2026-10-27

Large language model (LLM) serving relies on batching to improve inference throughput, but the key-value (KV) cache grows with sequence length and limits GPU-resident concurrency. CPU offloading can expand KV capacity using host memory, yet existing systems remain largely GPU-gated: CPU-resident requests make progress only when the GPU scheduler exposes suitable overlap opportunities. As a result, host memory and CPU compute cannot act as an independent serving resource, leading to bursty CPU-side progress and inflated tail inter-token latency (ITL) under load. We present DualPace, an LLM serving runtime that turns the host into an independently scheduled decode engine. DualPace separates scheduling authority across two engines: the GPU engine executes prefills and GPU-resident decodes, while an AMX-enabled CPU engine independently batches and advances CPU-resident decodes. The engines are decoupled in scheduling but cooperative in execution: the CPU can execute a complete decode iteration locally or selectively invoke GPU Dense Assist for dense stages, while offline profiles guide hardware-specific batching and routing. Our evaluations show that, in decode-oriented scenarios, DualPace improves serving throughput by 6.9--43.2\% over NEO, a state-of-the-art open-source CPU-offloading system, and by 11.5--90.6\% over vLLM, a widely used GPU serving system, while maintaining the same level of latency. DualPace also preserves NEO's P99 time-to-first-token (TTFT) advantage over vLLM while sustaining 7.3--50.5\% higher request rates than NEO at the same level of P99 ITL.

Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To avoid recomputation, a model stores the attention keys and values of all past tokens in a key-value (KV) cache. Each layer maintains its own cache, so KV-cache memory can become the main bottleneck for GPU serving. Most compression methods operate within a single layer, by quantizing the cache entries, reducing the channel dimension, or evicting less important tokens. All three lose substantial accuracy at high compression ratios. Cross-layer methods share one compressed cache across adjacent layers to go beyond this limit. However, these methods also degrade sharply at high compression ratios. We find that the KV-cache representations of adjacent layers are quite different, whereas the input hidden states of adjacent layers are highly similar. We also find that, under the same memory budget, reducing the dimension of this shared latent loses far more accuracy than quantizing it. Based on these findings, we propose BellowsKV, a training-free method that maps the hidden states of n adjacent layers to one shared latent per token. At inference time, each layer reads its own keys and values from this latent through a layer-specific projection. We keep the latent at full dimension and compress it in two other ways: BellowsKV-Quant stores the latent at low precision, which is nearly lossless because the latent has no dominant outlier channels; BellowsKV-Evict makes one shared eviction decision per block, because the layers rank token importance very similarly. Across Llama-3.1-8B, Mistral-7B and Qwen2.5-7B, BellowsKV-Quant retains about 99% of full-cache accuracy at 3.6x compression and 97-98% at 13-14x, where the strongest cross-layer method retains at most 84% at 16x. Code: https://github.com/MorrisZJ/BellowsKV

On-policy self-distillation (OPSD) trains a student on its own reasoning trajectories using token-level supervision from a reference-conditioned teacher initialized from the same model. However, access to the reference can induce teacher behaviors that undermine independent reasoning in thinking models, leading to reference hallucinations, where the student cites an unavailable reference to justify reasoning steps, and reconsideration suppression, where the teacher discourages verification and revision at uncertain steps. We introduce Steered-Teacher OPSD (ST-OPSD) to jointly address these failure modes through activation steering. ST-OPSD extracts two directions from model-generated reasoning trajectories and applies them to the teacher's hidden states during distillation target computation, suppressing reference-invoking behavior and encouraging reconsideration with marker-family calibration. The student learns from the resulting teacher distributions without being steered. Experiments on AIME and HMMT show that ST-OPSD improves accuracy over vanilla OPSD and other interventions while restoring reconsideration and reference-hallucination rates to base-model levels.

While large-scale models achieve strong results in remote sensing scene classification, deploying them on edge devices remains difficult due to limited compute, memory, and energy budgets. Compact CNN backbones are therefore a practical choice in many real-world settings, offering efficient inference with well-optimized implementations on embedded hardware. To push this efficiency further, we investigate how far a knowledge-distilled student model can be compressed while preserving high accuracy. We propose a Distillation-Aware Compression Loop (DAC-Loop) that integrates (i) knowledge distillation applied both before and after structured channel pruning, and (ii) an iterative compression schedule that alternates structured pruning with unstructured sparsification. Across EuroSAT and NWPU-RESISC45, the DAC-Loop consistently improves the accuracy-compression trade-off relative to standard one-shot pruning and single-stage KD recovery, maintaining strong accuracy even beyond 90% parameter reduction. These results demonstrate that combining distillation and pruning in a unified iterative loop offers an effective path toward compact remote sensing models.

A looped language model runs one block of layers $R$ times per token, so its key-value (KV) cache holds $R$ planes per recurrent layer, and self-speculative decoding (draft with a few loops, verify with all $R$) reads them all. We find that prefix keys and values matter late: at equal perturbation size, loops 25-32 are $e^{15.0}$ (Huginn-0125) and $e^{12.8}$ (Raven-Llama) times more sensitive than loops 2-9, and the planes approach the last one geometrically, at 0.88 and 0.81 per loop. Whether the deepest plane alone suffices then depends on how ripe the late planes are: on Raven, storing only it is indistinguishable from a bf16-to-fp32 change; on Huginn, 1% of positions carry 74% of its error. We turn this into RIPE, which stores the deepest plane everywhere except a window of recent positions kept at full depth. On Huginn, a 64-position window removes 89% of the tail positions at 11.9% of the full cache's memory (1024-token context), and any window at least as long as the draft block makes batched verification block-independent by construction. Nine pre-registered GPU batches back these results. Simple models of acceptance, decoding cost and memory, two of them checked on held-out measurements, account for the measured self-speculative speed-ups ($2.09\times$ with the full cache, $2.58\times$ with the deepest plane) and decompose the $2.24\times$ fixed-memory throughput gain of one shared plane.

Effective discrete representation learning with VQ-VAE serves as a fundamental component of modern autoregressive generative models. However, it is widely known to suffer from codebook collapse and inefficient code utilization. To address this issue, we propose FlowVQ, a simple yet effective alternative to standard vector quantization by formulating neural quantization as a joint modeling and clustering process. Specifically, we learn the latent space using a lightweight flow matching model alongside training, corresponding to a probability flow ODE that transports an often fixed well-behaved source distribution while preserving probability mass. Based on this, we define the codebook C_1 by transporting even-quantile partitioned prior codes C_0 from the source space through the learned flow, and obtaining the Voronoi centroids under the induced distribution, leading to a structured discretization. We evaluated on image tokenization and generation, as well as on two tasks requiring high-precision tokenization, demonstrating scalability and improved expressivity of the codebook with full utilization.

Transformers achieve strong performance across modalities but are bottlenecked by the quadratic cost of attention. Sparse attention can reduce this cost, but effective sparsity must preserve useful long-range communication and map well to GPU hardware. We introduce INTRA--Interleaved Non-contiguous Token spaRse Attention--a hardware-aware static sparse attention design based on a relay-communication view under Computational Query Set (CQS) constraints. Rather than directly covering all token-to-token interactions within each sparse layer, INTRA uses complementary sparse groups to provide short mediated communication paths across layers. Scatter forms relay groups that sample broadly across local neighborhoods, while Gather reconnects these groups in subsequent layers. Under a partition model, this structure guides CQS-compatible sparse patterns with global reachability. Our CQS-aware FlashAttention-style kernel computes structured indices at runtime, loads selected non-contiguous tokens, and packs them densely for computation. On an A100, INTRA achieves a $1.53\times$ speedup in $2K$ FLUX.1-dev generation, with better measured image quality than evaluated CLEAR configurations after LoRA self-distillation. On LLaMA-3.1-8B, it improves long-context prefill efficiency with strong teacher-forced retrieval performance under matched LoRA adaptation. Attention-only benchmarks further show up to $8.4\times$ and $10.4\times$ speedups over MoBA and density-constrained NSA, respectively.