On-policy self-distillation enables the same model to act as student and teacher, with dense token-level supervision obtained by the teacher being given access to privileged information. However, conditioning the teacher on a single reference can concentrate its predictions on one reasoning trajectory and limit transfer to the student. We propose Multi-Mask Selective Self-Distillation (MASS), which approximates the unavailable marginal over diverse reference realizations using multiple independently masked views of one verified solution. MASS aggregates these complementary teacher predictions, selectively transfers reliable token-level corrections, and adaptively focuses masking on influential reference spans. Across mathematical reasoning benchmarks and model families, MASS consistently outperforms strong self-distillation baselines.
Shufan Shen, Zhongni Hou, Chao Bi et al. · cs.LG, cs.AI
Compression 96%
Abstract
On-Policy Self-Distillation (OPSD), which aligns a student model's token-level distribution with that of its self-teacher prompted by privileged context, has emerged as a promising direction to mitigate reward sparsity in sequence-level post-training paradigms. However, OPSD often fails to achieve continuous performance gains during training. In this paper, we find that the model's limited self-teaching capability underlies this phenomenon by generating unreliable guidance on certain samples. To address this problem, we propose Teacher-Evolution Self-Distillation (TESD), which enables the self-teacher to generate rollouts and receive environmental feedback based on task correctness and context independence. First, using these rewards as training objectives, we perform teacher-oriented optimization to continuously enhance the model's self-teaching capability during distillation. Second, by estimating teacher guidance reliability from these rewards, we apply adaptive student-oriented optimization to down-weight samples with unreliable guidance. Extensive experiments across multiple datasets and model scales show that TESD consistently outperforms representative token-level and sequence-level methods, highlighting the necessity of improving supervision quality in OPSD.
Enda Yu, Dezun Dong, Zhaoning Zhang et al. · cs.DC
Inference speed 96%Serving systems 85%
Abstract
Local deployment of MoE-based LLMs on CPU-GPU nodes is attractive for cost and privacy reasons, yet limited GPU memory forces most experts into CPU memory. Prefill is then compute-bound and benefits from aggressive GPU loading, while decode is I/O-bound and requires conservative PCIe transfers. These demands cannot be amortized under low batching---typical of local assistants and document tools---forcing phase-agnostic schedulers into a seesaw: favoring prefill's aggressive loading injects unmaskable I/O that stalls decode, while favoring decode's conservative transfers starves prefill's GPU of work. We observe that expert routing is predictable in phase-specific ways, allowing us to decouple the two phases and optimize prediction, scheduling, and execution independently for each phase. PDScope predicts expert activations to guide placement and prefetching, adapts these policies to each phase's bottleneck via cross-layer cost modeling, and compiles decisions into overhead-free pipelines using a static computation graph, cache prefetching, and zero-copy asynchronous I/O. On mainstream MoE models, PDScope achieves 2.5-10.3x higher prefill throughput and 2.1-6.1x higher decode throughput than existing systems, showing that phase-aware co-design unlocks efficient MoE inference under memory constraints and extends naturally to multi-batch settings.
Xin Huang, Teng Ma, Zhuoran Li et al. · cs.DC, cs.OS
Inference speed 95%Serving systems 82%
Abstract
KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We present FlatCake, a flat KV storage architecture that exposes host DRAM and NVMe SSD as peer media in a unified 64-bit address space. Exploiting the fact that KV blocks are write-once, immutable, and reused at session scope, FlatCake places each block directly onto its target medium in session-contiguous, layer-major order at creation time and never migrates it. FlatCake streams coalesced SSD reads directly into a bounded GPU window via GPUDirect Storage and decouples DRAM and SSD request queues to prevent head-of-line blocking. On a production Claude Code trace with DeepSeek-V4-Flash, FlatCake sustains a 94% cache hit rate as load increases (versus 64% for Mooncake), eliminates Mooncake's $6.5\times$--$6.7\times$ SSD read amplification, and reduces peak-load mean TTFT by $1.6\times$ while improving total token throughput by $1.6\times$. Across 8K--128K contexts on Llama-3.1-70B, FlatCake further reduces mean TTFT by up to $2.4\times$ over Mooncake and $1.8\times$ over Mooncake-GDS.
Mads Toftrup, Marco Federici, Boris van Breugel et al. · cs.LG, cs.AI
Compression 95%Inference speed 69%
Abstract
Low-bit quantization enables efficient deployment of large language models by reducing memory footprint and computational cost. Achieving high accuracy at low precision requires selecting quantization ranges, linear transformations, and weight rounding decisions. Existing approaches typically optimize these components independently using local reconstruction objectives, which fail to account for the varying impact of quantization errors on downstream predictions. Direct optimization of task-level metrics would provide a more faithful objective, but is challenging due to the discontinuous nature of quantization operations and the high computational cost of evaluating their downstream effects. In this work, we derive a Hessian-guided approximation of the output change induced by joint weight and activation quantization. Based on a tractable Fisher approximation, we formulate a unified post-training quantization framework that enables the joint optimization of quantization ranges, linear transformations, and weight rounding under a common output-sensitivity objective. The resulting method remains practical for modern large language models and consistently outperforms state-of-the-art transformation and rounding methods across both channel-wise and block-wise quantization settings.
Kwun Fung Lau, Frost Mitchell, Kanya Mo et al. · cs.LG
Inference speed 94%Serving systems 89%
Abstract
Agentic AI drives significant productivity gains across diverse domains, but concurrent long-context prefills from multiple agents can saturate GPU compute resources. Although CPUs offer a viable alternative, existing CPU attention kernels fall short in adapting to multi-tier cache hierarchies and lack effective thread scheduling. To address this, we introduce CACPU, a high-performance attention kernel featuring tiered cache blocking and an in-kernel task dispatching mechanism designed for causal full-sequence prefilling. CACPU achieves up to a 2.01x speedup over PyTorch SDPA, FlexAttention, and SGLang attention kernels. Integrated into SGLang, it reduces Time to First Token (TTFT) of Llama-3.1-8B by up to 1.74x while preserving model accuracy.
Unlike autoregressive large language models (AR-LLMs), diffusion LLMs (dLLMs) generate text by iteratively denoising many tokens in parallel, offering a promising path to efficient inference. Existing serving frameworks, however, schedule dLLM requests with policies designed for AR models, namely first-come-first-served (FCFS) admission and prefill priority. We find that both policies are ill-suited to dLLMs, wasting much of their speed. FCFS ignores that a dLLM request's decoding cost is highly heterogeneous and unknown at arrival, causing severe head-of-line blocking. Prefill priority, widely used in AR-LLM serving to keep time-to-first-token low, sharply cuts dLLM efficiency, as rounds that run only prompts leave most of the batch idle. We present DWISE, a Denoising Workload-Informed Scheduling Engine for efficient dLLM serving. DWISE schedules requests by their remaining work, measured in a block-level unit shared by prefill and denoising, so both phases share a round. To predict the unknown denoising workload, DWISE combines a lightweight prompt-based predictor at admission with a single refinement from the first denoising round's hidden states, which capture model-specific information absent from the prompt and require no additional model execution. Implemented in SGLang and evaluated on five models from three dLLM families under bursty workloads, DWISE reduces mean end-to-end latency over SGLang by 5.4$\times$ on average and up to 14.9$\times$ while preserving task accuracy. The code will be released publicly.
Prefill and decode are complementary in resource usage, but often divergent in parallelism preferences. Executing the two phases separately preserves their preferred parallelism but leaves complementary resources underutilized, while co-batching improves utilization but often requires a uniform parallel configuration that hurts execution efficiency. To resolve this utilization-efficiency dilemma, we propose Conflux Parallelism (XP), a parallel execution abstraction that co-executes prefill and decode with different parallelism within the same batch. Conflux Parallelism exploits operator-level Transformer structure to construct a heterogeneous execution path with different batching and parallelism strategies across operators, combining the best of co-batching and phase-specific execution. Building on Conflux Parallelism, we design ConfluxServe, an SLO-aware serving system that maximizes cluster goodput under TTFT and TBT constraints. Evaluation on production workloads shows that ConfluxServe improves serving goodput by 1.26-1.62x over the strongest baselines.
Linear attention stores historical information in a recurrent state of fixed size, providing a compact memory representation for long-context inference. Reducing state precision further lowers per-request storage requirements, but quantized writeback errors enter subsequent recurrences, degrade prediction quality on long-reasoning tasks, and can be accompanied by repetitive generation and excessively long outputs. We propose Global Gradient-aware Allocation (GGA) and the Pure-Write Sliding Window (PW) to address precision allocation and recent-information maintenance. GGA combines candidate quantization residuals with gradient sensitivity to future prediction loss to allocate head bit widths under a model-wide mixed-precision budget. PW independently maintains recent pure writes as FP8 vector pairs, propagates them through the full GDN state transition, and merges them into the low-bit historical state only after they leave the window. Across seven evaluations on Qwen3.5-4B/9B, GGA allocation alone improves over STEPQuant@4 by 19.48/29.84 percentage points. Adding PW-S8 yields seven-task macro-averages of 69.83%/78.63%, improving over STEPQuant@4 by 38.39/41.21 percentage points, respectively. Both precision reallocation and window maintenance improve task scores while mitigating excessively long outputs: at @6, GGA reduces mean generation length, averaged equally across tasks, by 10.45%/14.65% relative to STEPQuant; adding PW-S8 to GGA@4 further reduces mean generation length by 32.32%/36.77%. Comparisons at similar storage budgets further show that using additional capacity to maintain recent writes yields higher task scores and shorter mean outputs than solely increasing the bit width of the accumulated state.
Neehal Tumma, Philipp Nazari, Rohan Cherukuri et al. · cs.LG
Compression 91%Inference speed 56%
Abstract
Linear recurrences are emerging as practical alternatives to softmax attention in large language models, appearing in both standalone and hybrid architectures. However, despite their rapid emergence, low-precision inference remains much less developed for these models than for softmax attention. During prefill, chunkwise-parallel algorithms cast much of the recurrence as dense matrix multiplications, suggesting a natural path to low-precision execution. Simply quantizing these operations in their native form, however, can be fragile. In this work, we identify functional invariances that preserve the full-precision computation while making the operands more amenable to quantization. Building on this observation, we introduce LinQuant, a framework for low-precision recurrent prefill based on seven such invariances, including rotations, diagonal rescalings, and alternative placements of recurrent decay and write factors. We derive when these transformations remain valid across common recurrence families spanning linear, gated, and delta-rule updates, and show that they provide complementary degrees of freedom for improving low-precision accuracy. Finally, we translate this framework to production by augmenting SSDi8, an existing low-precision prefill pipeline for Mamba-2, with the applicable invariances. LinQuant improves perplexity at matched latency on Mamba-2 1.3B, 2.7B, and 8B and Nemotron-H 8B, demonstrating the practical utility of invariance-based quantization for recurrent models.
Prefill-decode (PD) disaggregation isolates the two phases of large language model serving, but request-level path selection must account for prefill work, cache reuse, and decode demand. We present DolphinRouting, which learns this choice from end-to-end SLO feedback under a given P/D resource allocation. Requests are clustered by raw prompt length, cache-adjusted input lengths on both worker pools, and output-length information. Offline policy search learns one binary action per cluster. Online feature extraction, cluster assignment, and table lookup select local or remote prefill. We evaluate DolphinRouting on a real PD-disaggregated cluster using three public benchmarks, a mixed workload, and real data. Across the mixed workload and real data, DolphinRouting with $k$-means reduces P90 TTFT by 5.7--7.9\% and P90 TPOT by 2.6--10.1\% relative to Dynamo.