LLM Inference Efficiency

KV cache eviction methods typically use a single retention-rule family throughout a model, making eviction-method identity a model-level design choice. Yet Transformer layers differ substantially in their attention behavior, representations, and sensitivity to compression, suggesting that a uniform rule may overlook useful layer-wise structure. This raises a basic question: should eviction methods themselves vary across layers? We investigate this question by composing existing eviction methods across Transformer layers and systematically exploring the resulting layer-wise design space. Using simple offline profiles, we construct fixed routes and study how their quality varies with method placement and cache budget. On LongBench, heterogeneous routing improves performance on a majority of tasks over homogeneous policies at the same cache budget. Even when method counts are held fixed, the profile-guided placement ranks second among 100 evaluated assignments, demonstrating that routing quality depends strongly on where methods are placed. Moreover, the same fixed route outperforms the best of nine standalone baselines across all five tested cache budgets. Together, these results establish layer-wise method composition as an exploitable, placement-sensitive design dimension for KV cache compression.

Researchers and practitioners working with large language models face a fragmented landscape: local models are free and private but hardware limits the model size and context windows a researcher can use; institutional HPC centers offer powerful GPU resources at no marginal cost and keep data within institutional boundaries, but operate behind firewalls and are designed for batch jobs rather than interactive use; commercial cloud APIs provide frontier-model quality on demand but impose significant cost and data retention policies unsuitable for sensitive research data. No existing system unifies all three. STREAM (Smart Tiered Routing Engine for AI Models) addresses this gap with four contributions: (1) a three-tier routing architecture combining local, HPC, and cloud inference with a local LLM-based complexity judge; (2) a dual-channel HPC streaming architecture that separates the Globus Compute control plane (authentication and job dispatch) from a WebSocket relay data plane (token delivery), enabling sub-second TTFT (0.54 s median, 21.1x over batch mode's 11.40 s) through institutional firewalls without VPN or firewall rule changes, with end-to-end AES-256-GCM encryption ensuring the relay operator cannot read token payloads; (3) tier-aware context summarization that prevents long conversations from forcing simple queries onto expensive tiers; and (4) an HPC-as-API proxy mode that exposes HPC inference as an OpenAI-compatible endpoint callable from any standard client with no HPC expertise, a deployment pattern made practical only by the sub-second TTFT of contribution (2). Llama 3.2 3B achieves 85.1% free-tier retention on a 1,200-query benchmark spanning ten domains. Measured TTFT: 0.26 s local, 0.54 s HPC (relay), 1.68 s cloud.

Long Chain-of-Thought (CoT) reasoning improves LLM problem-solving but is computationally expensive due to sequential token generation. While recent works explore reasoning in continuous latent spaces to bypass discrete token generation, they often struggle with training stability and fail to scale to complex, long-horizon tasks due to lack of supervision signal. We propose SuperThoughts, which compresses pairs of consecutive CoT tokens into single latent representations and decodes two tokens per step via a lightweight Multi-Token Prediction (MTP) module. This preserves discrete token supervision at training time while doubling throughput at inference time. We finetune Qwen2.5-Math-1.5B-Instruct, Qwen2.5-Math-7B-Instruct, Qwen2.5-Math-14B-Instruct, and evaluate on MATH500, AMC, OlympiadBench, and GPQA-Diamond. With a confidence-based adaptive mechanism that falls back to standard decoding when uncertain, SuperThoughts achieves $\sim$20--30% CoT length reduction while maintaining accuracy with minimal degradation (1-2 points accuracy drop on most tasks).

Speculative decoding speeds up LLM inference by using a draft model to generate tokens, with an acceptance-rejection scheme that ensures that the output matches the target distribution. Adapting this to continuous diffusions is difficult because speculative sampling requires drawing from a residual distribution. While straightforward in discrete spaces, efficiently sampling this residual in continuous space is non-trivial. Consequently, existing diffusion adaptations either use computationally inefficient sampling techniques or rely on an alternative scheme. In this work, we introduce a novel scheme that efficiently implements the original speculative sampling mechanism for diffusion models. Our approach offers a critical advantage over current methods: it enables us to adapt block verification from LLMs to diffusions -- which provably improves the acceptance rate of drafts. Furthermore, we formalize and analyze the Free Drafter, a heuristic self-speculative drafter for diffusions that requires no training. By enabling block verification, our Free Drafter yields up to a 6.3% speedup over existing speculative methods with no additional training and negligible overhead beyond the existing parallel verification pass.

Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.

Diffusion large language models (dLLMs) accelerate generation by denoising multiple tokens in parallel, making them attractive for latency-sensitive mobile inference. However, repeated denoising introduces substantial computation on smartphones. Mobile neural processing units (NPUs) offer high-throughput dense matrix computation, but efficiently exploiting them remains challenging: token commitment shrinks per-block effective workloads, token revision complicates KV cache reuse, and limited NPU-visible address space incurs costly remapping and data transfer overheads. In this paper, we propose llada.cpp, the first NPU-aware inference framework for accelerating dLLMs on smartphones. llada.cpp aligns block-wise dLLM inference with the execution characteristics of mobile NPUs through three techniques. (1) Multi-Block Speculative Decoding fills the shrinking workload in late-stage current-block decoding with speculative future-block tokens. (2) Dual-Path Progressive Revision keeps committed tokens revisable until stable and refreshes unstable tokens through a CPU-side path without stalling dense NPU execution. (3) Swap-Optimized Memory Runtime compacts NPU-visible address layouts and overlaps data staging with NPU computation to reduce remapping and transfer overheads. We implement llada.cpp as an end-to-end framework and evaluate it across diverse hardware platforms and dLLM workloads. llada.cpp reduces LLaDA-8B generation latency by 17x-42x over the CPU baseline with prefix KV cache reuse, while preserving generation quality.

Large language models (LLMs) exhibit exceptional general language processing capabilities, but their memory and compute costs hinder deployment. Ternarization has emerged as a promising compression technique, offering significant reductions in model size and inference complexity. However, existing methods struggle with heavy-tailed activation distributions and therefore keep activations in high precision, fundamentally limiting end-to-end inference acceleration. To overcome this limitation, we propose TWLA, a post-training quantization (PTQ) framework that achieves 1.58-bit weight compression and 4-bit activation quantization while maintaining high accuracy. TWLA comprises three components: (1) Euclidean-to-Manifold Asymmetric Ternary Quantizer (E2M-ATQ) minimizes layer-output error under weight ternarization via a two-stage optimization from Euclidean initialization to manifold relocation; (2) Kronecker Orthogonal Tri-Modal Shaping (KOTMS) applies a Kronecker-structured orthogonal rotation to reshape weights into ternary-friendly tri-modal distributions, while the shared rotation statistically suppresses activation outliers; and (3) Inter-Layer Aware Activation Mixed Precision (ILA-AMP) explicitly introduces adjacent-layer second-order interaction costs in bit allocation and jointly optimizes for the layer-wise disparity of activation quantization gains induced by the shared orthogonal transform, preventing cascades triggered by a few weak layers. Extensive experiments demonstrate that TWLA maintains high accuracy under W1.58A4, while delivering significant inference acceleration. The code is available at https://github.com/Kishon-zzx/TWLA.

As large language models (LLMs) are increasingly deployed across heterogeneous hardware with varying resource constraints, the ability to adaptively manage the trade-off between performance and efficiency without retraining is critical. We propose Drop-by-Drop, a novel multi-bitwidth post-training quantization framework that enables inference-time precision control over LLM weights from a single trained model. Our method is theoretically grounded in information theory and successive refinement. We establish that LLM weights, which commonly follow a Gaussian distribution, can be optimally reconstructed with increasing fidelity as additional bits are incorporated, under a weighted mean squared error distortion motivated by LLM loss functions. To realize this in practice, Drop-by-Drop incorporates Matryoshka-style supervision into the loss function, exploiting the structure of additive codebooks. Drop-by-Drop produces a single model where ordered subsets of codebooks yield accurate partial reconstructions at each precision level. This approach significantly reduces storage and memory overhead by allowing a single checkpoint to serve multiple bitwidths, while maintaining competitive perplexity and accuracy across major architectures, such as Qwen, LLaMA, Gemma, and Mistral.

Arbor is a multi-agent framework that introduces structured tree search as a cognition layer for autonomous agents operating in large, stateful action spaces. Prior autonomous optimization systems operate on isolated targets with stateless evaluation. Arbor instead maintains an explicit search tree of scored hypotheses that serves as the shared working memory across agents, evolving with every measurement, treating failures as diagnostic signal that reshapes subsequent exploration, and expanding as prior successes shift the bottleneck distribution. We validate Arbor on full-stack LLM inference optimization, a domain where achieving peak performance has historically required coordinated effort from engineering teams across the application, framework, compiler, kernel, and hardware stack. Arbor pairs an Orchestrator agent, which drives optimization by delegating to Domain Specialists across the inference stack, with a Critic agent that safeguards stability through root-cause analysis, introspection, and measurement validation -- a checks-and-balances architecture where neither agent can unilaterally drive the system. Agent capabilities are decomposed into hard skills (domain expertise) and soft skills (coordination protocols that determine how contributions compose), enabling fully autonomous multi-day campaigns. Arbor achieves up to 193% inference throughput-latency Pareto improvement over vendor-optimized baselines, while a single agent without the harness plateaus at +33% throughput improvement and crashes irrecoverably within hours. Arbor generalizes to multiple generations of hardware platform, and run-to-run variance is within 2 percentage points demonstrating that the method is hardware-agnostic and reproducible.

Speculative decoding (SD) addresses the high inference costs of LLMs by having lightweight drafters generate candidates for large verifiers to validate in parallel. Existing draft-verify methods use binary decisions: accept or fully recompute. Yet we find that many rejected tokens can be verified correctly by a slim submodel derived from the full verifier via intra-model routing, instead of the full verifier. This motivates our slim-verifier to handle tokens requiring moderate verification resources, reducing expensive large-model calls. We propose Verification via Intra-Model Routing for Speculative Decoding (VIA-SD), a multi-tier framework using a routed slim-verifier. Draft tokens are processed hierarchically: direct acceptance for high-confidence cases, slim-verifier regeneration for medium-confidence cases, and full-model verification for uncertain cases. Across four representative tasks and multiple model families, VIA-SD reduces rejection rates by 0.10-0.22 and delivers 10-20% speedups over strong SD baselines, while achieving 2.5-3x acceleration over non-drafting decoding. Moreover, VIA-SD is compatible with existing SD frameworks without modifying their training procedures. Our results suggest multi-tier SD as a general paradigm for scalable and efficient LLM inference. Project page: https://zju-xyc.github.io/VIA-SD-Project-Page/

Large language models (LLMs) are widely used to tackle complex tasks with autonomous workflows. Recently, reusable natural language skills have emerged as a popular paradigm to inject procedural knowledge into LLM applications. Since popular skills are often invoked repeatedly, placing their full text in every context significantly increases prefill cost and latency. While text compression techniques have the potential to solve this problem, most existing methods are designed to compress factual knowledge in documents instead of procedural knowledge, making them insufficient for skill compression. In this paper, we argue that an effective skill compression method should: 1) preserve logical dependencies among workflows and tool protocols, 2) enable lightweight, offline compression for frequently updated community skills, and 3) be adaptable to varying complexities across skills. To address this, we present SKIM (SKIll coMpression), an adaptive multi-resolution soft token compression framework for procedural skills. Depending on the complexity of each skill, SKIM creates different numbers of soft tokens that not only improve the efficiency of LLM inference, but also preserve the effectiveness of skill usage. Experiments indicate that SKIM compresses skills to 30 to 60 percent of their original token length while preserving task performance better than existing compression methods.We have released our code at https://github.com/bebr2/SKIM .

Post-training quantization (PTQ) is essential for efficient large language model inference, but reliably quantizing activations remains challenging when weights, activations, and KV caches are all quantized to 4-bit precision. A key difficulty lies in massive activations, whose extreme values dominate the activation range and amplify quantization errors. State-of-the-art methods mainly mitigate massive activations through transformation-based smoothing, such as orthogonal rotations and affine scaling, but overlook the cross-layer dynamics of the residual stream. In this paper, we show that massive activations emerge and disappear in a phase-wise pattern across network depth, triggering large residual changes. These changes cause newly injected layer-wise updates to dominate the 4-bit quantization scale and weaken historical residual information. To characterize this behavior, we introduce Jump Ratio and Historical Feature SNR. This suggests that static transformation-based smoothing cannot fully resolve dynamic quantization instability caused by cross-layer residual changes. Based on this analysis, we propose DynamicPTQ, a Dynamic Post-Training Quantization policy for phase-aware mixed-precision activation quantization. DynamicPTQ identifies quantization-sensitive layers from residual-stream dynamics and assigns 8-bit activation precision only to these layers, while keeping weights, KV caches, and other activations in 4-bit precision. It can be directly integrated with strong PTQ baselines such as QuaRot, SpinQuant, and FlatQuant. Experiments on LLaMA-2 and LLaMA-3 show that DynamicPTQ consistently improves perplexity and zero-shot QA performance under W4A4KV4 quantization, while achieving 1.05 to 1.07 times throughput improvement with modest memory overhead. These results demonstrate a practical path toward robust low-bit LLM inference.

The prefill stage of Large Language Model (LLM) inference is a growing contributor to cloud-scale energy cost. Many consumer-support and conversational prompts contain social scaffolding: politeness markers, apologetic preamble, repetition, and rapport-building language that is important for human communication but carries low marginal information for machine reasoning. We call this discrepancy the Social-Semantic Gap. We present SPSD (Sentiment Preserving Semantic Distillation), an edge-based pipeline that compresses user prompts using a 4-bit quantised Small Language Model before transmission to a cloud-deployed LLM. Evaluation on a 248-prompt corpus using Gemma-2-2B-Instruct (Q4_K_M) as the SLM and Llama-3.1-8B-Instruct as the cloud evaluation model yields a mean input token saving of 99.9 tokens per distilled call, with all 146 distilled calls yielding positive savings. Response quality, assessed by blind LLM-as-judge scoring across 121 pairs, is non-inferior to the raw path within a pre-specified 1-point margin on a 15-point rubric; the judge awarded 43 percent ties, 28 percent distilled wins, and 29 percent raw wins. Cosine similarity is mixed: mean 0.682, median 0.712, with 54.1 percent of pairs above the 0.70 reference threshold. Safety-critical domains are conservatively routed to passthrough via rule-based gates. Per-call net energy saving is estimated at 70-270 uWh under stated assumptions. SPSD shows that on-device prompt distillation can reduce cloud LLM input-token cost while preserving response quality within a practical non-inferiority margin.

With the growing demand for on-device LLM inference, edge SoCs increasingly integrate NPUs to improve performance and energy efficiency under tight power and thermal budgets. However, practical LLM deployment on current client NPUs remains difficult: widely used quantization formats such as AWQ do not map cleanly onto many existing NPU software stacks, which are often proprietary and expose limited low-level control. In this work, we present TileFuse, a close-to-metal mixed-precision kernel library for AMD XDNA2 NPUs that targets GEMM/GEMV-based operators in quantized LLM inference. TileFuse brings practical low-bit formats such as AWQ-style W4A16 and W8A16 directly onto XDNA2, rather than forcing the model to be reshaped around an NPU-specific quantization scheme. TileFuse co-designs weight layout, metadata placement, mixed-precision microkernels, and array-level dataflow. Specifically, it fuses unpacking, dequantization, and GEMM/GEMV execution into a single kernel flow, introduces an interleaved pre-tiling layout that supports GEMM dimensions up to 32K, and redesigns GEMV dataflow to utilize the full 4x8 AIE array. Across kernel-level evaluations, TileFuse improves performance by up to 121.6% for GEMM and 281% for GEMV over full-precision baselines, while delivering more than 2x performance and energy-efficiency gains over strong iGPU baselines on GEMM. In end-to-end LLM experiments on Ryzen AI laptops, TileFuse achieves up to 2.0x lower prefilling latency with more than 64.6% lower energy consumption. Together, these results show that XDNA2 is a practical target for AWQ-style edge LLM inference and that native NPU support for off-the-shelf quantization can make NPUs substantially more usable in real client deployments.

Large language model inference is bottlenecked by autoregressive decoding, where each token requires a full forward pass. Multi-token prediction (MTP) offers a promising acceleration path, but existing approaches suffer from a fundamental architectural flaw: the MTP head for the first token competes with the backbone's own language model (LM) head, leading to severe quality degradation when predictions are accepted. We identify this head-backbone competition as the root cause of repetitive and incoherent outputs in prior MTP-based acceleration methods. To address this, we propose Backbone-as-Architect, a design principle where the backbone LM head always generates the first token, and MTP heads are responsible only for subsequent tokens. Building on this principle, we introduce CLP (Collocation-Length Predictor), a lightweight span-level decision layer that predicts how many additional tokens can be safely accepted at each decoding step. CLP uses only a single linear layer (4.6K--7.7K parameters), replacing the over-engineered 1M-parameter gate networks used in prior work. Experiments on Qwen2.5 models (0.5B, 1.5B, 7B) show that CLP achieves 1.20x--1.29x speedup on 1.5B and 1.14x--1.20x on 7B, with zero quality degradation (repetition ratio < 0.02), while gate-based approaches fail to accelerate (1.07x) or produce severely degraded outputs (repetition ratio > 0.5%). We further demonstrate that shorter prediction horizons (k=2) recover 24% higher MTP head accuracy on large models, establishing a scaling-aware design principle. We identify MTP head prediction accuracy as the binding constraint on acceleration and establish a clear roadmap for future improvements.

Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing recomputation that scales quadratically with context length and becomes prohibitive for long-context scenarios. We propose Prefilling-dLLM, a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding, showing that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. On LongBench and InfiniteBench, Prefilling-dLLM achieves state-of-the-art quality among dLLM acceleration methods, and an attention kernel that parallelizes decoding over the non-contiguously cached chunk KV yields 9.1--28.0x speedup at 8K--32K contexts. We further show that beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon. Code is available at https://github.com/menik1126/Prefilling-dLLM.

Post-training quantization at the 2-bit level enables low-cost deployment and inference acceleration for large language models (LLMs). Scalar quantization (SQ) and vector quantization (VQ) are two primary quantization methods, however, the former suffers from significant performance degradation, and the latter incurs computational and storage overhead. We propose UniSVQ, a unified 2-bit quantization framework that bridges scalar and vector quantization by parameterizing codewords as an affine transform of integer lattices. This structure preserves compatibility with optimized integer kernels while retaining much of VQ's flexibility. We further introduce a data-driven block-wise fine-tuning strategy to directly minimize quantization reconstruction error. Extensive experiments across multiple LLM families and zero-shot benchmarks demonstrate that UniSVQ consistently outperforms state-of-the-art SQ methods and achieves performance comparable to advanced VQ methods, while providing higher inference throughput. Codes are publicly available at https://github.com/AI9Stars/UniSVQ.

Semi-structured 2:4 sparsity is widely supported by modern accelerators, providing up to a 2x theoretical speedup. However, its strict 50% sparsity constraint often causes non-negligible accuracy degradation under post-training pruning. Meanwhile, existing relaxed sparsity formats either require specialized compiler support or introduce runtime overheads that limit end-to-end speedup. We propose Spense, a practical hybrid sparse-dense format that splits each weight matrix into a 2:4 sparse region and a dense region. This design relaxes the effective sparsity constraint while remaining compatible with existing high-performance sparse and dense GEMM libraries, avoiding both custom compiler support and input activation expansion. Building on this format, we introduce SpenseGPT, a one-shot post-training pruning method that produces sparse and dense regions. Notably, we show that selecting the right dense regions is important, and we devise two different strategies to choose them. Experiments on Qwen3-32B and Seed-OSS-36B demonstrate that our method achieves up to 1.2x end-to-end decoding speedup on B200 GPUs with FP8 precision, while preserving accuracy. To the best of our knowledge, this is the first one-shot pruning demonstration of real-world end-to-end LLM decoding speedup from semi-structured sparse tensor cores on recent GPUs such as B200s, while maintaining model quality.

Cross-domain recommendation is a core problem in content-to-e-commerce platforms. Its objective is to leverage user interactions with content to infer potential purchasing intent on the e-commerce side, thereby enhancing conversion rates and commercial value. However, in real industrial scenarios, cross-domain recommendation faces multiple challenges: significant semantic gaps exist between different domains, and user cross-domain behavior sequences are often massive in scale and rich in noise. Although large language models (LLMs) possess powerful semantic understanding and reasoning capabilities, their millisecond-level inference latency makes direct application in online recommendation systems difficult. To address these issues, this paper introduces AIR (Atomic Intent Reasoning), an LLM-driven cross-domain recommendation framework designed for industrial-grade deployment. By migrating LLM inference to the offline phase and dynamically constructing user intent representations through efficient retrieval and composition during online operations, it achieves approximately 400* inference acceleration while maintaining semantic consistency. Experimental results across multiple public datasets demonstrate that our method achieves state-of-the-art performance in cross-domain recommendation tasks. Furthermore, large-scale online A/B testing conducted in Kuaishou E-commerce's real-world business scenarios shows that our approach delivers stable and significant improvements across multiple core business metrics, including a +3.446% increase in GMV, fully validating its effectiveness and practical value in industrial-scale recommendation systems.

Retrieval-Augmented Generation (RAG) pipelines are compute-intensive, combining embedding, retrieval, reranking, and large language model (LLM) generation. Running them entirely on-device benefits privacy, latency, and offline use, but the energy cost of CPU inference is a major barrier. We present what is, to our knowledge, the first end-to-end RAG pipeline that runs all neural stages -- embedding, reranking, and LLM generation -- on the Qualcomm Hexagon NPU of the Snapdragon X Elite. Profiling on a Dell XPS 13 laptop, we compare NPU-accelerated RAG against CPU and OpenCL/Adreno GPU baselines on indexing and query workloads. On indexing, the NPU achieves 9.1x higher embedding throughput and 12.3x less system energy. On a 120-query Wikipedia-passage benchmark, it delivers 18.1x faster LLM prefilling, 4.0x lower end-to-end query latency, and 4.0x less system energy than the CPU baseline; the same workload on the integrated GPU is 1.7x slower than CPU and uses 6.5x more energy than the NPU. A GPT-4.1 LLM-as-judge evaluation finds NPU answer quality on par with CPU and GPU within evaluator noise (mean 9.32 vs. 8.95 vs. 9.03 on a 1-10 rubric), with 86.7% of queries scoring identically across all three backends. On the Snapdragon X Elite / Hexagon class of laptop SoC, the NPU thus enables practical, energy-efficient on-device RAG without quality regression -- a sustainable path toward green edge intelligence that we expect to generalize to comparable mobile NPUs (Apple Neural Engine, Intel NPU, MediaTek APU) as their software stacks mature.

AutoMegaKernel (AMK) compiles a HuggingFace Llama-family model into a single persistent cooperative CUDA kernel that runs the whole forward pass in one launch, with no per-model hand-written CUDA. The contribution is the system, not raw speed. A frozen schedule-IR validator statically certifies deadlock-freedom and race-freedom via static graph checks (not a mechanized proof), so an unsafe agent-proposed schedule is rejected before launch: across 7,160 adversarial schedules (6,091 unsafe) it had zero false-accepts and accepted all 360 real lowerings. The same source retargets sm_80/sm_90/sm_120 from one codebase, auto-generates correct megakernels for 10 of 10 supported models, and on a real SmolLM2-135M checkpoint reproduces HuggingFace greedy decode token-for-token (perplexity match 2.5e-7). An unattended, agent-drivable autoresearch loop self-improves the megakernel over its own baseline (1.25-1.72x). A search-found int8 (W8A16) megakernel beats CUDA-graphed cuBLAS bf16 at batch-1 decode across NVIDIA's datacenter inference fleet: L4 up to 1.33x, the current-gen L40S 1.25-1.27x, A10G up to 1.08x at scale, and the consumer RTX 5090 1.19-1.23x. The ordering is not a clean function of bandwidth (the 864 GB/s L40S beats the 600 GB/s A10G); the divide is inference-class vs training-class. AMK trails cuBLAS on the high-bandwidth training-class A100/H100, where the harness localizes the cross-SM-sync bottleneck; we report the gap plainly. This is a precision-asymmetric (W8A16 vs bf16) comparison at decode position 0; the largest real checkpoint is TinyLlama-1.1B. Code and the harness: https://github.com/RightNow-AI/AutoMegaKernel

Long-context language model inference is bottlenecked by memory, as the KV cache grows with context length. Recent techniques to compress the KV cache fall short: they either degrade model quality substantially or require considerable time and compute to compress a single long prompt. Furthermore, many methods require the input to fit within the target model's context window, and are generally incompatible with modern production inference engines. Encoder-decoder compressors, which map a long token sequence to a shorter sequence of latent embeddings consumed by a decoder, are an appealing alternative in principle. However, existing approaches are not competitive with KV cache compression on the accuracy-efficiency frontier. In this work, we revisit encoder-decoder compression and close this gap. We first perform an architecture search, pre-training many variants from scratch to determine how best to design and train encoder-decoder compressors. Guided by our findings, we continually pre-train a family of 0.6B-encoder, 4B-decoder models on over 350B tokens each, at compression ratios of 1:4, 1:8, and 1:16. We introduce Latent Context Language Models (LCLMs), a family of compressors that improve the Pareto frontier across general-task performance, compression speed, and peak memory usage. We demonstrate that LCLMs serve as efficient backbones for long-horizon agents, letting the agent skim through a compressed long context and adaptively expand relevant segments on demand.

Ensuring the reliability of Large Language Models (LLMs) under distribution drift requires inference-time adaptation. While inference-time alignment methods such as Best-of-$N$ and rejection sampling are widely used, they frame the task as a sampling-intensive, reward-guided search, leading to two key limitations: their performance is bounded by the base model's generation quality, and their reliance on imperfect reward models makes them vulnerable to reward hacking. To address these challenges, we introduce Gradient-Guided Reward Optimization (GGRO), a lightweight inference-time method that performs targeted, minimal intervention during decoding via gradient guidance. Specifically, GGRO monitors token-level entropy to identify high-uncertainty regions indicative of drift or misalignment. Upon detection, it responds by injecting nudging tokens, generated using gradient signals from an off-the-shelf reward model, to steer the generation trajectory rather than merely re-ranking samples. Experiments show that GGRO consistently improves inference-time alignment across safety, helpfulness, and reasoning benchmarks. It also increases coverage of high-quality responses and robustness to reward hacking, with minimal computational overhead. Code is available at https://github.com/lhk2004/GGRO.

Two-server secure inference allows a client to query a hosted large language model (LLM) without revealing prompts or embeddings. Recent GPU systems based on function secret sharing (FSS) make linear layers efficient, but fixed-point nonlinearities and helper operations remain a bottleneck because each operator is typically implemented as a bespoke protocol with its own comparisons, wrap-around corrections, and preprocessing material. We present FuseFSS, a compiler that replaces per-operator protocol design with a single compilation pipeline. For each scalar fixed-point operator, a compact specification lists its interval partition, low-degree arithmetic pieces, and required predicate bits. The compiler emits two batched FSS evaluations on the public masked value: one packed comparison that returns all predicate bits, and one vector interval lookup that returns the active coefficients and constants. Compared to the current state-of-the-art FSS-based GPU secure inference, FuseFSS preserves accuracy while achieving a $1.24\times$--$1.50\times$ end-to-end speedup and reducing online communication by $9\%$--$16\%$ on BERT and GPT-style models; preprocessing is also lighter, with $14\%$--$23\%$ lower key-generation time and $20\%$--$24\%$ smaller keys.

Large language models (LLMs) incur high inference cost due to their depth and parameter scale. Depth pruning can reduce latency by skipping redundant Transformer blocks, but existing methods (i) provide limited control under user-specific compute budgets and (ii) typically fix the routing path, failing to adapt as the context grows during decoding. We propose Buddy, a budget-driven dynamic depth routing framework. Buddy uses a lightweight Decision Module to score intermediate layers conditioned on the input and deterministically executes the top-k layers to satisfy a given budget. To support decode-time adaptation, Buddy reuses the first-layer KV cache as a low-overhead global context source and pools it together with the newest token representation before each routing decision. When no explicit budget is provided, an optional Budget Predictor estimates an input-dependent compute level to balance quality and efficiency. Experiments on Llama-family and Qwen models show that Buddy is competitive with strong static pruning baselines and often improves the accuracy-compute trade-off, while uniquely supporting strict budget control, decode-time rerouting, and multiple budgets within a single trained model.

Existing sparse attention and KV cache compression methods for long-context LLM inference typically apply fixed sparsity patterns or uniform budgets across all attention heads, overlooking the substantial variation in attention behavior among heads and contexts. We observe two distinct entropy patterns among attention heads: Rigid Heads, whose entropy stays near zero across input segments, and Dynamic Heads, whose entropy fluctuates significantly. Crucially, the distribution of these types is context-dependent and cannot be predetermined offline. We therefore propose EntropyInfer, a training-free framework that uses attention entropy to adaptively allocate compute at the granularity of individual heads and segments during prefilling. For decoding, we introduce a latent KV cache compression scheme that leverages generated output tokens, rather than prefill tokens alone, to identify and retain the most critical cache entries. Extensive experiments on Llama, Qwen and openPangu model series show that EntropyInfer consistently outperforms baselines including SnapKV, AdaKV, and CritiPrefill, achieving up to 2.39$\times$ end-to-end speedup beyond 100k tokens with minimal quality degradation compared to full attention. The code is released in https://github.com/SHA-4096/EntropyInfer.

Retrieval-Augmented Generation (RAG) injects LLM queries with relevant documents to improve response quality. This injection increases prompt length and slows time to first token (TTFT). Unlike standard queries, RAG queries have a unique property of context reuse where the same documents recur across user queries. Thus, fully recomputing documents for every RAG query does redundant compute and increases TTFT. Prior works precompute KV tensors of RAG documents offline and coarsely recompute some tokens during online prefill. However, such KV reuse is often slower than full recomputation on modern GPUs due to high-latency disk transfers. Further, such a coarse-grained recomputation degrades accuracy. To address these limitations, this paper proposes SIFT: Selective-Index For Fast Compute of RAG Prefill by Exploiting Attention Invariance. SIFT processes documents offline and extracts fine-grained locations of high attention scores for each document. Next, we identify the following attention invariance insights that enable us to exploit the extracted locations during runtime: (1) Local-Attention Invariance: The location of high attention scores within a document remain invariant to surrounding documents. This helps us predict the location of high scores where the document attends to itself. (2) Cross-Attention Consistency: Keys with high intra-document attention also attract cross-attention from subsequent documents. This helps us predict the location of high scores where the document attends to future documents. Critically, SIFT stores no KV data and only stores locations of high scores in the form of two compact bit vectors. SIFT's storage is up to 24,000x smaller than KV tensors, obviating costly disk transfers. During prefill, SIFT computes the attention only for the marked locations and improves TTFT by 1.71x while holding accuracy within 1% of full recompute.

Pruning has emerged as a dominant paradigm for accelerating large language model (LLM) inference, spanning a broad spectrum of methods that remove computation across tokens, layers, heads, dimensions, and attention patterns. Despite sharing the same objective, these pruning approaches induce fundamentally different execution behaviors, causing realized speedups to depend heavily on hardware and kernel implementations. Consequently, the practical acceleration benefits of different pruning families remain poorly understood. In this work, we introduce a GEMM-centric taxonomy that reorganizes existing pruning methods according to the logical \textbf{M}, \textbf{N}, and \textbf{K} dimensions of general matrix multiplication (GEMM). Leveraging this abstraction, we build a unified benchmarking framework that enables implementation-consistent comparison across the pruning design space and systematically characterizes the acceleration--quality Pareto frontier. Our results show that static depth pruning remains the strongest Pareto-optimal baseline and stays closest to its theoretical acceleration upper bound in memory-bounded scenarios. During prefill, the frontier transitions from static depth at low quality loss (0%--4%), to dynamic depth at moderate loss (5%--16%), and finally to static width pruning at higher loss levels (17%--26%). These findings establish the first unified view of the practical limits of pruning-based LLM acceleration and provide guidance for future pruning research.\footnote{Code is available at https://github.com/EIT-NLP/LLM-Pruning/tree/main/PruningInferSim}

Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose Lookahead Sparse Attention (LSA), a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a backbone-free decoupled training strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this "less is more" paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), FM-DS-V4 compresses the average physical KV cache footprint down to merely 13.5% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6% absolute margin on average). Crucially, at extreme 500K scales, FlashMemory suppresses the physical KV cache overhead by over 90% without destabilizing the backbone's core reasoning capacities.

We introduce RKSC (Reasoning-Aware KV Cache Sharing), a training-free inference framework that eliminates two structural redundancies in multi-branch LLM reasoning pipelines. ASKS (Attention-Similarity KV Sharing) computes the prefix KV cache once and broadcasts it to all semantically similar branches via hidden-state cosine similarity, strictly generalising the token-exact prefix caching used by vLLM and SGLang. CGEE (Confidence-Gated Early Exit) applies two complementary exit mechanisms: (1) it skips the verification forward pass entirely when generation confidence is decisive across branches, and (2) it terminates the verification pass at an intermediate layer when per-layer entropy stabilises, using lightweight hooks on the transformer backbone. RSBCM (Reasoning-Selective Block Cache Manager) prevents unbounded cache growth via attention-weighted depth-priority eviction. Across five model families (7B-10B), four benchmarks, and 1,000 evaluated problems, RKSC achieves a mean speedup of 3.008x over the No-KV baseline (peak 3.990x), a 1.66x mean improvement over vLLM-equivalent prefix caching, with a CGEE-induced error rate of only 0.37% (6 errors out of 1,616 verify calls). No fine-tuning or architecture changes are required. Code is available at https://github.com/AnirudhSekar/RKSC.

W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this bottleneck. Through controlled benchmarks across four GPUs from Ampere and Ada architectures, we identify the Tensor Cores to CUDA Cores throughput ratio ($ρ$) as the primary hardware indicator: the W4A4-g128 kernel yields $2.0$--$2.5\times$ speedup on RTX3090 ($ρ=16$) yet degrades to $0.43$--$0.47\times$ on A100 ($ρ=64$) in compute-bond scenarios, establishing W4A4 viability as platform-dependent rather than universally infeasible. Guided by this finding, we build \textbf{APEX4}, which co-designs pure INT4 GEMM kernels with $ρ$-aware granularity adaptation to mitigate the CUDA Cores dequantization bottleneck. APEX4 achieves perplexity within 0.63 of FP16 on LLaMA-2-70B and outperforms W4Ax Atom-g128 by 4.0%--4.4% in zero-shot accuracy. Deployed as a drop-in replacement in unmodified vLLM, it delivers up to $1.66\times$ end-to-end speedup on L40S ($ρ=8$), and $1.78\times$ on RTX3090 ($ρ=16$), $2.09\times$ on A40 ($ρ=16$), while recovering A100 ($ρ=64$) to $1.20$--$1.40\times$ via the mixed-granularity mode. Our code is available at https://github.com/APEX4-W4A4/APEX4-W4A4.

Block-wise semi-autoregressive decoding is the standard inference paradigm for diffusion large language models (DLMs), but it imposes a strict dependency between blocks: the next block cannot begin until the current block is fully decoded or its denoising budget is exhausted. We observe that once a block exposes a reliable delimiter boundary or stable semantic prefix, continuation generation need not wait for every residual token to be resolved. We propose AsyncLane, a training-free decoding scheduler that decouples refinement from advancement. AsyncLane forks a generate lane at observed delimiter boundaries into a refine lane and a continuation generate lane: the prefix remains editable, while the continuation advances before prefix refinement finishes. The resulting lane tree records decoding dependencies and output order, while execution proceeds over the active lane set. To make this asynchronous schedule efficient under bidirectional attention, AsyncLane combines shared-prefix lane batching, lookahead draft reuse, cascading termination, and compact cache refresh with refresh-logit reuse, preventing model-call cost from scaling directly with the number of lanes. AsyncLane is a drop-in replacement for block-wise DLM samplers and requires no retraining. Experiments on mathematical reasoning and code generation show that AsyncLane consistently improves throughput while maintaining competitive quality. Across LLaDA and Dream backbones, AsyncLane achieves the highest TPS in all evaluated benchmark-length settings; relative to the fastest competing baseline, it reaches peak speedups of 2.95x on LLaDA and 3.04x on Dream, with especially large gains under longer generation budgets.

Multi-turn LLM agents fan short queries into long trajectories of tool calls, search results, and intermediate reasoning. Both KV memory and KV read bandwidth grow by orders of magnitude across a single trajectory, making the key-value (KV) cache, not parameter compute, the dominant serving bottleneck for long-horizon agents. We introduce IntentKV, learned KV pruning that keeps the base LLM frozen. IntentKV maintains a session-level QueryMemory of cross-turn intent, scores live history tokens with a memory-attention rule, and adds a zero-initialized residual head with cross-attention over current-query K-vectors. To stay composable with prefix caches, eviction is a slot-map redirection: dropped positions route to a sentinel dead slot while surviving K/V rows, RoPE phases, and slot identities stay in place. IntentKV matches the no-pruning full-cache baseline with almost no accuracy drop under tight KV budgets: at an 8k KV budget, mean peak request tokens drop 23.9% on Qwen3-8B and 30.7% on Qwen2.5-14B. On the 100 longest BCP queries that all methods complete on Qwen2.5-14B, IntentKV-8k further cuts worst-case peak request tokens from 92.3k to 20.5k, a 77.8% reduction, and worst-case raw KV reads from 411M to 31M, a 92.6% reduction.

Majority voting is the default unsupervised aggregator for multi-sample LLM inference, but it discards two signals: within-group answer entropy and between-group reasoning geometry. We aggregate by delegation instead (Propagational Proxy Voting, PPV): each group of samples keeps weight on its own answer in proportion to its entropy-based confidence (When) and routes the rest to peers by reasoning-embedding similarity (Whom); the stationary distribution of the resulting delegation matrix picks the consensus answer. This requires neither gold labels nor training. On MMLU-Pro with 128 samples per question, delegation beats majority by +1.5 pp overall and +2.24 pp on non-trivial questions (McNemar p ~ 1.0e-14, n = 8,099), overturning wrong majorities whose answer cluster is geometrically incoherent while the correct minority is tight. We then characterize exactly when delegation overturns majority: a two-option model gives a closed-form flip condition on each option's confidence and the weight it routes to the other, with a do-no-harm corollary for near-unanimous questions. The condition calls the realized winner on 96.5% of non-trivial questions, and its predicted mass gap tracks the realized gap at r = 0.97. We did not find any other unsupervised ensemble methods that close the oracle gap.

The autoregressive nature of large language models (LLMs) remains a significant bottleneck for inference, particularly in complex agentic workloads. While speculative decoding (SD) accelerates inference, current approaches rely on static drafting paradigms, utilising either autoregressive drafting models for reasoning or diffusion-based parallel drafting models for structured outputs. We empirically find that drafting accuracy fluctuates dramatically within a single sequence, leaving significant performance unrealised by static paradigms and coarse-grained routing. To address this volatility, we introduce WhiFlash, the first cross-paradigm SD method that unifies autoregressive and diffusion-based parallel drafting under a single token-level controller. WhiFlash adopts a fine-grained routing mechanism that employs either a lightweight entropy-based or a learned neural policy, both parametrised to provide a tunable balance between expected token gain and latency. To make high-frequency switching computationally viable, we introduce novel cache-management optimisations, Lazy Catch-up and KV-only Prefill, reducing switching overhead to below 7% of per-round latency. By capitalising on the complementary strengths of fundamentally distinct drafting architectures, WhiFlash achieves significantly higher acceptance lengths, yielding category-specific throughput gains of up to 69.6% over the state-of-the-art autoregressive EAGLE-3 and 37.3% over the diffusion-based DFlash.

Low-bit quantization has been widely adopted to accelerate the inference of large language models (LLMs) by significantly reducing computational cost and memory usage. However, activation outliers pose a major challenge to effective quantization, often leading to notable performance degradation. In this paper, we introduce OffQ, a method designed to mitigate activation outliers in low-bit quantization through a novel offsetting mechanism. Specifically, OffQ first identifies a low-dimensional outlier subspace in the activations using a proposed top-1 PCA, and then concentrates high-magnitude activations into 1 channel via rotation. OffQ then absorbs this concentrated outlier channel by converting its magnitude into a shared offset, thereby reducing the standard deviation of the activations. This offsetting strategy enables effective W4A4KV4 quantization of LLMs using deployment-friendly uniform-grid and uniform-precision quantization. Extensive experiments across diverse LLM architectures and benchmarks demonstrate that OffQ outperforms state-of-the-art baselines, consistently improving model accuracy while preserving low-bit efficiency.

Large language models (LLMs) perform inference by following a fixed depth and order, non-recurrent execution of all layers. We reveal the wide existence of training-free, flexible, dynamic program-of-layers (PoLar), where pretrained layers can be packed as modules and then skipped or looped to form a customized program for each input. For most inputs, substantially shorter program executions can achieve the same or better accuracy, while incorrect predictions of the original LLM can be corrected by alternative programs with fewer layers. These observations indicate that inference admits multiple valid latent computations beyond the standard forward pass. To efficiently achieve PoLar in practice, we propose a lightweight PoLar prediction network, which learns to generate execution programs that dynamically skip or repeat pretrained layers for each input. Experiments on mathematical reasoning benchmarks demonstrate that PoLar consistently improves accuracy over standard inference and prior dynamic-depth methods, often while executing fewer layers, and that these gains persist under out-of-distribution evaluation. Our results suggest that fixed-depth execution captures only a narrow subset of an LLM's latent reasoning capacity.

Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought. Existing sparse attention methods often face a practical efficiency-quality trade-off. Structured block sparse methods typically provide stronger acceleration but incur noticeable quality loss, while token sparse methods are usually more accurate yet deliver limited end-to-end speedup because top-k routing over the full cache remains expensive. In this work, we propose cross-layer sparse attention (CLSA), which is built on top of KV-sharing architectures such as YOCO. The core idea is to share not only the KV cache across cross-decoder layers, but also the routing index. A single indexer computes token-level top-k selection once and reuses the resulting index across layers, thereby preserving the fine-grained selectivity of token sparse attention while amortizing the routing overhead. The resulting architecture improves all major inference bottlenecks jointly, including pre-filling, KV-cache storage, and long-context decoding. Experiments across short-context and long-context benchmarks show that CLSA is both accurate and efficient, achieving up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context. These results suggest a more complete architectural solution for long-context LLMs that jointly advances model quality and inference efficiency.

Causal graphs provide a high-level language for making mechanisms transparent. Recent work uses Large Language Models (LLMs) to recover causal graphs of external-world processes. Instead, in this paper, we use causal graphs to model LLM inference itself, providing stakeholders with a transparent view of how the model perceives and organizes high-level concepts to produce a prediction. We propose a four-phase method for constructing such graphs. Given a target LLM and a set of textual examples, our method discovers class-discriminative, human-interpretable concepts and maps each input to LLM-perceived concept states. We then introduce an MCMC-inspired counterfactual augmentation procedure that expands the sparse observational data through chains of counterfactuals. This enables stable causal discovery with $σ$-CG, yielding informative, interpretable graphs. We apply our method to three LLMs across disease diagnosis, sentiment analysis, and LLM-as-a-judge classification tasks. We evaluate the learned graphs for predictive fidelity and structural stability, and the MCMC-inspired augmentation for convergence and downstream utility. Our results show that the discovered causal graphs capture meaningful dependencies consistent with LLMs' reasoning. Together, this paper provides a foundation for concept-level explainability of LLMs.

With the widespread deployment of public large language models (LLMs) such as ChatGPT, protecting user prompt privacy has become an increasingly critical issue. Existing privacy-preserving inference methods sacrifice either utility or efficiency, and often require model-specific modifications that limit their compatibility. In this paper, we propose SharedRequest, a model-agnostic framework for privacy-preserving LLM inference that reformulates privacy protection at the batch level rather than the individual-prompt level. The key idea is to obscure sensitive information by mixing original prompts with noisy variants, while grouping semantically equivalent instructions to amortize the inference cost over a large batch of queries with minimal impact on LLM response quality. This design is independent of the LLM architecture, requiring no access to model parameters or architectural modification. Empirical results demonstrate that SharedRequest achieves over $20\%$ higher utility compared to prior differential privacy baselines, and its shared-prompt mechanism reduces query cost by up to $5\times$ compared to non-batched inference.

Sparse attention reduces compute and memory bandwidth for long-context LLM inference. However, two key challenges remain: (1) KV cache capacity still grows with sequence length, and offloading to CPU memory introduces a PCIe transfer bottleneck; (2) the sparse selection step itself retains $O(T^2)$ complexity and can dominate attention cost at long contexts. We propose SparDA, a decoupled sparse attention architecture that introduces a fourth per-layer projection, the Forecast, alongside Query, Key, and Value. The Forecast predicts the KV blocks needed by the next layer, enabling lookahead selection that overlaps CPU-to-GPU prefetch with current-layer execution. Because Forecast is decoupled from the attention query, our GQA implementation uses one Forecast head per GQA group, reducing selection overhead versus the original multi-head selector. SparDA adds $<$0.5% parameters and trains only the Forecast projections by matching the original selector's attention distribution. On two sparse-pretrained 8B models, SparDA matches or slightly improves accuracy and delivers up to 1.25$\times$ prefill speedup and 1.7$\times$ decode speedup over the sparse-attention offload baseline. By enabling larger feasible batch sizes on a single GPU, SparDA further reaches up to 5.3$\times$ higher decode throughput than the non-offload sparse baseline. Our source code is available at https://github.com/NVlabs/SparDA.

Speculative decoding accelerates autoregressive large language model inference by drafting multiple tokens and verifying them in a single target-model forward pass. Recent diffusion-based drafters generate an entire block of tokens in parallel but usually commit to a single draft sequence per verification: once the first mismatch occurs, all subsequent draft tokens are discarded, resulting in a limited acceptance rate. Naively batching more draft candidate sequences only introduces a marginal improvement, as redundant or poorly placed branches increase the cost of drafting and verification without proportionally increasing the number of accepted tokens. We propose D^2SD, a dual diffusion draft speculative decoding framework that organizes candidates into a confidence-guided prefix tree, where the first diffusion drafter generates a block along with per-position confidence scores that are used to identify the most likely rejection boundary and select the top-K prefix ranges for recovery; the second variable-prefix diffusion drafter re-anchors at each selected prefix and proposes alternative continuations in one batched pass; the resulting shared-prefix candidates are jointly verified via cascade attention. Empirically, D^2SD shows clear improvements over both the underlying diffusion approach and strong autoregressive speculative decoding baselines.

This study addresses on-device inference bottlenecks of Transformer models on Tenstorrent's Tensix architecture and proposes an operator fusion strategy that enhances data locality. RMSNorm is fused with matrix multiplication in self-attention and in the FFN, enabling back-to-back execution of memory-bound and compute-bound operators in on-chip SRAM to significantly reduce DRAM reads/writes of intermediate results and scheduling overhead. To support multi-core parallelism, a NoC-based multicast mechanism is leveraged in which row/column master nodes efficiently distribute inputs and weights across the core mesh, alleviating DRAM bandwidth contention. Experiments on the Wormhole platform with Qwen2.5-0.5B, Qwen3-0.6B, and Qwen3-4B show up to 37.44% latency reduction for attention and 15.89% for MLP, with up to 7.91% reduction per decoder layer, while Pearson Correlation Coefficient (PCC) remains above 98.75%, confirming significant end-to-end efficiency gains under numerical consistency.

Key-value (KV) caching accelerates inference of large language models (LLMs) by reusing past computations for generated tokens. Its importance becomes even greater in long-context applications such as retrieval-augmented generation (RAG) and in-context learning (ICL). However, conventional KV caching embeds positional information directly into the cache, limiting its reusability. Existing solutions either restrict reuse to prefixes or require expensive memory materialization for positional re-encoding. We introduce LazyAttention, a novel attention mechanism that kernelizes deferred positional encoding to enable zero-copy, position-agnostic KV reuse. By adjusting positional encoding within attention kernels on-the-fly, LazyAttention resolves the materialization bottleneck, allowing a single physical KV copy to serve multiple logical requests at arbitrary positions. Leveraging attention kernels tailored for prefilling and decoding, our system achieves significant efficiency improvements: under skewed document distributions, it reduces time-to-first-token (TTFT) by 1.37$\times$ and increases inference throughput by 1.40$\times$ compared to the state-of-the-art Block-Attention, while maintaining comparable output quality.

Aggressive weight quantization to 2-bit precision offers substantial throughput and memory gains for large language model (LLM) inference, but typically incurs severe accuracy degradation. These gains are particularly relevant for edge and on-device deployment, where memory capacity and bandwidth are primary constraints. In this work, we extend Recover-LoRA -- a lightweight, data-free accuracy recovery method originally developed for general model weight corruption -- to the setting of ultra-low-bit quantization. We propose a selective mixed-precision strategy in which only gate and up projection layers of the MLP are quantized to 2-bit (W2), while all other linear layers remain at higher precision, yielding a mixed-precision GateUp configuration. We demonstrate via roofline analysis across three model families (4B--20B) and two hardware platforms that a W4/W2-GateUp deployment (4-bit base with 2-bit gate/up) delivers 7.5--23.3% TPS improvement over uniform W4 depending on model and context length, while confining quantization error to a predictable subset of layers. We then apply Recover-LoRA -- training low-rank adapters on the quantized layers via logit distillation with synthetic data -- to recover accuracy lost from 2-bit quantization of the gate and up layers. In a case study on Qwen3-4B, Recover-LoRA achieves 80--95% accuracy recovery on 9 of 12 benchmarks, using only 10k synthetic training samples and no labeled data. We further demonstrate that synthetic data performs comparably to curated labeled data for distillation-based recovery, and that recovery generalizes to out-of-distribution evaluation tasks. Our results present Recover-LoRA as a practical post-quantization accuracy recovery tool for aggressive weight compression in deployment settings.

Disaggregated LLM inference forces the KV cache to traverse the datacenter network before decoding begins, so transfer time enters directly into the Time to First Token (TTFT) budget. Current schedulers route on compute load and prefix-cache locality alone, ignoring the topological distance and dynamic congestion between prefill and decode instances. We close this gap with a thin operator-to-scheduler interface, the network cost oracle, and we prove that ignoring the network term renders cache-aware-only scheduling arbitrarily suboptimal as context length grows. NetKV, the O(|D|) per-request greedy that consumes this oracle, has tier rankings that are provably robust to stale telemetry. On a 64-GPU four-tier fat-tree simulator driven by Mooncake traces, NetKV reduces mean TTFT by up to 21.2% over round-robin and 17.6% over a tuned cache+load-aware scheduler, lifts SLO attainment by up to 20.1 percentage points, and keeps the Time Between Tokens overhead below 0.5 ms in every condition tested, with no changes to the transport, inference engine, or hardware.

Large Language Models (LLMs) have become integral to modern applications, yet their deployment remains challenging. Beyond executing the models themselves, practical deployment must address cost efficiency, low latency, and optimal resource utilization. Conventional approaches typically assume that an entire model can be hosted on a single device, which does not hold in many real-world scenarios, particularly in Edge and Fog environments where device resources are constrained. In this paper, we introduce E2LLM, a framework designed to enable efficient LLM deployment in such resource limited settings. Rather than simply partitioning a single model across all available devices, E2LLM replicates the full model across multiple groups of devices (replicas) and applies model parallelism within each replica. Each replica is assigned a specialized role PREFILL or DECODER based on its efficiency in handling input and output tokens. This separation leverages the inherent differences between these two phases of LLM inference. To effectively organize devices, we utilize a Genetic Algorithm to form clusters that maximize system performance. Within each cluster, we apply Dynamic Programming to determine an optimal partitioning strategy that minimizes bottlenecks in model-parallel execution. Experimental results demonstrate that our approach adapts robustly to varying workloads, including scenarios with significant variation in input and output token lengths. Compared to the Splitwise baseline, E2LLM reduces average waiting time by over 50% under high-demand conditions

Deploying large language models (LLMs) is challenging due to their significant memory and computational requirements. While some methods address this by developing small or tiny language models from scratch, these approaches demand extensive GPU training. Compressing pre-trained LLMs for edge devices offers a compelling alternative. Beyond pruning and quantization, Neural Architecture Search (NAS) enables effective compression, yet prior NAS approaches often limit the search space and decouple architecture from quantization. We introduce a differentiable NAS framework that explores the entire space and jointly optimizes architectural configurations alongside mixed-precision quantization for linear layers of LLMs. Experiments demonstrate superior accuracy-latency trade-offs: our models achieve up to 1.4x faster inference than sequential NAS-then-quantization baselines at comparable accuracy, or up to 6% higher average accuracy across seven reasoning tasks at equivalent latency.

Benchmark contamination, where evaluation examples appear in a model's training data, threatens the validity of LLM assessment. Statistical tools for detecting training-data membership exist, but have been validated almost exclusively in controlled academic regimes: large, homogeneous pre-training corpora and transparent, single-stage training pipelines. Whether these methods remain reliable in realistic auditing scenarios remains unclear. We identify two under-studied failure modes: distribution shift, which arises when suspect and validation sets violate the IID assumption, and scale constraints, which arise because benchmarks are orders of magnitude smaller than pre-training corpora. We systematically evaluate three leading paradigms, LLM Dataset Inference, Post-Hoc Dataset Inference, and CoDeC, across 25 models from multiple families (including Pythia, OLMo 2, and specialised cultural and medical LLMs) and scales (up to 27B). We then further extend our analysis to frontier industry models. Across 335 evaluations, only 201 yield correct outcomes. LLM Dataset Inference results in false positives under distribution shift, Post-Hoc Dataset Inference is underpowered at benchmark scale, and CoDeC provides only coarse provenance signals that are insufficient to verify individual benchmark splits. Our results reveal a systematic reliability gap between controlled validation and practical benchmark auditing, and show that statistical detection cannot yet replace transparent data provenance. We open-source our benchmark for further research.

Large Language Models suffer from slow autoregressive inference. While self-speculative decoding accelerates this process, its efficiency is hampered by static configurations like fixed exit layers and speculation lengths. We reframe this optimization as a \textbf{Markov Decision Process} and propose \textbf{LEDE}, a framework that uses offline reinforcement learning. LEDE learns a policy to dynamically select the optimal exit layer and speculation length based on the local context of the generated sequence at each step, balancing computational cost and draft quality. Comprehensive evaluations on Llama-2 and Llama-3 models show LEDE achieves up to a 2.0\times$$\sim$$2.7\times speedup over autoregressive decoding and and provides an additional 17% speedup over the static speculative baselines.

Binary spike activations allow a language-model runtime to read only active weight columns and replace multiplications by weight sums. We implement this execution strategy in C++ for an 874M-parameter spike-gated language model. Sparse projections use column-major INT8 weights, integer accumulation, and one scale application per output channel; dense projections retain row-major access and FP32 activations. In a single-thread comparison using an early checkpoint, INT8 achieves 23.31 tokens/s versus 9.82 for FP32, while reducing weight storage from 3355.2 to 1087.4 MiB. A variant using INT4 on dense projections saves a further 17.4% of storage but reduces decode throughput by 46.6%. On an AMD Ryzen 7 5800X, the final INT8 checkpoint achieves 22.63 tokens/s on one thread and 47.90 on four threads; 512-token prefill reaches 94.68 tokens/s on eight threads. A separate ARM output-head case study records higher trimmed decode-window energy metrics for two candidate-verification configurations. The results characterize how activation-specific layouts and quantized kernels support CPU deployment of a spike-gated language model.

Large Language Model (LLM) inference relies on key-value (KV) caches to avoid redundant attention computation. While approximate KV cache retention techniques reduce memory usage by sacrificing model accuracy, lossless approaches instead evict KV cache blocks from GPU memory and reconstruct them on demand to preserve exact outputs. Existing lossless KV cache management systems primarily base eviction decisions on access frequency or positional heuristics, without considering how different KV cache blocks affect the execution efficiency of GPU attention kernels. In this paper, we propose AsymCache, a computation-latency-aware KV cache management system for LLM inference that explicitly aligns cache residency decisions with GPU attention kernel performance, including three key components: Multi-Segment Attention (MSA) for efficient non-contiguous KV context processing, a cache eviction policy that jointly optimizes hit rate and position-aware recomputation cost, and an adaptive chunking scheduler for high hardware utilization. Experiments show that AsymCache reduces TTFT by up to 1.90-2.03x and time-per-output-token (TPOT) by 1.62-1.71x over latest baselines, confirming the effectiveness of the method in common workloads and validating its design goal of balancing computational efficiency with cache hit rate. Moreover, the low-level design of AsymCache allows seamless integration into agent serving systems such as Continuum, where it further reduces average job latency by up to 18.1%.

Production inference increasingly targets a heterogeneous mix of accelerators. Agentic pipelines interleave reasoning, tool calls, and multi-agent coordination, each with distinct compute and memory profiles. For optimal efficiency, each stage should run on the accelerator best suited to it. This creates a systems challenge: each pipeline now requires high-performance kernels across a growing set of hardware backends and programming models. Writing these kernels by hand is time-consuming, demands deep low-level expertise, and does not scale as kernel complexity grows. Recently, Large Language Models (LLMs) have been leveraged for automatic kernel generation, but challenges in low-level code generation and cross-backend generalization persist. We present KForge, a cross-platform framework built around an iterative refinement loop driven by two collaborating LLM-based agents: a generation agent that produces and progressively refines kernels using compilation and correctness feedback, and a performance-analysis agent that interprets profiling data, from programmatic APIs to GUI-based tools, and emits recommendations that steer the next round of synthesis. The loop alternates between functional passes, which drive a candidate to correctness, and optimization passes, which close the performance gap to hand-tuned baselines. We evaluate KForge on two backends with very different baseline reference availability. On NVIDIA B200, KForge achieves a 2.12$\%$ improvement in end-to-end throughput compared to TensorRT-LLM on the gpt-oss-20b inference speed benchmark. On Intel Arc B580, KForge generates Triton kernels achieving a 5.13$\times$ geometric mean speedup over the faster of PyTorch eager and torch.compile on 37 GEMM + tail-ops workloads from KernelBench Level 2, primarily via operator fusion and mixed-precision execution.

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles. We propose \textbf{Fast-dLLM++}, a training-free extension that introduces \emph{Fréchet profile decoding}: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable \emph{heterogeneity bonus} when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding. Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy--throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37% higher throughput at comparable accuracy. Our code release is at https://github.com/Ringo-Star/FastdLLM_plusplus.

Two-bit weight quantization is attractive for memory-efficient LLM inference, but the standard W2 level set {-2,-1,0,+1} often collapses under aggressive W2A4/KV4 settings. We study the scalar level-set geometry of two-bit weights in a Hadamard-rotated quantization pipeline. Conventional asymmetric W2 substantially improves over the standard level set, indicating that W2A4 failure is not only a bit-width problem but also a reconstruction-level problem. Across all 224 linear modules in each of LLaMA-2-7B and LLaMA-3.1-8B, pretrained weights are already nearly zero-centered, while Hadamard rotation primarily Gaussianizes their standardized shape: excess kurtosis and Q-Q error drop by orders of magnitude. Based on this approximate zero-centered Gaussian-like source model, we propose Qift, a fixed no-zero W2 level set for rotated W2A4/KV4 inference. The main level set is {+/-0.5, +/-1.5}, equivalently {+/-1, +/-3} under a half-scale reparameterization; a power-of-two variant uses {+/-1, +/-4} for sign-and-shift decoded weight application. Qift redesigns the fixed two-bit code-to-level mapping and is training-free, learned-codebook-free, group-grid-free, and zero-point-free, retaining the standard per-channel scale. A scale-invariant ratio analysis identifies an effective inner/outer centroid ratio range of 0.25 to 0.33, explaining why mirror no-zero (MNZ), Lloyd, NF2, and PoT-MNZ perform well while {+/-1, +/-2} does not. On both models, the no-zero level sets consistently improve pure W2A4 perplexity, L-layer mixed W2/W4 perplexity, downstream accuracy, and GPTQ residual behavior over the standard W2 level set. At L=16 mixed precision, they substantially narrow the gap to W3A4 while keeping half of the transformer layers at two-bit precision, giving a simple, source-aware, and deployment-friendly alternative to more complex learned W2 codebooks.

Diffusion large language models (dLLMs) have recently emerged as a promising alternative to autoregressive (AR) LLMs, offering faster inference through parallel or blockwise decoding. However, their masked language modeling formulation remains incompatible with standard token-level speculative decoding, one of the most effective acceleration techniques for AR models. In AR decoding, the causal mask preserves temporally valid token-level contexts, enabling a target model to verify multiple drafted tokens in a single forward pass. In contrast, dLLMs rely on mask tokens and bidirectional attention, causing the effective context to change across denoising steps and preventing direct token-level speculative verification. To bridge this gap, we propose a simple but effective speculative decoding algorithm for diffusion language models, named SimSD, which mainly adopts a plug-and-play masking strategy that equips dLLMs with temporally valid token-level contexts for speculative decoding. Our method explicitly introduces reference tokens from draft-model predictions and designs an attention mask that regulates their interaction with current-step tokens, allowing dLLMs to compute valid logits for drafted tokens in a single forward pass. This restores the key verification ability provided by causal masking in AR models while preserving the parallel decoding advantages of dLLMs. The proposed method is training-free and can be flexibly integrated with other acceleration techniques such as KV cache and blockwise decoding. Experiments on SDAR-family dLLMs across four benchmarks show that our method achieves up to 7.46x higher decoding throughput while maintaining and even improving average generation quality.

Large language models (LLMs) are increasingly integrated into high-performance computing (HPC) workflows, accelerating scientific discovery through diverse perspectives such as code generation and domain-specific decision-making. Yet, how soft errors propagate and affect LLM inference remains largely unexplored. To bridge this gap, we present a comprehensive study on error propagation in LLM inference, enabled by our proposed LLMFI, a configurable and deterministic fault-injection framework. Using LLMFI, we systematically inject faults across three open-weighted LLMs and thirteen representative tasks, covering reasoning, multilingual, mathematical, and coding domains. In addition, we conduct fine-grained case studies that reveal critical vulnerability patterns. Overall, our study yields 17 takeaways that advance the understanding of error propagation in LLM inference and introduces four low-overhead directions to improve reliability through software-only modification, offering practical guidance for future error detection and mitigation.

Quantum LDPC code discovery requires searching large algebraic design spaces while reliably certifying the parameters and equivalence classes of any candidates found. We introduce an LLM-guided evolutionary workflow in which language models mutate Python programs that generate bivariate-bicycle and perturbed bivariate-bicycle code ansätze. Across five campaigns, the system performed approximately 1{,}650 evolutionary iterations, screened about $2 \times 10^5$ candidate codes, and required ${\sim}140$ hours of computation and ${\sim}$US$400 in LLM inference cost. Candidate codes are evaluated through a staged validation pipeline combining $\mathrm{GF}(2)$ rank computation, distance estimation and certification, mixed-integer linear programming, BLISS Tanner-graph deduplication, decomposability analysis, and local-Clifford equivalence checks. At block length $n \leq 360$, the workflow identifies 465 distinct candidate codes: 97 CSS bivariate-bicycle codes and 368 non-CSS perturbed variants. The CSS search recovers known high-performing codes and finds new finite-length representatives, including an indecomposable [[288,16,12]] code and higher-weight codes with up to $k = 50$ at distance $d = 8$. The non-CSS search produces perturbed codes matching the gross-code figure of merit at [[144,12,12]], along with additional high-distance candidates reported as certified values or upper bounds according to MILP status. Overall, these results show that LLM-guided program evolution can serve as a practical tool for structured quantum-code discovery when paired with independent evaluation.

Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present \modelname, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, \modelname attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11%, 8%, and 5% respectively. Our code is available at https://github.com/Tencent/AngelSlim.

Large Language Models (LLMs) frequently exhibit "contextual disregard" when faced with input evidence that conflicts with their internal parametric memory, leading to persistent factual hallucinations. Existing mitigation strategies primarily rely on suppressing specific neuron activations or employing computationally expensive contrastive decoding mechanisms, which often result in increased perplexity or significantly elevated inference latency. To address these limitations, we propose Resonant Context Anchoring (RCA), a lightweight inference-time intervention method grounded in the perspective of residual stream signal dynamics. RCA aims to resolve the signal attenuation of external evidence during its propagation through deep networks. The core mechanism involves the orthogonal decoupling of routing logic and information magnitude within the self-attention module. By utilizing raw pre-softmax attention scores as an instantaneous metric of semantic alignment, we construct a dynamic gain field via non-linear rectification to selectively amplify the norms of value vectors corresponding to context tokens, without altering the attention probability distribution. This mechanism effectively elevates the signal-to-noise ratio (SNR) of input evidence within the residual stream mixture, thereby robustly anchoring the generation trajectory to the truthful context during inference. Extensive experiments on the Llama-3 model series demonstrate that RCA significantly improves contextual faithfulness across multiple factual consistency and strong knowledge-conflict tasks, effectively suppressing parametric hallucinations. Furthermore, results confirm that as a training-free and computationally negligible plug-and-play module, RCA achieves a Pareto improvement in faithfulness and fluency while maintaining the model's general language understanding capabilities.