation, we propose Shortcut-Resistant CAM Distillation (SRCD), a plug-and-play framework that transfers object-focused explanations from head to tail classes. SRCD operates in the Class Activation Map (CAM) space, where a CAM provides a class-specific spatial evidence map for a prediction. SRCD aggregates CAMs from a small set of head-class candidates into a shortcut-resistant teacher using an energy-model weighting based on coherence and concentration, and distills it to the tail-class CAM. We provide a theoretical analysis that quantifies shortcut reliance as shortcut-region evidence mass in CAM space and shows that SRCD suppresses tail shortcuts. Extensive experiments on long-tailed benchmarks consistently improve strong baselines. The code is available at https://github.com/Haifeng3/SRCD.
Jay Agarwal, Siddharth Khare, Dhruv Kumar · cs.LG, hep-ex, hep-ph
Compression 92%
Abstract
Jet tagging is one of the most important applications of machine learning at the Large Hadron Collider. Recent models build in more physics knowledge or rely on large-scale pretraining, but the experiments place strict limits on their inference cost. Knowledge distillation can transfer their performance to small students that are cheap to run. We ask which properties of a teacher help the student, how small a student can get and how close it can come to the state of the art. On a jet tagging benchmark, we distil six teachers that differ in architecture, symmetry and pretraining, including two foundation models, into five small students. Every teacher improves every student, and the gain depends mainly on the strength of the teacher. We then shrink our two strongest distilled students, Transformer and L-GATr-slim, in width and in depth. With 20 thousand parameters, a distilled L-GATr-slim outperforms every published jet tagger trained from scratch at a fraction of the inference cost of every teacher. Even with two thousand parameters, it matches ParT, a model with a thousand times as many parameters. We also study what the students learn from their teachers. Our benchmark of inference cost shows that students of the same size can differ by almost an order of magnitude in speed. We release the code and the weights of all trained models.
Yubing Wu, Wei Chen, Junmei Yang et al. · cs.LG, cs.AI
Compression 92%Inference speed 67%
Abstract
Direct preference optimization (DPO) incurs training costs in output scoring, sequence execution, and backpropagation, so reducing supervision alone does not necessarily save computation. We introduce a unified framework for stochastic pruning of the DPO computation graph that independently samples response positions for scoring, response prefixes for execution, and preference branches for backpropagation. By making these selections before the corresponding operations, the framework omits computation at each stage of a training step. Our analysis characterizes the objective changes induced by position and sequence pruning and uses inverse-probability weighting to obtain a conditionally unbiased single-branch gradient estimator for the sampled objective. Controlled measurements show that position and sequence pruning primarily reduce peak memory, with most output-memory savings attributable to an implementation that retains full response supervision. Branch pruning yields a $1.426\times$ step-time speedup over a concatenated DPO baseline and reduces training time by about 30\% relative to stock DPO across two model families, with slightly lower mean accuracy and benchmark-dependent changes in open-ended scores. These findings relate the resource benefits of stochastic graph pruning to the computation omitted and the resulting changes in the training objective.
Junhyeok Park, Jungeun Yoon, Kyungmin Kim et al. · cs.AR, cs.DC
Inference speed 91%Serving systems 39%
Abstract
Large language model (LLM) inference exhibits distinct computational characteristics across its two phases: prefill is compute-intensive, whereas decode is more memory-bound. As a result, AI accelerators achieve varying levels of efficiency across the two phases. GPUs are well suited to prefill, but are often underutilized during decode. By contrast, NPUs can utilize their compute resources more effectively during decode. Motivated by this observation, we explore a heterogeneous architecture that integrates GPUs and NPUs, mapping prefill to GPUs and decode to NPUs to achieve high performance and efficiency. However, realizing such heterogeneous integration is challenging because LLM phase disaggregation requires KV cache transfer from prefill to decode (that is, from GPUs to NPUs), which is complicated by differences in inter-device communication mechanisms and KV cache tensor layouts. To address this, we propose HAWK, a transfer mechanism that enables rapid KV cache transfer across heterogeneous devices. HAWK establishes a remote direct memory access (RDMA) path for direct cross-device transfers while handling inter-device tensor layout discrepancies as part of the transfer process. In particular, HAWK allocates tensor-layout-aware RDMA buffers on the NPU side to match the target tensor layout. Leveraging these buffers, HAWK orchestrates data transfer from a conventional contiguous GPU RDMA buffer to the corresponding NPU buffers, allowing transferred tensors to be consumed directly without additional data rearrangement while maintaining transfer efficiency. We implement and evaluate HAWK on a system built with commercially available GPUs and NPUs, demonstrating performance within 7.88% of a GPU-only system under the same SLO constraints while achieving up to 1.26$\times$ higher power efficiency and 1.72$\times$ higher cost efficiency.