2026-10-20

As LLM adoption surges, scaling model serving requires unlocking stranded GPU capacity alongside deploying more GPUs. Speculative decoding brings two models into the serving process: a small draft model proposes tokens that a larger target model verifies. Yet compiling these models individually leaves their shared execution outside the compiler's optimization boundary. We introduce CONJUGATE COMPILATION, which reframes speculative decoding as a cross-model compilation problem and jointly optimizes the draft and target models within their shared execution rather than treating each in isolation. Our key insight is that cross-model compilation should be informed by the measured performance of the co-located shared execution it produces: an optimization that improves a kernel in isolation need not improve the models when they execute together. To expand cross-model opportunities, we devise operand layouts that interleave blocks from both models within common tiles, co-locating GEMM and non-GEMM work while preserving their mathematics. We realize CONJUGATE COMPILATION through a self-improving agentic search system in which an LLM edits cross-model optimization plans using the compiler and GPU execution as its environment. Each proposed plan is compiled and evaluated through its co-located shared execution, whose measured score, together with kernel-level verdicts and compiler feedback, informs subsequent proposals, while Monte Carlo tree search accumulates rewards to guide exploration. Across six single-GPU speculative-decoding configurations on NVIDIA B300 and RTX 5090 GPUs, CONJUGATE COMPILATION achieves a geometric-mean per-round speedup of 1.37x over conventional execution. Interleaved kernels raise this speedup to 1.45x, reaching 1.59x in the best case. CONJUGATE COMPILATION also improves GPU utilization on average by 1.46x on RTX 5090 and 1.50x on B300.

The growing adoption of large language models (LLMs) has significantly increased the compute and memory demands of inference serving. A key challenge arises from the mismatch between compute-bound prefill and memory-bound decoding: when executed together, the two phases interfere, while separating them via prefill-decode (PD) disaggregation leads to substantial resource underutilization. Specifically, prefill leaves memory bandwidth underutilized, whereas decoding underutilizes compute resources. Prior approaches, such as GPU Streaming Multiprocessor (SM) partitioning, attempt to improve utilization by spatially dividing GPU resources. However, these approaches remain fundamentally limited, as memory-bound decoding still requires a large number of SMs to saturate bandwidth, reducing the compute resources available for prefill and resulting in an inherent trade-off. To address this limitation, we propose DuoPD, a hybrid PD-disaggregation architecture that integrates lightweight Processing-Near-Memory (PNM) units into the HBM logic die. Unlike prior PIM/PNM designs that rely on costly internal bandwidth scaling and offload only partial decoding operations, DuoPD enables full decoding execution on PNM by leveraging bandwidth sharing with the GPU over a shared HBM interface. This allows compute-bound prefill and memory-bound decoding to execute concurrently with minimal interference, improving both compute and memory utilization. To realize this design, we develop a compact PNM microarchitecture, a TSV-level bandwidth sharing mechanism, and a hierarchical data mapping scheme tailored for hybrid GPU-PNM execution. Evaluations on server-scale LLMs demonstrate that DuoPD achieves up to 2.23x higher end-to-end throughput than conventional PD-disaggregation.

Autoregressive language models emit one token per forward pass. This causes high inference latency, in particular as models are scaled up for quality. In this paper, we introduce Parallel Language Models (PLMs), which generate mutually consistent tokens in one forward pass by routing dependencies through latent auxiliary variables. PLMs are trained using a variational bound that reduces to an autoregressive loss plus a parallel decoding loss. Empirically, PLMs are on a faster and better Pareto frontier than autoregressive and discrete diffusion baselines. Pretrained PLMs achieve near-autoregressive pretraining loss on model scales of 80M to 3B parameters, and speedup follows a power law in model size. When post-trained, our 1.3B model matches or exceeds an identically trained autoregressive model on math, coding, and instruction-following benchmarks, speeding up inference $2.3\text{-}5.7\times$. Together, we show that parallel generation can be trained from a single training objective at autoregressive quality. We will make our code available at https://github.com/mandt-lab/plm.

Teacher takeover in on-policy distillation (OPD) can produce incorrect responses even when the student and teacher independently answer the same problem correctly, a failure we quantify using Collaboration Failure Rate (CFR). To address it, we propose Contextual Digest for On-Policy Distillation (CD-OPD). CD-OPD structures the earlier context of teacher guidance through a teacher-generated dependency graph, curating the required problem conditions and preceding reasoning as node summaries alongside the unchanged guidance in a Contextual Digest before student continuation. We further develop online meta-reweighting, which compares Digest segment gradients with a task gradient from independent student responses to improve student-only performance, favoring updates predicted to reduce task loss. Across eight benchmarks, CD-OPD improves average accuracy by 1.32--7.05 percentage points over representative existing OPD methods. It also reduces CFR by 27.4\%--56.4\% relative to existing takeover-based OPD methods.

Multi-teacher on-policy distillation (MOPD) aggregates the capabilities of multiple domain-expert teachers into a single student model. However, all domains in MOPD share the student's finite budget of optimization steps and rollouts. Fixed uses constant domain sampling ratios based on corpus sizes; by contrast, D$^3$-MOPD adjusts the ratios using the teacher-student reverse-KL divergence and its recent decrease. However, favoring large recent decreases can direct more training toward domains that already improve quickly, while domains that improve more slowly may receive too few examples. To address inflexible allocation and the risk of giving insufficient training to domains that improve more slowly, we propose Nash-MOPD, an adaptive domain scheduler based on a local minimax game using existing reverse-KL statistics. The equilibrium maximizes the smallest predicted proportional reduction under the local model and gives more prompts to domains that need more examples for comparable improvement. Consequently, Nash-MOPD responds to changing learning progress and supports domains that improve more slowly without extra teacher queries or rollouts. At step 100, Nash-MOPD exceeds both baselines' final mathematics and code checkpoint scores on DeepSeek-R1-Distill-Qwen-1.5B and reaches Fixed's on Qwen3-8B-Base. Furthermore, it leads eight of nine reported metrics and is best or tied best on all six mathematics and code accuracy metrics, respectively.

The size of the KV cache grows linearly with context length, making it a major memory bottleneck in LLM serving. Motivated by recent progress in identifying the functional roles of attention heads, a growing body of work compresses the KV cache non-uniformly across heads. Existing methods either fix the important heads via offline profiling or select them using attention signals from the prefill stage. In both cases, however, accuracy fluctuates substantially across workloads. While investigating the cause of this instability, we find that the active heads shift between the prefill and decode stages even within a single input. Along the decode path, by contrast, the activehead composition remains stable across content tokens once format tokens are filtered out. Based on this observation, we propose PeekKV. PeekKV observes a single content token immediately after prefill - a one-token peek - to jointly obtain head-group scores and token-importance scores. Guided by these scores, PeekKV allocates KV cache budgets to head groups in proportion to their group scores and retains only the top-scoring tokens within each budget. PeekKV requires no offline profiling or training and thus applies directly to unseen workloads and new models. Across four public benchmarks, PeekKV achieves the highest accuracy in more than half of the per-model comparisons against state-of-the-art methods, and stays closest to the full-cache accuracy on RULER and LongBench. On Llama-3.1-8B-Instruct, PeekKV retains 95.7% of the full-cache accuracy at a 10% memory budget on RULER with 16K contexts and 95.9% at a 1.5% budget on LongBench.

Offline compaction of the key--value cache (KV-cache) reduces the cost of repeatedly querying a fixed context by constructing a smaller cache once and reusing it subsequently. The central challenge of compaction is deciding which cache entries to delete and which to retain. Existing methods score each candidate for deletion on its own, with the rest of the cache being held fixed. However, softmax renormalization of the attention operation couples these decisions in a nonlinear manner: removing one entry changes the attention weights of every remaining one. In this work, we introduce KV-Surgeon, which uses calibration data to take interactions into account in a tractable way. We derive a quadratic surrogate that admits closed-form expressions for jointly making the optimal greedy removal decision as well as compensation within the remaining entries. On QuALITY, LongHealth, and Long Code Arena, and at every evaluated compaction ratio, KV-Surgeon substantially advances the cost--accuracy Pareto frontier of dense and mixture-of-experts models from 4B to 30B parameters. At $50\times$ compaction, KV-Surgeon improves question-answering accuracy by up to four points over attention matching (Zweiger et al., 2026), the state of the art, at comparable cost. Lastly, we introduce an even faster variant, which also exceeds the state of the art while requiring only two fifths of its cost.

Self-distillation policy optimization provides dense token-level supervision for LLM post-training through feedback-conditioned self-teachers. However, directly matching the self-teacher can disrupt successful behaviors on correct rollouts while providing insufficient correction on incorrect ones. We propose Asymmetric Policy Optimization via Self-Distillation (APSD), which constructs correctness-dependent targets by mixing student and self-teacher distributions: interpolation preserves successful behaviors on correct rollouts, while extrapolation beyond the self-teacher strengthens correction on incorrect ones. We analyze APSD through divergence decomposition, reverse-KL barycentric geometry, and gradient analysis, highlighting the importance of gradient flow through student-dependent targets. Experiments under reinforcement learning with verifiable rewards and rich feedback show that APSD achieves strong overall performance against reinforcement-learning and self-distillation baselines and matches SDPO performance with 2.6x fewer training steps on average across Qwen3 model scales.