Daily Papers

As large language models serve ever more requests, cumulative inference cost is growing relative to the one-time cost of training. In typical serving, prompt prefill runs in parallel and is compute-bound, whereas autoregressive decode is sequential and memory-traffic-bound. Conventional width or depth scaling raises both costs together, since every added layer is evaluated in both phases and enlarges the weights read at each decode step. We instead ask whether additional learned computation can be allocated to continuation prediction while preserving prompt-wide primary computation and a single KV cache. We realize this with the Decode-Branch Transformer. Its primary path alone processes the prompt and writes the KV cache; the decode branch is omitted during prefill and activated only from the final prompt position onward, adding continuation computation without writing state or affecting the primary path. The paths share attention, MLP, and output matrices, using separate token embeddings with lightweight coupling. Grouped decode reuses loaded weight tiles and the primary KV cache across both paths, so the added arithmetic does not proportionally increase dominant memory traffic or decode latency. Across matched-token comparisons, Decode-Branch achieves lower validation loss across architectures and data settings. In MoE models, the primary and branch expert fan-outs become independent knobs for trading prompt cost, decode cost, and predictive quality. We study two expert-allocation regimes, holding prefill or decode computation fixed, and expose a prefill-decode-quality trade-off enabled by phase-specific expert allocation.

Physical AI systems, including robots, autonomous vehicles, embodied agents and edge copilots, often run a different inference workload from cloud LLM serving: single-stream, batch-1 autoregressive decode, where one robot, camera feed or user session waits on the next token. This workload is usually described as memory-bandwidth-bound. Each decode step streams model weights and the active KV cache, so latency should scale with peak HBM bandwidth. We show that this account is true but incomplete. We measure batch-1 decode for three 7 to 8B-class GQA transformers across four NVIDIA GPUs: H100 SXM5, A100-80GB SXM4, L40S and L4. We evaluate context lengths from 2048 to 16384, producing 44 valid cells under a controlled bf16 SDPA setup. The achieved fraction of peak HBM bandwidth falls as peak bandwidth rises. On the headline Qwen-2.5-7B ctx=2048 cell, an L4 reaches roughly 81 percent of its analytic memory floor, while an H100 reaches only 27 percent. Physical-AI decode is memory-dominated, but faster memory does not translate into proportional latency gains. We test the missing term with a CUDA Graphs A/B experiment. On H100 at ctx=2048, CUDA Graphs improves decode latency by 1.259x across N=10 fresh sessions, with a 95 percent bootstrap confidence interval of 1.253 to 1.267. On L4, the same intervention gives only 1.028x. This isolates a launch-side overhead that becomes visible on fast GPUs but remains mostly hidden on slower, bandwidth-bound GPUs. The deployment implication is that memory savings matter only when the runtime realises them. On L4, bf16 decode sits close to the memory floor, but common quantised paths do not recover the expected 4x weight-traffic reduction: bnb-nf4 reaches 59.36 ms/step and AutoAWQ+Marlin reaches 45.24 ms/step from a 62.32 ms bf16 baseline. GPTQ+ExLlamaV2, with Ada-tuned int4 kernels, reaches 17.36 ms/step.

Speculative decoding accelerates autoregressive generation by having a cheap draft propose tokens that a target verifies in parallel. Frontier models increasingly ship a built-in Multi-Token-Prediction (MTP/NEXTN) draft head under the assumption that the draft is negligibly cheap. At million-token context this breaks: an MTP draft head typically runs full attention over the entire KV cache at every draft step, so its read grows linearly with context and comes to dominate the draft cost -- precisely where speculation is most valuable. The effect compounds with draft length (a deep native draft can turn net-negative, slower than no speculation) and sharpens under hybrid/linear-attention targets, where cheaper verification leaves the draft's full-attention read exposed. We apply a StreamingLLM-style sliding window plus attention sink to the draft's attention only (Windowed-MTP), leaving full-attention verification intact. It is training-free, drop-in, and lossless by construction: the full-attention target still decides every accepted token, so windowing changes only which tokens are proposed, never which are accepted. It bounds the draft's KV working set to a constant, dropping ~99% of KV entries at 1M. Across three architecture families (Qwen GDN-MoE 35B/122B and a Mamba2-hybrid NoPE 120B) at 1M context on a single GPU in SGLang, windowing cuts the per-decode-step cost over the shipping native MTP draft by +28% to +44%, an input-invariant margin that widens with context. Since per-token latency is this cost divided by acceptance length, at matched acceptance end-to-end decode latency improves by the same amount, and more where windowing also lifts acceptance, while preserving the target's verified output distribution. Finally, the unread draft KV -- 7.7-11% of total KV at 1M -- is reclaimed via a compact ring buffer at no acceptance or quality cost.

AI agents are increasingly used to automate research and development tasks, yet existing benchmarks typically evaluate them on prescribed workflows or narrow action spaces. Even nominally open-ended tasks can often be solved by retrieving a well-known recipe and tuning a few hyperparameters, making it unclear whether strong results reflect genuine optimization or memorized solutions. We introduce InferenceBench, where an agent must deploy an OpenAI-compatible inference server and optimize the speed of LLM inference. Each agent receives a target LLM, one H100 GPU, an optimization scenario, and a wall-clock time budget of two hours. Three optimization scenarios isolate distinct bottlenecks of inference (prefill latency, decode latency, and concurrent request throughput) and a fourth balances all three at the same time. Across 15 frontier agent configurations, agents reliably improve over a naive PyTorch baseline (up to 8.08times) and often match or exceed serving engines with default settings (4.05times for vLLM), but still fall below a simple hyperparameter search under the same time budget (up to 11.53times). Qualitative analysis of agent trajectories shows that although agents enumerate many relevant optimization techniques, they overwhelmingly converge on a single inference framework. They test only a few distinct configurations and spend the remaining budget re-measuring, repairing, or optimizing hyperparameters rather than exploring substantially different strategies. This suggests the bottleneck is not domain knowledge, but the ability to propose diverse configurations, evaluate them systematically, and submit the best identified solution. Overall, InferenceBench reflects the ability of agents to operate in an open-ended AI engineering setting, where memorized solutions lead to limited improvements.

Emerging Large Language Model (LLM) system patterns, such as disaggregated inference, Mixture-of-Experts (MoE) routing, and asynchronous reinforcement fine-tuning, require flexible point-to-point communication beyond simple collectives. Existing implementations are locked to specific Network Interface Controllers (NICs), hindering integration into inference engines and portability across hardware providers. We present TransferEngine, which bridges the functionality of common NICs to expose a uniform interface. TransferEngine exposes one-sided WriteImm operations with a ImmCounter primitive for completion notification, without ordering assumptions of network transport, transparently managing multiple NICs per GPU. We demonstrate peak throughput of 400 Gbps on both NVIDIA ConnectX-7 and AWS Elastic Fabric Adapter (EFA). We showcase TransferEngine through three production systems: (1) KvCache transfer for disaggregated inference with dynamic scaling, (2) RL weight updates achieving 1.3 seconds for trillion-parameter models, and (3) MoE dispatch/combine implementation exceeding DeepEP decode latency on ConnectX-7, with the first viable latencies on EFA. We demonstrate that our portable point-to-point communication complements collectives while avoiding lock-in.

Smaller activation outliers do not necessarily imply better low-bit quantization: their alignment with the quantizer matters. We introduce PrismQuant, a quantizer-aware rotation framework that aligns the leading activation eigenspace with the constant group subspace of asymmetric grouped INT4. The affine offsets represent the energy in this subspace without widening the range within the group. We formulate rotation design as a Ky Fan trace maximization and derive a closed-form solution that is provably optimal for this alignment objective. Compact Householder transformations and their compact-WY representation enable gradient-free construction and efficient application at both foldable and online sites. A predictive range law further connects unaligned activation energy and group size to quantization-relevant variation. Experiments on Llama, Qwen, and Mistral span dense models up to 70B parameters and a 30B mixture-of-experts model. Under W4A4KV4, PrismQuant sets the state of the art on Llama-3.2-3B among the compared methods in both perplexity and accuracy. On Llama-3.1-70B, it attains 3.85 perplexity and 72.46% average zero-shot accuracy, only 0.22 percentage points below full precision. In the deployment study on Llama-3.1-8B, our optimized implementation achieves 1.51x prefill and 1.22x CUDA Graph decode speedups over matched FP16 baselines, with 56.34% lower decode peak memory and only 2.35% additional Graph decode latency over Hadamard. Code is available at https://github.com/ForeverBlue816/PrismQuant.

Long-context LLM inference is bottlenecked by quadratic attention computation and growing KV-cache costs. Existing sparse attention and KV-compression methods typically decide which tokens or heads to preserve from runtime attention scores, observation windows, calibration prompts, or learned gates, making head diagnosis input-dependent and costly to deploy. We propose Autonomy-of-Heads (AoH), a data-free method that identifies retrieval and streaming heads from the spectral geometry of query-key projections. AoH defines the kernel attention operator M_h = W_K^{htop}W_Q^h and uses its effective-rank as a weight-space measure of head function: concentrated spectra indicate a small number of dominant query-key matching directions and are associated with retrieval heads, whereas diffuse spectra indicate the absence of a dominant global matching direction and are associated with streaming heads. We further derive an efficient d_head-dimensional computation that avoids constructing the full d_modeltimes d_model matrix. We conducted extensive experiments across models demonstrating that at 50\% sparsity, AoH retains 96.5\% of Full Attention performance on average while reducing prefill and decode latency by up to 41.4\% and 66.0\%, respectively, and KV-cache memory by 50.0\% at 256K tokens.

Mixture-of-experts models make nearly trillion-parameter capacity accessible with sparse per-token computation, provided that the serving system can distribute the weights and coordinate their execution. We present Cascadia's resident execution of Inkling, a 975B-total/41B-active-parameter model, on eleven Intel Core Ultra X7 358H AI PCs, each with 64 GB of memory, Arc B390 integrated graphics and gigabit Ethernet. We contribute a custom resident MoE engine that preserves Inkling's routing rules, constructs compressed graphs for OpenVINO's fused iGPU primitives, and coordinates FP16 expert computation with FP32 output restoration. The engine fits six consecutive decoder layers per machine and represents dense feed-forward blocks as all-active expert slices, reducing measured dense-layer call time from approximately 8.1 to 4.5 ms. A streaming pipeline coordinates concurrent generation, while captured-state draft evaluation measures agreement with the deployed numerical path. Paired measurements at fifteen concurrency levels from 1 to 176 streams reach 60.29 aggregate decode tokens/s at 88 streams, with 46.87 tokens/s over the complete serving phases. At fifteen streams, median first-token latency is 6.05 s. Raising the context budget from the 1,024-position default, real prompts of 1k to 64k tokens recover the embedded code in all 19 measured answers, with first-token time growing as aN+bN^2 and decode latency growing approximately linearly, both bounded by a single-threaded CPU attention loop rather than by memory, which holds 512k positions per stream. Evaluation on captured fleet states separates the effects of vocabulary selection and weight quantization on draft agreement. Together, these contributions establish an execution and evaluation approach for large sparse models on distributed client systems with shared CPU-GPU memory.

Scaling inference-time compute has emerged as an important driver of LLM performance, making inference efficiency a central focus of model design alongside model quality. While the current Transformer-based models deliver strong model quality, their quadratic compute and linear memory make inference expensive. This has spurred the development of sub-quadratic models with reduced linear compute and constant memory requirements. However, many recent linear models trade off model quality and capability for algorithmic efficiency, failing on tasks such as state tracking. Moreover, their theoretically linear inference remains hardware-inefficient in practice. Guided by an inference-first perspective, we introduce three core methodological improvements inspired by the state space model (SSM) viewpoint of linear models. We combine: (1) a more expressive recurrence derived from SSM discretization, (2) a complex-valued state update rule that enables richer state tracking, and (3) a multi-input, multi-output (MIMO) formulation for better model performance without increasing decode latency. Together with architectural refinements, our Mamba-3 model achieves significant gains across retrieval, state-tracking, and downstream language modeling tasks. At the 1.5B scale, Mamba-3 improves average downstream accuracy by 0.6 percentage points compared to the next best model (Gated DeltaNet), with Mamba-3's MIMO variant further improving accuracy by another 1.2 points for a total 1.8 point gain. Across state-size experiments, Mamba-3 achieves comparable perplexity to Mamba-2 despite using half of its predecessor's state size. Our evaluations demonstrate Mamba-3's ability to advance the performance-efficiency Pareto frontier.

Multi-turn LLM serving accumulates dialogue history whose Key-Value (KV) cache grows with every turn and every user, quickly exceeding the model weights themselves and making memory -- not compute -- the binding constraint on throughput. Non-uniform KV compression, which allocates heterogeneous budgets across attention heads, preserves accuracy far better than uniform schemes, yet remains impractical: modern serving stacks assume identical KV lengths across heads, so heterogeneity traps freed memory as page fragmentation, spends up to 25% of prefill time reclaiming scattered pages, and skews GPU workloads that inflate decode latency by up to 1.7times or burn 15--20% of each decode step on re-planning. We observe that this heterogeneity need not be discovered at runtime: head-wise retention follows a two-level structural regularity -- an input-invariant head ranking with narrowly bounded per-head ratios -- that can be calibrated offline from as few as 50 samples. Building on this insight, we present Tangram, a serving framework that statically resolves what prior systems handle dynamically: Budget Reservation fixes each head's post-compression footprint at scheduling time, eliminating page reclamation; Ragged Paging clusters similar-budget heads into independent page tables, turning fragmentation into reclaimable memory; and Ahead-of-Time Load Balancing precomputes balanced GPU partitions with zero runtime planning. Implemented on vLLM, Tangram serves as a drop-in substrate for existing non-uniform compression methods, matching their accuracy while improving end-to-end throughput by up to 2.6times over the full-KV baseline. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

Serving large language models at long context is bottlenecked by the key-value (KV) cache, which is read in full at every decode step. Attention keys are locally low-rank though globally high-rank: a fixed low-rank sketch shared across pages is provably blind to page-specific directions, while at the same summary size a page's own basis ranks pages and keeps carriers far better. LOCKS gives every page its own rank-r spectral summary (resident, a tenth of the cache at r{=}8 and a twenty-fifth at r{=}2), reconstructs within-page logits, estimates each page's attention mass by log-sum-exp, and attends only the top pages; selection itself reads no candidate keys or values. Selecting on this summary alone stays within about a point of the full cache on long-document QA (LongBench-v1; Llama-3.1-8B), tracks the read-every-key exact-LSE oracle on retrieval-dense RULER down to the smallest budgets, and holds quality furthest into the small-budget regime on long-form reasoning (AIME26, MATH-500; Qwen3-4B), where selectors and eviction-based reasoning compressors both fall away. At a 2048-token budget LOCKS matches FullKV aggregate quality at 100K+ context (GLM-4-9B-Chat-1M) while attending about 2% of the tokens; since the summary is scanned in full each step, the per-step KV read falls by about 10--25times across that rank range, and this halves per-token decode latency (2.0times at 1M tokens on one H200 NVL, measured at r{=}8) against dense attention. LOCKS ships as a drop-in plugin for unmodified vLLM, with batched decode running in full CUDA graphs.

Large language models (LLMs) excel across diverse tasks but face significant deployment challenges due to high inference costs. LLM inference comprises prefill (compute-bound) and decode (memory-bound) stages, with decode dominating latency particularly for long sequences. Current decoder-only models handle both stages uniformly, despite their distinct computational profiles. We propose OverFill, which decouples these stages to optimize accuracy-efficiency tradeoffs. OverFill begins with a full model for prefill, processing system and user inputs in parallel. It then switches to a dense pruned model, while generating tokens sequentially. Leveraging more compute during prefill, OverFill improves generation quality with minimal latency overhead. Our 3B-to-1B OverFill configuration outperforms 1B pruned models by 83.2%, while the 8B-to-3B configuration improves over 3B pruned models by 79.2% on average across standard benchmarks. OverFill matches the performance of same-sized models trained from scratch, while using significantly less training data. Our code is available at https://github.com/friendshipkim/overfill.

Most practical high-resolution text-to-image systems, including latent diffusion and autoregressive models, perform generation in a compact latent space, and a decoder maps the generated latents back to pixels. Yet the latent-to-pixel decoder is reconstruction-oriented, optimized to invert the encoder rather than synthesize more details, and becomes increasingly costly at megapixel scale. This drawback calls for a more expressive and efficient decoding paradigm. Motivated by recent progress in scalable pixel-space diffusion, we introduce PiD, a Pixel diffusion Decoder that reformulates latent decoding as conditional pixel diffusion, unifying decoding and upsampling into one generative module. By denoising directly in high-resolution pixel space, PiD synthesizes 4times and even 8times upscaled images with low latency. For latent conditioning, a lightweight sigma-aware adapter injects noise-corrupted latents into the pixel diffusion backbone, enabling PiD to decode partially denoised latents and terminate the latent diffusion process early. To further improve efficiency, we distill the model using DMD2, reducing inference to just 4 steps. PiD applies to both conventional VAE latents and semantic latents (e.g., SigLIP, DINOv2) used in recent RAE-based models. PiD decodes latents of 512 times 512 images into 2048 times 2048 pixels in under 1 second with 13 GB peak memory on a consumer RTX 5090, and as fast as 210 ms on a GB200 GPU, about 6times faster than cascaded diffusion-based super-resolution pipelines with better visual fidelity.

End-to-end (E2E) models have shown to outperform state-of-the-art conventional models for streaming speech recognition [1] across many dimensions, including quality (as measured by word error rate (WER)) and endpointer latency [2]. However, the model still tends to delay the predictions towards the end and thus has much higher partial latency compared to a conventional ASR model. To address this issue, we look at encouraging the E2E model to emit words early, through an algorithm called FastEmit [3]. Naturally, improving on latency results in a quality degradation. To address this, we explore replacing the LSTM layers in the encoder of our E2E model with Conformer layers [4], which has shown good improvements for ASR. Secondly, we also explore running a 2nd-pass beam search to improve quality. In order to ensure the 2nd-pass completes quickly, we explore non-causal Conformer layers that feed into the same 1st-pass RNN-T decoder, an algorithm called Cascaded Encoders [5]. Overall, we find that the Conformer RNN-T with Cascaded Encoders offers a better quality and latency tradeoff for streaming ASR.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

Different real-time speech applications impose distinct latency budgets, often requiring separately trained enhancement models for each scenario. In this paper, we propose a one-for-all, real-time universal speech enhancement model that provides explicit control over both algorithmic and computational latency. Algorithmic latency is flexibly adjusted via configurable look-ahead frames. To avoid learning inefficiency caused by varying padding configurations, we introduce parallel convolutional layers corresponding to different look-ahead settings. Computational latency is controlled through an early-exit mechanism, enabling inference at different network depths. To narrow the performance gap between specialized and flexible models, we propose a two-stage training strategy with a shared-to-multiple decoder transition. Overall, the proposed framework enables a single model to be deployed across diverse latency budgets without retraining separate models.

Auto-regressive Large Language Models (LLMs) demonstrate remarkable performance across different domains such as vision and language processing. However, due to sequential processing through a stack of transformer layers, autoregressive decoding faces significant computation/latency challenges, particularly in resource-constrained environments like mobile and edge devices. Existing approaches in literature that aim to improve latency via skipping layers have two distinct flavors - 1) Early exit, and 2) Input-agnostic heuristics where tokens exit at pre-determined layers irrespective of input sequence. Both the above strategies have limitations - the former cannot be applied to handle KV Caching necessary for speed-ups in modern framework and the latter does not capture the variation in layer importance across tasks or more generally, across input sequences. To address both limitations, we propose FiRST, an algorithm that reduces inference latency by using layer-specific routers to select a subset of transformer layers adaptively for each input sequence - the prompt (during the prefill stage) decides which layers will be skipped during decoding. FiRST preserves compatibility with KV caching enabling faster inference while being quality-aware. FiRST is model-agnostic and can be easily enabled on any pre-trained LLM. Our approach reveals that input adaptivity is critical - indeed, different task-specific middle layers play a crucial role in evolving hidden representations depending on tasks. Extensive experiments show that FiRST significantly reduces latency while outperforming other layer selection strategies in quality metics. It retains competitive performance to base model (without layer skipping) and in some cases, even improves upon it. FiRST is thus a promising and efficient solution for LLM deployment in low-resource environments.

Large Language Models (LLMs) are increasingly deployed in both latency-sensitive online services and cost-sensitive offline workloads. Co-locating these workloads on shared serving instances can improve resource utilization, but directly applying this approach to Prefill/Decode (P/D) disaggregated systems introduces severe load imbalance, as fluctuating request mixes alter the intrinsic P/D ratio. Existing dynamic adjustment techniques cannot keep up with the bursty traffic patterns of online services. We propose a latency-constraint disaggregated architecture, which separates cluster resources into latency-strict and latency-relaxed pools based on task latency requirements. This design enables flexible placement of offline decode tasks, mitigating P/D imbalance while preserving online performance. To fully exploit this flexibility, we propose (1) a bottleneck-based scheduler guided by a Roofline-based performance model for performance bottleneck based scheduling, and (2) a fast preemption mechanism that strictly enforces Service Level Objectives (SLOs) for online requests. Experiments on real-world traces show that compared to existing offline system approaches, our method improves offline throughput by up to 3x, while maintaining online request SLOs.

Speculative decoding has emerged as a powerful method to improve latency and throughput in hosting large language models. However, most existing implementations focus on generating a single sequence. Real-world generative AI applications often require multiple responses and how to perform speculative decoding in a batched setting while preserving its latency benefits poses non-trivial challenges. This paper describes a system of batched speculative decoding that sets a new state of the art in multi-sequence generation latency and that demonstrates superior GPU utilization as well as quality of generations within a time budget. For example, for a 7.8B-size model on a single A100 GPU and with a batch size of 8, each sequence is generated at an average speed of 5.8ms per token, the overall throughput being 1.1K tokens per second. These results represent state-of-the-art latency and a 2.15X speed-up over optimized regular decoding. Within a time budget that regular decoding does not finish, our system is able to generate sequences with HumanEval Pass@First of 43% and Pass@All of 61%, far exceeding what's feasible with single-sequence speculative decoding. Our peak GPU utilization during decoding reaches as high as 15.8%, more than 3X the highest of that of regular decoding and around 10X of single-sequence speculative decoding.

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt to produce one output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler Sarathi-Serve inspired by the techniques we originally proposed for optimizing throughput in Sarathi. Sarathi-Serve leverages chunked-prefills from Sarathi to create stall-free schedules that can add new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Our evaluation shows that Sarathi-Serve improves serving throughput within desired latency SLOs of Mistral-7B by up to 2.6x on a single A100 GPU and up to 6.9x for Falcon-180B on 8 A100 GPUs over Orca and vLLM.

GPU collective communication is typically optimized for bandwidth, yet many emerging workloads are increasingly limited by latency. Long-context decode-heavy large language model (LLM) inference is a prime example, where serving large models requires multiple GPUs, and many small collectives lie directly on the critical path of token generation. Therefore, even microsecond of overhead can impact performance and cost. In this work, we study how to approach the hardware Speed-of-Light (SoL) lower bound for GPU collectives within a scale-up network. We identify key principles for near-optimal designs, including barrier-free synchronization and efficient use of symmetric memory and multicast. Building on NCCL's device-side API, we develop low-latency interfaces for constructing custom collective kernels and use them to implement new symmetric collectives in NCCL. Microbenchmarks show substantial latency reductions for small and medium messages, reducing overhead to within 7% of the absolute SoL lower bound. When integrated into real applications, these kernels improve inter-token latency and throughput in LLM inference and accelerate cuSOLVERMp, demonstrating benefits for both AI inference and traditional HPC workloads.

Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.

Test-Time Scaling (TTS) has proven effective in improving the performance of Large Language Models (LLMs) during inference. However, existing research has overlooked the efficiency of TTS from a latency-sensitive perspective. Through a latency-aware evaluation of representative TTS methods, we demonstrate that a compute-optimal TTS does not always result in the lowest latency in scenarios where latency is critical. To address this gap and achieve latency-optimal TTS, we propose two key approaches by optimizing the concurrency configurations: (1) branch-wise parallelism, which leverages multiple concurrent inference branches, and (2) sequence-wise parallelism, enabled by speculative decoding. By integrating these two approaches and allocating computational resources properly to each, our latency-optimal TTS enables a 32B model to reach 82.3% accuracy on MATH-500 within 1 minute and a smaller 3B model to achieve 72.4% within 10 seconds. Our work emphasizes the importance of latency-aware TTS and demonstrates its ability to deliver both speed and accuracy in latency-sensitive scenarios.

Recent advancements in Mixture of Experts (MoE) models have significantly increased their parameter scale as well as model performance. Extensive offloading techniques have been proposed to address the GPU memory limitations of MoE inference. However, due to the I/O bottleneck and sparse computation of MoE models, existing offloading techniques still suffer from low hardware utilization. To fully utilize the hardware resources, we propose SpecMoEOff, which employs the speculative decoding technique to enlarge the workload of each expert. SpecMoEOff orchestrates the GPU and CPU by both theoretical and empirical roofline analysis. In addition, we develop a dedicated CPU chunked attention verification kernel to fit the speculative decoding in offloading scenarios as well as minimizing the additional overhead led by draft models. SpecMoEOff further integrates an optimizer to automatically tune the hyperparameters of speculative decoding under given hardware and workload. Experimental results show that SpecMoEOff achieves up to 2.5x decode throughput improvement over the state-of-the-art MoE offloading techniques.

Transformer-based models have emerged as one of the most widely used architectures for natural language processing, natural language generation, and image generation. The size of the state-of-the-art models has increased steadily reaching billions of parameters. These huge models are memory hungry and incur significant inference latency even on cutting edge AI-accelerators, such as GPUs. Specifically, the time and memory complexity of the attention operation is quadratic in terms of the total context length, i.e., prompt and output tokens. Thus, several optimizations such as key-value tensor caching and FlashAttention computation have been proposed to deliver the low latency demands of applications relying on such large models. However, these techniques do not cater to the computationally distinct nature of different phases during inference. To that end, we propose LeanAttention, a scalable technique of computing self-attention for the token-generation phase (decode-phase) of decoder-only transformer models. LeanAttention enables scaling the attention mechanism implementation for the challenging case of long context lengths by re-designing the execution flow for the decode-phase. We identify that the associative property of online softmax can be treated as a reduction operation thus allowing us to parallelize the attention computation over these large context lengths. We extend the "stream-K" style reduction of tiled calculation to self-attention to enable parallel computation resulting in an average of 2.6x attention execution speedup over FlashAttention-2 and up to 8.33x speedup for 512k context lengths.

Disaggregated inference has become an essential framework that separates the prefill (P) and decode (D) stages in large language model inference to improve throughput. However, the KV cache transfer faces significant delays between prefill and decode nodes. The block-wise calling method and discontinuous KV cache memory allocation increase the number of calls to the transmission kernel. Additionally, existing frameworks often fix the roles of P and D nodes, leading to computational imbalances. In this paper, we propose FlowKV, a novel disaggregated inference framework, which reduces the average transmission latency of KV cache by 96%, from 0.944s to 0.053s, almost eliminating the transfer time relative to the total request latency by optimizing the KV cache transfer. FlowKV introduces the Load-Aware Scheduler for balanced request scheduling and flexible PD node allocation. This design maximizes hardware resource utilization, achieving peak system throughput across various scenarios, including normal, computational imbalance, and extreme overload conditions. Experimental results demonstrate that FlowKV significantly accelerates inference by 15.2%-48.9% on LongBench dataset compared to the baseline and supports applications with heterogeneous GPUs.

In prefill-decode (PD) disaggregated LLM serving, each request is assigned to a decode worker after prefill. Existing decode routers balance only load; for mixture-of-experts (MoE) models this is incomplete: equally loaded workers can differ in latency, since each decode step loads the weights of every distinct expert its batch activates. We present ELDR, an expert-locality-aware decode router for PD-disaggregated MoE serving. From a request's prefill expert activations, ELDR builds an expert signature predicting the experts it will activate during generation. Offline, balanced K-means partitions signature space across decode workers; online, locality-band routing sends each request to the least-loaded worker among those best matching its signature. A signature cache, co-indexed with the KV cache at KV-block granularity, keeps signatures exact under prefix caching. Implemented in vLLM and evaluated on deployments of up to 40 GPUs, ELDR reduces median TPOT by 5.9-13.9% over the strongest of four load-balancing baselines across three MoE models and two workloads, with model outputs unchanged.

Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency.

Monolithic serving with chunked prefill improves GPU utilization by batching prefill and decode together, but suffers from fine-grained phase interference. Engine-level prefill-decode (PD) disaggregation avoids interference but incurs higher hardware and coordination overhead. Prior intra-GPU disaggregation approaches multiplex prefill and decode within a single GPU, using SLO-based tuning guided by heuristics from offline profiling or reactive feedback loops. However, these methods respond reactively to performance issues rather than anticipating them, limiting adaptability under dynamic workloads. We ask: can we achieve proactive intra-GPU disaggregation that adapts effectively to dynamic workloads? The key challenge lies in managing the conflicting resource demands of prefill and decode under varying conditions. We first show that GPU resources exhibit diminishing returns -- beyond a saturation point, more allocation yields minimal latency benefit. Second, we observe that memory bandwidth contention becomes a critical bottleneck. These insights motivate a design that dynamically partitions GPU resources across prefill and decode phases, while jointly considering compute capacity, memory footprint, and bandwidth contention. Evaluated on diverse LLMs and workloads, our system Nexus achieves up to 2.2x higher throughput, 20x lower TTFT, and 2.5x lower TBT than vLLM; outperforms SGLang by up to 2x; and matches or exceeds disaggregated vLLM.

Existing streaming multimodal models process observations incrementally but still follow a turn-based prefill-then-decode pattern, making them non-duplex: new observations cannot naturally enter an active generation stream. Proactive alternatives use micro-turn polling or external response gates, which fragment continuous interaction, decouple response timing from language generation, and complicate KV-cache-friendly serving. We introduce Aero Realtime, a 4B streaming multimodal model with a duplex architecture for realtime generation. Aero Realtime aligns video, audio, and textual output on a shared temporal grid, where each approximately 80-ms audio slot predicts either a lexical token or a silence token. This allows input and output to advance together, enabling one autoregressive objective to learn both when to respond and what to generate. During inference, Aero Realtime appends only the newest multimodal slot, carries forward the previous output state, and reuses the KV cache for efficient incremental execution. We further provide a complete training and serving recipe, including realtime QA construction, slot-aligned supervision, hardware-aware distributed training, and resumable inference. On four NVIDIA A6000 workstation GPUs, Aero Realtime maintains 84-ms median and 173-ms P95 processing lag over 20 minutes of a continuously streamed video, remaining within 200~ms of the source timeline. These results demonstrate the feasibility of fully aligned input-output modeling for duplex, proactive, and hardware-aligned multimodal interaction.

With the emergence of wearable devices and other embedded systems, deploying large language models (LLMs) on edge platforms has become an urgent need. However, this is challenging because of their high computational and memory demands. Although recent low-bit quantization methods (e.g., BitNet, DeepSeek) compress weights to as low as 1.58~bits with minimal accuracy loss, edge deployment is still constrained by limited on-chip resources, power budgets, and the often-neglected long latency of the prefill stage. We present TeLLMe, the first table-lookup-based ternary LLM accelerator for low-power edge FPGAs that fully supports both prefill and autoregressive decoding using 1.58-bit weights and 8-bit activations. TeLLMe incorporates several novel techniques, including (1) a table-lookup-based ternary matrix multiplication (TLMM) engine utilizing grouped activations and online precomputation for low resource utilization and high throughput; (2) a fine-grained analytic URAM-based weight buffer management scheme for efficient loading and compute engine access; (3) a streaming dataflow architecture that fuses floating-point element-wise operations with linear computations to hide latency; (4) a reversed-reordered prefill stage attention with fused attention operations for high memory efficiency; and (5) a resource-efficient specialized decoding stage attention. Under a 5~W power budget, TeLLMe delivers up to 25~tokens/s decoding throughput and 0.45--0.96~s time-to-first-token (TTFT) for 64--128 token prompts, marking a significant energy-efficiency advancement in LLM inference on edge FPGAs.

Long-context inference in decoder-only language models is costly because long prompts are processed during Prefill, cached at every layer, and repeatedly attended to during autoregressive Decode. We introduce Shallow Prefill, dEEp Decode (SPEED), a phase-asymmetric KV-visibility policy that materializes non-anchor prompt-token KV states only in lower layers while keeping Decode-phase tokens full-depth. Unlike previous approaches that make upper-layer prompt KV states cheaper to store or construct, SPEED removes prefill tokens from the upper-layer Decode visibility set altogether. With a minimal BoS anchor, this simple change preserves broad benchmark quality while reducing long-context cost. In a controlled Llama-3.1-8B instruction-tuning study, SPEED using only 75\% of layers for prefill tokens reaches 51.2 average score on OLMES-style benchmarks, compared with 51.4 for the full-depth baseline, while improving TTFT by 33\%, TPOT by 22\%, and reducing active KV memory by 25.0\% at 128K context. Layer-wise diagnostics suggest that this cutoff retains the main prompt-selection and representation-stabilization regions of the full-depth model. These results show that long-context prompt tokens need not always persist as full-depth KV-cache objects when Decode-phase tokens remain full-depth.

Thus far, end-to-end (E2E) models have not been shown to outperform state-of-the-art conventional models with respect to both quality, i.e., word error rate (WER), and latency, i.e., the time the hypothesis is finalized after the user stops speaking. In this paper, we develop a first-pass Recurrent Neural Network Transducer (RNN-T) model and a second-pass Listen, Attend, Spell (LAS) rescorer that surpasses a conventional model in both quality and latency. On the quality side, we incorporate a large number of utterances across varied domains to increase acoustic diversity and the vocabulary seen by the model. We also train with accented English speech to make the model more robust to different pronunciations. In addition, given the increased amount of training data, we explore a varied learning rate schedule. On the latency front, we explore using the end-of-sentence decision emitted by the RNN-T model to close the microphone, and also introduce various optimizations to improve the speed of LAS rescoring. Overall, we find that RNN-T+LAS offers a better WER and latency tradeoff compared to a conventional model. For example, for the same latency, RNN-T+LAS obtains a 8% relative improvement in WER, while being more than 400-times smaller in model size.

Recent advances with large language models (LLM) illustrate their diverse capabilities. We propose a novel algorithm, staged speculative decoding, to accelerate LLM inference in small-batch, on-device scenarios. We address the low arithmetic intensity of small-batch inference by improving upon previous work in speculative decoding. First, we restructure the speculative batch as a tree, which reduces generation costs and increases the expected tokens per batch. Second, we add a second stage of speculative decoding. Taken together, we reduce single-batch decoding latency by 3.16x with a 762M parameter GPT-2-L model while perfectly preserving output quality.

Recent work in speech-to-speech translation (S2ST) has focused primarily on offline settings, where the full input utterance is available before any output is given. This, however, is not reasonable in many real-world scenarios. In latency-sensitive applications, rather than waiting for the full utterance, translations should be spoken as soon as the information in the input is present. In this work, we introduce a system for simultaneous S2ST targeting real-world use cases. Our system supports translation from 57 languages to English with tunable parameters for dynamically adjusting the latency of the output -- including four policies for determining when to speak an output sequence. We show that these policies achieve offline-level accuracy with minimal increases in latency over a Greedy (wait-k) baseline. We open-source our evaluation code and interactive test script to aid future SimulS2ST research and application development.

Recent work leverages Vision Foundation Models as image encoders to boost the generative performance of latent diffusion models (LDMs), as their semantic feature distributions are easy to learn. However, such semantic features often lack low-level information (\eg, color and texture), leading to degraded reconstruction fidelity, which has emerged as a primary bottleneck in further scaling LDMs. To address this limitation, we propose LV-RAE, a representation autoencoder that augments semantic features with missing low-level information, enabling high-fidelity reconstruction while remaining highly aligned with the semantic distribution. We further observe that the resulting high-dimensional, information-rich latent make decoders sensitive to latent perturbations, causing severe artifacts when decoding generated latent and consequently degrading generation quality. Our analysis suggests that this sensitivity primarily stems from excessive decoder responses along directions off the data manifold. Building on these insights, we propose fine-tuning the decoder to increase its robustness and smoothing the generated latent via controlled noise injection, thereby enhancing generation quality. Experiments demonstrate that LV-RAE significantly improves reconstruction fidelity while preserving the semantic abstraction and achieving strong generative quality. Our code is available at https://github.com/modyu-liu/LVRAE.

Latent audio generative models are typically trained in two stages: an audio codec is learned first, followed by a latent generative model. This decomposition leads to a decoder train-generation mismatch: the codec decoder is trained on encoder-induced latents but deployed on generator-produced latents at inference time. Across diverse datasets and latent generative models, we observe clear reconstruction-generation gaps under both FD and FAD, showing that strong reconstruction quality does not necessarily translate into strong end-to-end generation quality. A natural remedy is to adapt the decoder on generation-produced latents, but generated latents lack correspondence with source audio and therefore cannot directly provide the paired supervision used for decoder fine-tuning. We introduce AudioGAR, which constructs intermediate latents by perturbing encoder latents and denoising them through the frozen latent diffusion model. These latents form a trajectory from reconstruction toward generation, with lower-noise latents retaining source correspondence and supporting paired decoder fine-tuning. We fine-tune only the codec decoder on these latents, while keeping the codec encoder and latent generative model frozen. When applied to AudioX, AudioGAR substantially improves generative performance. It requires only 1.5\% of the original training audio hours and 0.26\% of the original training cost.

Fast Automatic Speech Recognition (ASR) is critical for latency-sensitive applications such as real-time captioning and meeting transcription. However, truly parallel ASR decoding remains challenging due to the sequential nature of autoregressive (AR) decoders and the context limitations of non-autoregressive (NAR) methods. While modern ASR encoders can process up to 30 seconds of audio at once, AR decoders still generate tokens sequentially, creating a latency bottleneck. We propose Whisfusion, the first framework to fuse a pre-trained Whisper encoder with a text diffusion decoder. This NAR architecture resolves the AR latency bottleneck by processing the entire acoustic context in parallel at every decoding step. A lightweight cross-attention adapter trained via parameter-efficient fine-tuning (PEFT) bridges the two modalities. We also introduce a batch-parallel, multi-step decoding strategy that improves accuracy by increasing the number of candidates with minimal impact on speed. Fine-tuned solely on LibriSpeech (960h), Whisfusion achieves a lower WER than Whisper-tiny (8.3% vs. 9.7%), and offers comparable latency on short audio. For longer utterances (>20s), it is up to 2.6x faster than the AR baseline, establishing a new, efficient operating point for long-form ASR. The implementation and training scripts are available at https://github.com/taeyoun811/Whisfusion.

Compressed short-text generators can fail in two different places: the codec may discard information before generation starts, or the latent generator may produce weak codes. Without separating these failure modes, researchers can spend compute improving the wrong component. We study this problem in a controlled 64-to-16 TinyStories case study built from a hierarchical VQ-VAE-2 codec and a masked discrete diffusion generator (MDLM). We use a staged validation protocol that separates codec reconstruction fidelity, latent generation quality, and auxiliary latent diagnostics under one shared external GPT-2 scorer, while reporting complementary semantic metrics for the geometry study. In the tested configuration, codec reconstruction alone raises median external perplexity from 15.17 to 27.36 (+80.4%) and p95 from 25.10 to 98.91 (+294.1%), showing that the dominant quality loss appears before latent generation begins. Under the same scorer, code-space MDLM remains materially stronger than token-space diffusion, reducing mean, median, and p95 by 32.9%, 30.9%, and 36.6%, respectively. Geometry-aware regularization improves local latent proxies but does not improve decoded-text metrics in the available runs. The contribution is methodological rather than algorithmic: the paper presents a reusable staged diagnosis for one concrete pipeline and shows that, in this setting, codec fidelity rather than latent denoising sets the practical quality ceiling.

Autoregressive (AR) models remain the standard for natural language generation but still suffer from high latency due to strictly sequential decoding. Recent diffusion-inspired approaches, such as LlaDA and Dream, mitigate this by generating in parallel, yet they suffer from two core limitations: information loss, as predictive distributions for non-finalized tokens are discarded at each step, and premature commitment, where local decisions are made without sufficient global coordination. We introduce Latent Refinement Decoding (LRD), a two-stage framework with Latent Refinement and a Predictive Feedback Loop. The first stage maintains masked positions as distributional mixtures of predicted tokens and the mask embedding, allowing the model to establish more globally consistent beliefs. The second stage progressively finalizes confident tokens while retaining uncertain ones for iterative feedback. KL-divergence dynamics provide a principled and reliable criterion for convergence and early stopping. Experiments across coding (HumanEval +6.3, MBPP +2.6) and reasoning (GSM8K +2.9, MATH500 +3.8) show that LRD improves accuracy while delivering speedups of up to 10.6x, making it a strong and versatile alternative for parallel sequence generation.

Training-free acceleration has emerged as an advanced research area in video generation based on diffusion models. The redundancy of latents in diffusion model inference provides a natural entry point for acceleration. In this paper, we decompose the inference process into the encoding, denoising, and decoding stages, and observe that cache-based acceleration methods often lead to substantial memory surges in the latter two stages. To address this problem, we analyze the characteristics of inference across different stages and propose stage-specific strategies for reducing memory consumption: 1) Asynchronous Cache Swapping. 2) Feature chunk. 3) Slicing latents to decode. At the same time, we ensure that the time overhead introduced by these three strategies remains lower than the acceleration gains themselves. Compared with the baseline, our approach achieves faster inference speed and lower memory usage, while maintaining quality degradation within an acceptable range. The Code is available at https://github.com/NKUShaw/LightCache .

Autoregressive decoding in Large Language Models (LLMs) is inherently sequential, creating a latency bottleneck that scales linearly with output length. While ``Decomposition-and-Fill'' methods like Skeleton-of-Thought attempt to parallelize generation via external orchestration, they suffer from coherence drift due to the lack of cross-stream communication. In this work, we introduce the Parallel Decoder Transformer (PDT), a parameter-efficient architecture that embeds coordination primitives directly into the inference process of a frozen pre-trained model. Instead of retraining the base model, PDT injects lightweight Speculative Note Conditioning (SNC) adapters that allow parallel decoding streams to synchronize via a shared, dynamic latent space. We formulate coordination as a speculative consensus problem, where sibling streams broadcast semantic ``notes'' to a global bus, gated by a learned verification head. We validate our approach on a 50,000-step curriculum using a frozen 20B-parameter backbone. Our results demonstrate that PDT achieves effective self-correction, reaching 77.8\% precision in coverage prediction and recovering approximate serial semantics without modifying the trunk weights. This establishes PDT as a scalable, efficient alternative to full model fine-tuning for structured parallel generation.

Large-scale recordings of neural activity are providing new opportunities to study neural population dynamics. A powerful method for analyzing such high-dimensional measurements is to deploy an algorithm to learn the low-dimensional latent dynamics. LFADS (Latent Factor Analysis via Dynamical Systems) is a deep learning method for inferring latent dynamics from high-dimensional neural spiking data recorded simultaneously in single trials. This method has shown a remarkable performance in modeling complex brain signals with an average inference latency in milliseconds. As our capacity of simultaneously recording many neurons is increasing exponentially, it is becoming crucial to build capacity for deploying low-latency inference of the computing algorithms. To improve the real-time processing ability of LFADS, we introduce an efficient implementation of the LFADS models onto Field Programmable Gate Arrays (FPGA). Our implementation shows an inference latency of 41.97 mus for processing the data in a single trial on a Xilinx U55C.

We present LatentCSI, a novel method for generating images of the physical environment from WiFi CSI measurements that leverages a pretrained latent diffusion model (LDM). Unlike prior approaches that rely on complex and computationally intensive techniques such as GANs, our method employs a lightweight neural network to map CSI amplitudes directly into the latent space of an LDM. We then apply the LDM's denoising diffusion model to the latent representation with text-based guidance before decoding using the LDM's pretrained decoder to obtain a high-resolution image. This design bypasses the challenges of pixel-space image generation and avoids the explicit image encoding stage typically required in conventional image-to-image pipelines, enabling efficient and high-quality image synthesis. We validate our approach on two datasets: a wide-band CSI dataset we collected with off-the-shelf WiFi devices and cameras; and a subset of the publicly available MM-Fi dataset. The results demonstrate that LatentCSI outperforms baselines of comparable complexity trained directly on ground-truth images in both computational efficiency and perceptual quality, while additionally providing practical advantages through its unique capacity for text-guided controllability.

For deployment, neural architecture search should be hardware-aware, in order to satisfy the device-specific constraints (e.g., memory usage, latency and energy consumption) and enhance the model efficiency. Existing methods on hardware-aware NAS collect a large number of samples (e.g., accuracy and latency) from a target device, either builds a lookup table or a latency estimator. However, such approach is impractical in real-world scenarios as there exist numerous devices with different hardware specifications, and collecting samples from such a large number of devices will require prohibitive computational and monetary cost. To overcome such limitations, we propose Hardware-adaptive Efficient Latency Predictor (HELP), which formulates the device-specific latency estimation problem as a meta-learning problem, such that we can estimate the latency of a model's performance for a given task on an unseen device with a few samples. To this end, we introduce novel hardware embeddings to embed any devices considering them as black-box functions that output latencies, and meta-learn the hardware-adaptive latency predictor in a device-dependent manner, using the hardware embeddings. We validate the proposed HELP for its latency estimation performance on unseen platforms, on which it achieves high estimation performance with as few as 10 measurement samples, outperforming all relevant baselines. We also validate end-to-end NAS frameworks using HELP against ones without it, and show that it largely reduces the total time cost of the base NAS method, in latency-constrained settings. Code is available at https://github.com/HayeonLee/HELP.

Real-time video generation demands fast decoding as much as fast denoising, yet current latent video diffusion models rely on 3D convolutional decoders that are slow and memory-intensive at high resolutions or for long video. We introduce FlashDecoder, a fast, memory-efficient pure-Transformer video decoder that decodes latents to pixels frame by frame. At each step, the current frame attends only to a fixed-size window of past frames through a rolling KV cache. The fixed temporal window keeps decoding fast and memory bounded regardless of video length, enabling constant-latency streaming. Because frames are processed sequentially, temporal causality is enforced without explicit attention masks, enabling training at resolutions up to 1080p and matching the reconstruction quality of convolutional decoders. On the Wan2.1 and Wan2.2 latent spaces, FlashDecoder matches each convolutional decoder in reconstruction quality (e.g., 41.55dB vs. 41.49dB PSNR at 1080p) while decoding 3.6x-4.7x faster with up to 11x less memory on a single H100 GPU. With architecture-aware inference optimizations, the speedup widens to 12x.

Gaussian-corrupted sentence embeddings have no direct linguistic interpretation, yet continuous diffusion language models can generate fluent text from them. We study this puzzle through Embedded Language Flows (ELF) and identify a decoder-basin mechanism: our evidence suggests that denoising becomes reliable when trajectories reach regions where the native decoder can read stable tokens. We introduce a diagnostic protocol for denoisability, semantic recoverability, order sensitivity, decoder compatibility, and trajectory reliability. It exposes failures hidden by scalar metrics: low mean-squared error can discard linguistic content, low perplexity can reflect low-entropy collapse, and clean latent reconstruction can coexist with a narrow decoder basin. A decoder-margin bound explains why token recovery depends on margin and local decoder sensitivity, not latent error alone. Auditing public ELF checkpoints reveals an interface phase diagram: early predictions are weakly readable, mid-trajectory disagreement marks a competition region, and late predictions enter a high-margin decoder basin. Once inside, token realization is surprisingly simple on generated ELF states: frozen T5 (Text-to-Text Transfer Transformer) token-embedding lookup recovers 93--96% of native decoder decisions, and a single linear readout reaches 97.9% agreement at 32k samples, leaving an approx1.1--1.2 perplexity gap in a structured residual tail. Under conservative held-out gates, a margin rule exits roughly 17--28% earlier in denoising steps under an explicit diagnostic monitor. Boundary checks on LangFlow, BitstreamDiffusion, and the Continuous Latent Diffusion Language Model (Cola-DLM) show that the same interface questions remain meaningful when the state object and decoder change. Continuous and latent diffusion language models should therefore be evaluated as representation-decoder systems.

Simultaneous speech-to-text translation (SimulST) systems have to balance translation quality with latency--the delay between speech input and the translated output. While quality evaluation is well established, accurate latency measurement remains a challenge. Existing metrics often produce inconsistent or misleading results, especially in the widely used short-form setting, where speech is artificially presegmented. In this paper, we present the first comprehensive analysis of SimulST latency metrics across language pairs, systems, and both short- and long-form regimes. We uncover a structural bias in current metrics related to segmentation that undermines fair and meaningful comparisons. To address this, we introduce YAAL (Yet Another Average Lagging), a refined latency metric that delivers more accurate evaluations in the short-form regime. We extend YAAL to LongYAAL for unsegmented audio and propose SoftSegmenter, a novel resegmentation tool based on word-level alignment. Our experiments show that YAAL and LongYAAL outperform popular latency metrics, while SoftSegmenter enhances alignment quality in long-form evaluation, together enabling more reliable assessments of SimulST systems.

Latency-critical speech applications (e.g., live transcription, voice commands, and real-time translation) demand low time-to-first-token (TTFT) and high transcription accuracy, particularly on resource-constrained edge devices. Full-attention Transformer encoders remain a strong accuracy baseline for automatic speech recognition (ASR) because every frame can directly attend to every other frame, which resolves otherwise locally ambiguous acoustics using distant lexical context. However, this global dependency incurs quadratic complexity in sequence length, inducing an inherent "encode-the-whole-utterance" latency profile. For streaming use cases, this causes TTFT to grow linearly with utterance length as the encoder must process the entire prefix before any decoder token can be emitted. To better meet the needs of on-device, streaming ASR use cases we introduce Moonshine v2, an ergodic streaming-encoder ASR model that employs sliding-window self-attention to achieve bounded, low-latency inference while preserving strong local context. Our models achieve state of the art word error rates across standard benchmarks, attaining accuracy on-par with models 6x their size while running significantly faster. These results demonstrate that carefully designed local attention is competitive with the accuracy of full attention at a fraction of the size and latency cost, opening new possibilities for interactive speech interfaces on edge devices.

With the increased applications of automatic speech recognition (ASR) in recent years, it is essential to automatically insert punctuation marks and remove disfluencies in transcripts, to improve the readability of the transcripts as well as the performance of subsequent applications, such as machine translation, dialogue systems, and so forth. In this paper, we propose a Controllable Time-delay Transformer (CT-Transformer) model that jointly completes the punctuation prediction and disfluency detection tasks in real time. The CT-Transformer model facilitates freezing partial outputs with controllable time delay to fulfill the real-time constraints in partial decoding required by subsequent applications. We further propose a fast decoding strategy to minimize latency while maintaining competitive performance. Experimental results on the IWSLT2011 benchmark dataset and an in-house Chinese annotated dataset demonstrate that the proposed approach outperforms the previous state-of-the-art models on F-scores and achieves a competitive inference speed.

Fine-tuning large language models (LLMs) for recommendation in a generative manner has delivered promising results, but encounters significant inference overhead due to autoregressive decoding in the language space. This work explores bypassing language-space decoding by directly matching candidate items with the LLM's internal thought representations in the latent space, eliminating the time-consuming autoregressive process to reduce computational costs. Towards this, we introduce Light Latent-space Decoding (L2D), an effective and efficient latent-space decoding method. L2D represents user-preferred items by using the hidden states of test sequences reflecting the LLM's internal thought, and obtains candidate item representations from the hidden states of training sequences labeled with the corresponding candidate items. It then matches the two types of representations to decode items, achieving latent-space decoding. In this way, it enables efficient decoding without altering the LLM's generative tuning paradigm, thereby preserving performance. Extensive empirical results demonstrate that L2D is more than 10x faster than language-space decoding while maintaining or enhancing performance.

Multimodal language models that process both text and speech have a potential for applications in spoken dialogue systems. However, current models face two major challenges in response generation latency: (1) generating a spoken response requires the prior generation of a written response, and (2) speech sequences are significantly longer than text sequences. This study addresses these issues by extending the input and output sequences of the language model to support the parallel generation of text and speech. Our experiments on spoken question answering tasks demonstrate that our approach improves latency while maintaining the quality of response content. Additionally, we show that latency can be further reduced by generating speech in multiple sequences. Demo samples are available at https://rinnakk.github.io/research/publications/PSLM.

Recent advances in AudioLLMs have enabled spoken dialogue systems to move beyond turn-based interaction toward real-time full-duplex communication, where the agent must decide when to speak, yield, or interrupt while the user is still talking. Existing full-duplex approaches either rely on voice activity cues, which lack semantic understanding, or on ASR-based modules, which introduce latency and degrade under overlapping speech and noise. Moreover, available datasets rarely capture realistic interaction dynamics, limiting evaluation and deployment. To mitigate the problem, we propose FastTurn, a unified framework for low-latency and robust turn detection. To advance latency while maintaining performance, FastTurn combines streaming CTC decoding with acoustic features, enabling early decisions from partial observations while preserving semantic cues. We also release a test set based on real human dialogue, capturing authentic turn transitions, overlapping speech, backchannels, pauses, pitch variation, and environmental noise. Experiments show FastTurn achieves higher decision accuracy with lower interruption latency than representative baselines and remains robust under challenging acoustic conditions, demonstrating its effectiveness for practical full-duplex dialogue systems.

Large reasoning models (LRMs) improve complex problem-solving by generating long intermediate reasoning traces, but this substantially increases inference costs. NVFP4 inference offers a promising approach to reduce both computational and memory costs through hardware-supported low-precision execution. However, directly applying NVFP4 to LRMs introduces two practical limitations: reasoning accuracy degrades under quantization, and existing NVFP4 kernels do not fully realize latency benefits in small-batch autoregressive decoding. In this work, we analyze the effect of NVFP4 quantization on token-level uncertainty during reasoning. We show that quantization increases incorrect sampling at low-entropy symbolic tokens, while causing over-concentration on a small set of tokens in high-uncertainty reasoning steps. Based on this observation, we propose ReSET, a reasoning-step entropy-based temperature-scaling method that estimates step-level uncertainty online and adapts the decoding temperature using both token-level and step-level entropy signals. To address the latency gap, we further design a CUDA-core small-M NVFP4 kernel for latency-critical autoregressive decoding. Across reasoning benchmarks and model scales, ReSET improves NVFP4 reasoning accuracy by up to sim!2 points over the NVFP4 baseline. Our CUDA-core small-M kernel further improves latency-critical decoding, delivering up to 2.5!times kernel-level speedup over NVFP4 vLLM and approximately 2!times end-to-end decoding speedup over BF16. Code is available at https://github.com/aiha-lab/ReSET.

When Masked Diffusion Models (MDMs) generate sequences through iterative refinement, the rich internal computation over masked positions is discarded, forcing every subsequent refinement step to recompute the valuable internal information stored as model representations. To avoid a hard reset between denoising rounds, we propose Learned Relay Representations (Relay), a method that allows MDMs to be forward-thinking when denoising by explicitly learning how to propagate latent information for the benefit of future denoising steps. Relay introduces a differentiable per-token channel that passes information between forward passes and is trained via truncated backpropagation through time (BPTT). We show that this framework can be scaled to state-of-the-art Diffusion Language Models (DLMs), and is seamlessly compatible with techniques like block diffusion and KV caching. We first provide a thorough justification of the design choices in Relay on a challenging Sudoku-based planning task. We then scale Relay to Fast-dLLM v2, a state-of-the-art DLM, outperforming standard supervised finetuning on coding tasks while reducing inference latency by up to 32%. Our empirical results demonstrate that state-of-the-art DLMs can be explicitly trained to relay latent information forward across decoding steps, advancing the performance-latency Pareto frontier. We provide code for all our experiments.

Brain-computer interfaces (BCIs) hold great potential for aiding individuals with speech impairments. Utilizing electroencephalography (EEG) to decode speech is particularly promising due to its non-invasive nature. However, recordings are typically short, and the high variability in EEG data has led researchers to focus on classification tasks with a few dozen classes. To assess its practical applicability for speech neuroprostheses, we investigate the relationship between the size of EEG data and decoding accuracy in the open vocabulary setting. We collected extensive EEG data from a single participant (175 hours) and conducted zero-shot speech segment classification using self-supervised representation learning. The model trained on the entire dataset achieved a top-1 accuracy of 48\% and a top-10 accuracy of 76\%, while mitigating the effects of myopotential artifacts. Conversely, when the data was limited to the typical amount used in practice (sim10 hours), the top-1 accuracy dropped to 2.5\%, revealing a significant scaling effect. Additionally, as the amount of training data increased, the EEG latent representation progressively exhibited clearer temporal structures of spoken phrases. This indicates that the decoder can recognize speech segments in a data-driven manner without explicit measurements of word recognition. This research marks a significant step towards the practical realization of EEG-based speech BCIs.

Due to the auto-regressive nature of current video large language models (Video-LLMs), the inference latency increases as the input sequence length grows, posing challenges for the efficient processing of video sequences that are usually very long. We observe that during decoding, the attention scores of most tokens in Video-LLMs tend to be sparse and concentrated, with only certain tokens requiring comprehensive full attention. Based on this insight, we introduce Sparse-to-Dense (StD), a novel decoding strategy that integrates two distinct modules: one leveraging sparse top-K attention and the other employing dense full attention. These modules collaborate to accelerate Video-LLMs without loss. The fast (sparse) model speculatively decodes multiple tokens, while the slow (dense) model verifies them in parallel. StD is a tuning-free, plug-and-play solution that achieves up to a 1.94times walltime speedup in video processing. It maintains model performance while enabling a seamless transition from a standard Video-LLM to a sparse Video-LLM with minimal code modifications.

Speculative decoding is a relatively new decoding framework that leverages small and efficient draft models to reduce the latency of LLMs. In this study, we introduce GliDe and CaPE, two low-hassle modifications to vanilla speculative decoding to further improve the decoding speed of a frozen LLM. Specifically, GliDe is a modified draft model architecture that reuses the cached keys and values from the target LLM, while CaPE is a proposal expansion method that uses the draft model's confidence scores to help select additional candidate tokens for verification. Extensive experiments on different benchmarks demonstrate that our proposed GliDe draft model significantly reduces the expected decoding latency. Additional evaluation using walltime reveals that GliDe can accelerate Vicuna models up to 2.17x and further extend the improvement to 2.61x with CaPE. We will release our code, data, and the trained draft models.

Diffusion language models (DLMs) generate text through iterative denoising, but inference requires full-sequence attention at every iteration, resulting in substantial redundant computation on masked tokens. Block-wise diffusion can reduce this cost, yet it typically relies on retraining and constrained update orders, limiting its direct applicability to pretrained DLMs. Our token-level analysis reveals pronounced structural locality in DLM inference. Decoding is driven by a small set of prefix-localized active tokens; the influence of distant undecoded context diminishes rapidly, and decoded tokens exhibit stage-wise temporal stability, enabling reuse of intermediate representations except for a brief post-decode transient. Motivated by these observations, we propose \placeholderThe source code is available at https://github.com/vhicrgit/Window-Diffusion., a window-based token pruning and caching method for inference. We maintain a local computation window that slides rightward as denoising progresses, and partition undecoded tokens into: (i) active tokens that are computed online, (ii) buffer tokens whose KV states are cached and periodically refreshed, and (iii) far-field tokens that are pruned outside the window. Computation is restricted to active and buffer tokens within the window, while far-field tokens are omitted at each stage. Experiments on LLaDA and Dream show that, under matched compute budgets, our method achieves up to 99times inference speedup while largely preserving generation performance.

We present Unified Latents (UL), a framework for learning latent representations that are jointly regularized by a diffusion prior and decoded by a diffusion model. By linking the encoder's output noise to the prior's minimum noise level, we obtain a simple training objective that provides a tight upper bound on the latent bitrate. On ImageNet-512, our approach achieves competitive FID of 1.4, with high reconstruction quality (PSNR) while requiring fewer training FLOPs than models trained on Stable Diffusion latents. On Kinetics-600, we set a new state-of-the-art FVD of 1.3.

We propose a novel approach for time-scale modification of audio signals. Unlike traditional methods that rely on the framing technique or the short-time Fourier transform to preserve the frequency during temporal stretching, our neural network model encodes the raw audio into a high-level latent representation, dubbed Neuralgram, where each vector represents 1024 audio sample points. Due to a sufficient compression ratio, we are able to apply arbitrary spatial interpolation of the Neuralgram to perform temporal stretching. Finally, a learned neural decoder synthesizes the time-scaled audio samples based on the stretched Neuralgram representation. Both the encoder and decoder are trained with latent regression losses and adversarial losses in order to obtain high-fidelity audio samples. Despite its simplicity, our method has comparable performance compared to the existing baselines and opens a new possibility in research into modern time-scale modification. Audio samples can be found at https://tsmnet-mmasia23.github.io

There have been significant advances in deep learning for music demixing in recent years. However, there has been little attention given to how these neural networks can be adapted for real-time low-latency applications, which could be helpful for hearing aids, remixing audio streams and live shows. In this paper, we investigate the various challenges involved in adapting current demixing models in the literature for this use case. Subsequently, inspired by the Hybrid Demucs architecture, we propose the Hybrid Spectrogram Time-domain Audio Separation Network HS-TasNet, which utilises the advantages of spectral and waveform domains. For a latency of 23 ms, the HS-TasNet obtains an overall signal-to-distortion ratio (SDR) of 4.65 on the MusDB test set, and increases to 5.55 with additional training data. These results demonstrate the potential of efficient demixing for real-time low-latency music applications.

Automatic Speech Recognition (ASR) has seen remarkable progress, with models like OpenAI Whisper and NVIDIA Canary achieving state-of-the-art (SOTA) performance in offline transcription. However, these models are not designed for streaming (online or real-time) transcription, due to limitations in their architecture and training methodology. We propose a method to turn the transformer encoder-decoder model into a low-latency streaming model that is careless about future context. We present an analysis explaining why it is not straightforward to convert an encoder-decoder transformer to a low-latency streaming model. Our proposed method modifies the existing (non-causal) encoder to a causal encoder by fine-tuning both the encoder and decoder using Low-Rank Adaptation (LoRA) and a weakly aligned dataset. We then propose an updated inference mechanism that utilizes the fine-tune causal encoder and decoder to yield greedy and beam-search decoding, and is shown to be locally optimal. Experiments on low-latency chunk sizes (less than 300 msec) show that our fine-tuned model outperforms existing non-fine-tuned streaming approaches in most cases, while using a lower complexity. Additionally, we observe that our training process yields better alignment, enabling a simple method for extracting word-level timestamps. We release our training and inference code, along with the fine-tuned models, to support further research and development in streaming ASR.

Latent diffusion models have enabled high-quality video synthesis, yet their inference remains costly and time-consuming. As diffusion transformers become increasingly efficient, the latency bottleneck inevitably shifts to VAE decoders. To reduce their latency while maintaining quality, we propose a universal acceleration framework for VAE decoders that preserves full alignment with the original latent distribution. Specifically, we propose (1) an independence-aware channel pruning method to effectively mitigate severe channel redundancy, and (2) a stage-wise dominant operator optimization strategy to address the high inference cost of the widely used causal 3D convolutions in VAE decoders. Based on these innovations, we construct a Flash-VAED family. Moreover, we design a three-phase dynamic distillation framework that efficiently transfers the capabilities of the original VAE decoder to Flash-VAED. Extensive experiments on Wan and LTX-Video VAE decoders demonstrate that our method outperforms baselines in both quality and speed, achieving approximately a 6times speedup while maintaining the reconstruction performance up to 96.9%. Notably, Flash-VAED accelerates the end-to-end generation pipeline by up to 36% with negligible quality drops on VBench-2.0.

Differentiable neural architecture search methods became popular in recent years, mainly due to their low search costs and flexibility in designing the search space. However, these methods suffer the difficulty in optimizing network, so that the searched network is often unfriendly to hardware. This paper deals with this problem by adding a differentiable latency loss term into optimization, so that the search process can tradeoff between accuracy and latency with a balancing coefficient. The core of latency prediction is to encode each network architecture and feed it into a multi-layer regressor, with the training data which can be easily collected from randomly sampling a number of architectures and evaluating them on the hardware. We evaluate our approach on NVIDIA Tesla-P100 GPUs. With 100K sampled architectures (requiring a few hours), the latency prediction module arrives at a relative error of lower than 10%. Equipped with this module, the search method can reduce the latency by 20% meanwhile preserving the accuracy. Our approach also enjoys the ability of being transplanted to a wide range of hardware platforms with very few efforts, or being used to optimizing other non-differentiable factors such as power consumption.

Large language models have demonstrated exceptional capability in natural language understanding and generation. However, their generation speed is limited by the inherently sequential nature of their decoding process, posing challenges for real-time applications. This paper introduces Lexical Unit Decoding (LUD), a novel decoding methodology implemented in a data-driven manner, accelerating the decoding process without sacrificing output quality. The core of our approach is the observation that a pre-trained language model can confidently predict multiple contiguous tokens, forming the basis for a lexical unit, in which these contiguous tokens could be decoded in parallel. Extensive experiments validate that our method substantially reduces decoding time while maintaining generation quality, i.e., 33\% speed up on natural language generation with no quality loss, and 30\% speed up on code generation with a negligible quality loss of 3\%. Distinctively, LUD requires no auxiliary models and does not require changes to existing architectures. It can also be integrated with other decoding acceleration methods, thus achieving an even more pronounced inference efficiency boost. We posit that the foundational principles of LUD could define a new decoding paradigm for future language models, enhancing their applicability for a broader spectrum of applications. All codes are be publicly available at https://github.com/tjunlp-lab/Lexical-Unit-Decoding-LUD-. Keywords: Parallel Decoding, Lexical Unit Decoding, Large Language Model

Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive models. Unlike standard diffusion-based approaches, DLMs are not explicitly conditioned on a timestep, raising a natural question: do these models internally represent denoising progress, and how is such information used downstream? In this work, we show that DLMs do in fact encode a latent representation related to the diffusion timestep within their residual streams. We find that this signal can be reliably extracted using probes across layers, indicating that denoising progress is decodable from internal activations. We further demonstrate that steering the model along a low-dimensional subspace associated with the inferred timestep allows us to systematically modulate its notion of denoising progress, leading to predictable changes in model confidence and entropy. Finally, we analyse the geometry of the identified representation, showing that it exhibits structured and interpretable properties in activation space, and shedding light on how such a signal is processed by these models.

Large language model (LLM) decoding involves generating a sequence of tokens based on a given context, where each token is predicted one at a time using the model's learned probabilities. The typical autoregressive decoding method requires a separate forward pass through the model for each token generated, which is computationally inefficient and poses challenges for deploying LLMs in latency-sensitive scenarios. The main limitations of current decoding methods stem from their inefficiencies and resource demands. Existing approaches either necessitate fine-tuning smaller models, which is resource-intensive, or rely on fixed retrieval schemes to construct drafts for the next tokens, which lack adaptability and fail to generalize across different models and contexts. To address these issues, we introduce a novel methodology called ADED, which accelerates LLM decoding without requiring fine-tuning. Our approach involves an adaptive draft-verification process that evolves over time to improve efficiency. We utilize a tri-gram matrix-based LLM representation to dynamically approximate the output distribution of the LLM, allowing the model to adjust to changing token probabilities during the decoding process. Additionally, we implement a draft construction mechanism that effectively balances exploration and exploitation, ensuring that the drafts generated are both diverse and close to the true output distribution of the LLM. The importance of this design lies in its ability to optimize the draft distribution adaptively, leading to faster and more accurate decoding. Through extensive experiments on various benchmark datasets and LLM architectures, we demonstrate that ADED significantly accelerates the decoding process while maintaining high accuracy, making it suitable for deployment in a wide range of practical applications.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose ParallelBench, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We release our benchmark to help accelerate the development of truly efficient dLLMs.

One of the key components within diffusion models is the UNet for noise prediction. While several works have explored basic properties of the UNet decoder, its encoder largely remains unexplored. In this work, we conduct the first comprehensive study of the UNet encoder. We empirically analyze the encoder features and provide insights to important questions regarding their changes at the inference process. In particular, we find that encoder features change gently, whereas the decoder features exhibit substantial variations across different time-steps. This finding inspired us to omit the encoder at certain adjacent time-steps and reuse cyclically the encoder features in the previous time-steps for the decoder. Further based on this observation, we introduce a simple yet effective encoder propagation scheme to accelerate the diffusion sampling for a diverse set of tasks. By benefiting from our propagation scheme, we are able to perform in parallel the decoder at certain adjacent time-steps. Additionally, we introduce a prior noise injection method to improve the texture details in the generated image. Besides the standard text-to-image task, we also validate our approach on other tasks: text-to-video, personalized generation and reference-guided generation. Without utilizing any knowledge distillation technique, our approach accelerates both the Stable Diffusion (SD) and the DeepFloyd-IF models sampling by 41% and 24% respectively, while maintaining high-quality generation performance. Our code is available in https://github.com/hutaiHang/Faster-Diffusion{FasterDiffusion}.

Spatio-temporal video grounding (STVG) requires models to identify when a referred event occurs and localize the target entity throughout that interval. Existing multimodal large language models typically serialize dense localization trajectories autoregressively, causing decoding latency to grow with tube length and allowing localization errors to propagate across time. We introduce Parallel Tube Decoding (PTD), a generative formulation that decomposes grounding into a temporal block followed by time-conditioned spatial blocks decoded simultaneously. This removes both token-level and trajectory-level dependencies, reducing the sequential decoding depth to a fixed 1 + 1 rounds, independent of tube length. To enable parallel spatial generation, we introduce Decoupled Block Attention, which preserves access to shared video-query context while eliminating cross-box dependencies, together with localization-aware policy optimization for temporal boundaries and spatial geometry. On VidSTG, PTD reduces Tube Completion Latency by 79x and increases spatial decoding throughput by 92x over standard autoregressive decoding, while also improving grounding accuracy. With a compact 4B backbone, our model performs favorably well on VidSTG and HC-STVG, and generalizes zero-shot to temporal grounding, grounded VideoQA, and referring video object tracking. Our results show parallel tube generation is an efficient and effective alternative to autoregressive localization in videos.

In language modeling based music generation, a generated waveform is represented by a sequence of hierarchical token stacks that can be decoded either in an auto-regressive manner or in parallel, depending on the codebook patterns. In particular, flattening the codebooks represents the highest quality decoding strategy, while being notoriously slow. To this end, we propose a novel stack-and-delay style of decoding strategy to improve upon the flat pattern decoding where generation speed is four times faster as opposed to vanilla flat decoding. This brings the inference time close to that of the delay decoding strategy, and allows for faster inference on GPU for small batch sizes. For the same inference efficiency budget as the delay pattern, we show that the proposed approach performs better in objective evaluations, almost closing the gap with the flat pattern in terms of quality. The results are corroborated by subjective evaluations which show that samples generated by the new model are slightly more often preferred to samples generated by the competing model given the same text prompts.

To mitigate the high inference latency stemming from autoregressive decoding in Large Language Models (LLMs), Speculative Decoding has emerged as a novel decoding paradigm for LLM inference. In each decoding step, this method first efficiently drafts several future tokens and then verifies them in parallel. Unlike autoregressive decoding, Speculative Decoding facilitates the simultaneous decoding of multiple tokens per step, thereby accelerating inference. This paper presents a comprehensive overview and analysis of this promising decoding paradigm. We begin by providing a formal definition and formulation of Speculative Decoding. Then, we organize in-depth discussions on its key facets, including current leading techniques, the challenges faced, and potential future directions in this field. We aim for this work to serve as a catalyst for further research on Speculative Decoding, ultimately contributing to more efficient LLM inference.

Low-latency decoding for large language models (LLMs) is crucial for applications like chatbots and code assistants, yet generating long outputs remains slow in single-query settings. Prior work on speculative decoding (which combines a small draft model with a larger target model) and tensor parallelism has each accelerated decoding. However, conventional approaches fail to apply both simultaneously due to imbalanced compute requirements (between draft and target models), KV-cache inconsistencies, and communication overheads under small-batch tensor-parallelism. This paper introduces SwiftSpec, a system that targets ultra-low latency for LLM decoding. SwiftSpec redesigns the speculative decoding pipeline in an asynchronous and disaggregated manner, so that each component can be scaled flexibly and remove draft overhead from the critical path. To realize this design, SwiftSpec proposes parallel tree generation, tree-aware KV cache management, and fused, latency-optimized kernels to overcome the challenges listed above. Across 5 model families and 6 datasets, SwiftSpec achieves an average of 1.75x speedup over state-of-the-art speculative decoding systems and, as a highlight, serves Llama3-70B at 348 tokens/s on 8 Nvidia Hopper GPUs, making it the fastest known system for low-latency LLM serving at this scale.

Real-world images often suffer from spatially diverse degradations such as haze, rain, snow, and low-light, significantly impacting visual quality and downstream vision tasks. Existing all-in-one restoration (AIR) approaches either depend on external text prompts or embed hand-crafted architectural priors (e.g., frequency heuristics); both impose discrete, brittle assumptions that weaken generalization to unseen or mixed degradations. To address this limitation, we propose to reframe AIR as learned latent prior inference, where degradation-aware representations are automatically inferred from the input without explicit task cues. Based on latent priors, we formulate AIR as a structured reasoning paradigm: (1) which features to route (adaptive feature selection), (2) where to restore (spatial localization), and (3) what to restore (degradation semantics). We design a lightweight decoding module that efficiently leverages these latent encoded cues for spatially-adaptive restoration. Extensive experiments across six common degradation tasks, five compound settings, and previously unseen degradations demonstrate that our method outperforms state-of-the-art (SOTA) approaches, achieving an average PSNR improvement of 1.68 dB while being three times more efficient.

Learning latent representations from long text sequences is an important first step in many natural language processing applications. Recurrent Neural Networks (RNNs) have become a cornerstone for this challenging task. However, the quality of sentences during RNN-based decoding (reconstruction) decreases with the length of the text. We propose a sequence-to-sequence, purely convolutional and deconvolutional autoencoding framework that is free of the above issue, while also being computationally efficient. The proposed method is simple, easy to implement and can be leveraged as a building block for many applications. We show empirically that compared to RNNs, our framework is better at reconstructing and correcting long paragraphs. Quantitative evaluation on semi-supervised text classification and summarization tasks demonstrate the potential for better utilization of long unlabeled text data.

Tool-augmented speech assistants typically serialize automatic speech recognition, large language model inference, and external tool execution. As a result, tool latency is incurred only after the user has finished speaking and the LLM has identified the required tool calls. We present speculative tool execution for on-device cascaded voice agents, which predicts tool requests from partial ASR hypotheses and initiates tool execution while speech is still being received, thereby reducing end-to-end response latency. Our approach introduces a Predictor module that anticipates tool calls during speech recognition, executes them speculatively, and caches the results. The cached outputs are then injected into the LLM prompt, enabling faster responses. Additionally, to mitigate errors caused by user self-corrections during speech, we employ a rule-based validation mechanism that selectively injects only valid cached results. As a final safeguard, the LLM retains the ability to issue tool calls directly, ensuring that the latency of our framework is upper-bounded by the baseline serial execution pipeline in the worst case. We evaluate our method using live measurements from a fully implemented Android voice assistant. Our approach reduces the median time-to-first-audio from 5.79,s to 4.60,s and decreases the standard deviation from 3.49,s to 2.81,s, resulting in more predictable response latency.

We identify a previously overlooked failure mode of ALiBi positional encoding: its linear bias scaling underflows floating-point precision, which zeroes out a large fraction of attention weights and renders the affected attention heads partially blind. We analyze this failure mode, characterize its impact, and examine four mitigation strategies. We further demonstrate its occurrence in state-of-the-art pretrained models based on ALiBi. Comprehensive pretraining experiments with 148M-parameter decoder models help us to disentangle its effects from out-of-context degradation. We find that ALiBi's failure mode can substantially impair token retrieval while having only a minor effect on standard decoder benchmarks. We propose four training-time mitigation strategies and evaluate them individually and in combinations, finding that log-scaled distances yield the most consistent improvements in passkey retrieval. Despite this problem, default ALiBi slopes remain a surprisingly strong baseline, particularly for needle-in-a-haystack retrieval. Based on these findings we provide concrete recommendations on how to train models with ALiBi.

Real-time speech synthesis requires balancing inference latency and acoustic fidelity for interactive applications. Conventional continuous text-to-speech pipelines require computationally intensive neural vocoders to reconstruct phase information, creating a significant streaming bottleneck. Furthermore, regression-based acoustic modeling frequently induces spectral over-smoothing artifacts. To address these limitations, this paper proposes a novel end-to-end non-autoregressive architecture optimized for ultra-low latency block-wise generation, directly modeling the highly compressed discrete latent space of the Mimi neural audio codec. Integrating a modified FastSpeech 2 backbone with a progressive depth-wise sequential decoding strategy, the architecture dynamically conditions 32 layers of residual vector quantization codes. This mechanism resolves phonetic alignment degradation and manages the complexity of high-fidelity discrete representations without temporal autoregressive overhead. Experimental evaluations on English and Malay datasets validate its language-independent deployment capability. Compared to conventional continuous regression models, the proposed architecture demonstrates quantitative improvements in fundamental voicing accuracy and mitigates high-frequency spectral degradation. It achieves ultra-low latency inference, translating to a 10.6-fold absolute acceleration over conventional cascaded pipelines. Crucially, the system achieves an average time-to-first-byte latency of 48.99 milliseconds, falling significantly below the human perception threshold for real-time interactive streaming. These results firmly establish the proposed architecture as a highly optimized solution for deploying real-time streaming speech interfaces.

Neural audio codecs provide promising acoustic features for speech synthesis, with representative streaming codecs like Mimi providing high-quality acoustic features for real-time Text-to-Speech (TTS) applications. However, Mimi's decoder, which employs a hybrid transformer and convolution architecture, introduces significant latency bottlenecks on edge devices due to the the compute intensive nature of deconvolution layers which are not friendly for mobile-CPUs, such as the most representative framework XNNPACK. This paper introduces T-Mimi, a novel modification of the Mimi codec decoder that replaces its convolutional components with a purely transformer-based decoder, inspired by the TS3-Codec architecture. This change dramatically reduces on-device TTS latency from 42.1ms to just 4.4ms. Furthermore, we conduct quantization aware training and derive a crucial finding: the final two transformer layers and the concluding linear layers of the decoder, which are close to the waveform, are highly sensitive to quantization and must be preserved at full precision to maintain audio quality.

Long-context LLM decoding reads the key-value (KV) cache at every step. Loading it takes longer than computing attention over it, so throughput is bandwidth-bound. Hence, reducing the cache size can raise both decoding speed and serving capacity. The challenge is to reduce cache size while preserving the attention products, keeping reconstruction cheap, and using a fixed per-token bit count. At two bits per element, the most competitive methods rely on orthogonal transforms. However, existing techniques are either data-oblivious or use the query statistics without deriving the transform from a distortion criterion. Moreover, they rely on transforms built on top of random or Hadamard rotations, which equalize variances across entries rather than compacting energy, and fixed-width scalar quantizers, which are suboptimal at low rates. In this paper, we formulate KV cache quantization as a transform coding problem in which distortion is the error in the attention products. We derive closed-form optimal transforms for keys and values from calibration statistics, under a high-resolution model. We show that the optimal key transform is not orthogonal and satisfies a generalized Parseval relation: the attention-aware distortion becomes mean-squared error (MSE) in the transform domain. Thus, we can use MSE-optimal vector quantizers applied directly to the transformed key coefficients. To meet the fixed-width layout requirement, we show that grouping coefficients into equal-volume partitions makes equal-size codebooks attain the variable-rate optimum under the same high-resolution model. At two bits per element, our method, termed NOVA-KV, recovers most of the long-context retrieval accuracy lost by scalar quantization methods at comparable throughput.

Latent Diffusion Models (LDMs) typically adopt a two-stage pipeline: an auto-encoder (AE) is first pre-trained to define a latent space, then a diffusion model is trained to perform denoising within it. Such a two-stage design introduces a representation mismatch, as the latent space is optimized for reconstruction rather than adapting the denoising dynamics. We reveal that the LDM itself is an AE, and consequently present LDM-is-AE, an end-to-end one-stage LDM training framework that eliminates the need for a separately trained tokenizer. Our key observation is that the LDM backbone actually performs a latent-to-feature-to-latent transformation at each denoising step, which can be interpreted as an internal decoding--encoding process. Leveraging this structure, we split the DiT backbone into two reciprocal components, DiT-E (i.e., DiT Encoding) and DiT-D (i.e., DiT Decoding), and impose image-space supervision on the intermediate features across all timesteps. Our model encourages the internal representation to align with the image domain throughout denoising, thereby establishing an explicit latent-to-image-to-latent path. At the zero-noise timestep, our model further performs an image-to-latent-to-image mapping, corresponding to an auto-encoding process. As a result, LDM-is-AE jointly learns latent representations and denoising dynamics in an end-to-end manner, yielding a diffusion-native latent space tailored to the generation process. Experiments demonstrate that LDM-is-AE exhibits highly competitive generation performance, achieving an FID of 1.80 and 1.90 on 256x256 and 512x512 class-conditional image generation, respectively.

The real-time processing of time series signals is a critical issue for many real-life applications. The idea of real-time processing is especially important in audio domain as the human perception of sound is sensitive to any kind of disturbance in perceived signals, especially the lag between auditory and visual modalities. The rise of deep learning (DL) models complicated the landscape of signal processing. Although they often have superior quality compared to standard DSP methods, this advantage is diminished by higher latency. In this work we propose novel method for minimization of inference time latency and memory consumption, called Short-Term Memory Convolution (STMC) and its transposed counterpart. The main advantage of STMC is the low latency comparable to long short-term memory (LSTM) networks. Furthermore, the training of STMC-based models is faster and more stable as the method is based solely on convolutional neural networks (CNNs). In this study we demonstrate an application of this solution to a U-Net model for a speech separation task and GhostNet model in acoustic scene classification (ASC) task. In case of speech separation we achieved a 5-fold reduction in inference time and a 2-fold reduction in latency without affecting the output quality. The inference time for ASC task was up to 4 times faster while preserving the original accuracy.

Multimodal Large Language Models (MLLMs) have achieved strong performance across many tasks, yet most systems remain limited to offline inference, requiring complete inputs before generating outputs. Recent streaming methods reduce latency by interleaving perception and generation, but still enforce a sequential perception-generation cycle, limiting real-time interaction. In this work, we target a fundamental bottleneck that arises when extending MLLMs to real-time video understanding: the global positional continuity constraint imposed by standard positional encoding schemes. While natural in offline inference, this constraint tightly couples perception and generation, preventing effective input-output parallelism. To address this limitation, we propose a parallel streaming framework that relaxes positional continuity through three designs: Overlapped, Group-Decoupled, and Gap-Isolated. These designs enable simultaneous perception and generation, allowing the model to process incoming inputs while producing responses in real time. Extensive experiments reveal that Group-Decoupled achieves the best efficiency-performance balance, maintaining high fluency and accuracy while significantly reducing latency. We further show that the proposed framework yields up to 2x acceleration under balanced perception-generation workloads, establishing a principled pathway toward speak-while-watching real-time systems. We make all our code publicly available: https://github.com/EIT-NLP/Speak-While-Watching.

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to autoregressive generation by enabling parallel token prediction. However, practical dLLM decoding still suffers from high inference latency, which limits deployment. In this work, we observe that a substantial part of this inefficiency comes from recurring redundancy in the decoding process, including spatial redundancy caused by confidence clusters and positional ambiguity, and temporal redundancy caused by repeatedly remasking predictions that have already stabilized. Motivated by these patterns, we propose R^{2}-dLLM, a unified framework for reducing decoding redundancy from both inference and training perspectives. At inference time, we introduce training-free decoding rules that aggregate local confidence and token predictions, and finalize temporally stable tokens to avoid redundant decoding steps. We further propose a redundancy-aware supervised fine-tuning pipeline that aligns the model with efficient decoding trajectories and reduces reliance on manually tuned thresholds. Experiments demonstrate that R^{2}-dLLM consistently reduces the number of decoding steps by up to 88\% compared to existing decoding strategies, while maintaining competitive generation quality across different models and tasks. These results validate that decoding redundancy is a central bottleneck in dLLMs, and that explicitly reducing it yields substantial practical efficiency gains. Our code and models are available at https://github.com/GATECH-EIC/R2-dLLM.

Depth sensing is a critical function for robotic tasks such as localization, mapping and obstacle detection. There has been a significant and growing interest in depth estimation from a single RGB image, due to the relatively low cost and size of monocular cameras. However, state-of-the-art single-view depth estimation algorithms are based on fairly complex deep neural networks that are too slow for real-time inference on an embedded platform, for instance, mounted on a micro aerial vehicle. In this paper, we address the problem of fast depth estimation on embedded systems. We propose an efficient and lightweight encoder-decoder network architecture and apply network pruning to further reduce computational complexity and latency. In particular, we focus on the design of a low-latency decoder. Our methodology demonstrates that it is possible to achieve similar accuracy as prior work on depth estimation, but at inference speeds that are an order of magnitude faster. Our proposed network, FastDepth, runs at 178 fps on an NVIDIA Jetson TX2 GPU and at 27 fps when using only the TX2 CPU, with active power consumption under 10 W. FastDepth achieves close to state-of-the-art accuracy on the NYU Depth v2 dataset. To the best of the authors' knowledge, this paper demonstrates real-time monocular depth estimation using a deep neural network with the lowest latency and highest throughput on an embedded platform that can be carried by a micro aerial vehicle.

Large Language Models (LLMs), such as OpenAI-o1 and DeepSeek-R1, have demonstrated strong reasoning capabilities. To further enhance LLM capabilities, recent agentic systems, such as Deep Research, incorporate web interactions into LLM reasoning to mitigate uncertainties and reduce potential errors. However, existing research predominantly focuses on reasoning performance, often neglecting the efficiency of agentic systems. In this work, we present a comprehensive empirical study that identifies efficiency bottlenecks in web-interactive agentic systems. We decompose end-to-end latency into two primary components: LLM API latency and web environment latency. We conduct a comprehensive empirical study across 15 models and 5 providers to demonstrate high variability in API-based agentic systems. We observe that web environment latency can contribute as much as 53.7% to the overall latency in a web-based agentic system. To improve latency, we propose SpecCache, a caching framework augmented with speculative execution that can reduce web environment overhead. Extensive evaluations on two standard benchmarks show that our approach improves the cache hit rate by up to 58x compared to a random caching strategy, while reducing web environment overhead by up to 3.2x, without degrading agentic system performance.

Diffusion Language Models (DLMs) have been seen as a promising competitor for autoregressive language models. However, diffusion language models have long been constrained by slow inference. A core challenge is that their non-autoregressive architecture and bidirectional attention preclude the key-value cache that accelerates decoding. We address this bottleneck by proposing a KV-cache-like mechanism, delayed KV-Cache, for the denoising process of DLMs. Our approach is motivated by the observation that different tokens have distinct representation dynamics throughout the diffusion process. Accordingly, we propose a delayed and conditioned caching strategy for key and value states. We design two complementary variants to cache key and value step-by-step: (1) dKV-Cache-Decode, which provides almost lossless acceleration, and even improves performance on long sequences, suggesting that existing DLMs may under-utilise contextual information during inference. (2) dKV-Cache-Greedy, which has aggressive caching with reduced lifespan, achieving higher speed-ups with quadratic time complexity at the cost of some performance degradation. dKV-Cache, in final, achieves from 2-10x speedup in inference, largely narrowing the gap between ARs and DLMs. We evaluate our dKV-Cache on several benchmarks, delivering acceleration across general language understanding, mathematical, and code-generation benchmarks. Experiments demonstrate that cache can also be used in DLMs, even in a training-free manner from current DLMs.

Linear attention has recently gained significant attention for long-context inference due to its constant decoding cost with respect to context length. However, existing serving systems typically serve linear attention by recurrently computing and updating a large linear attention state in every decoding step. Since the state is much larger than the per-token key and value, recurrent decoding incurs substantial memory access and becomes inefficient for serving linear attention. In this paper, we propose KVBuffer, an IO-aware serving mechanism for linear attention. By buffering recent keys and values, KVBuffer enables serving systems to compute linear attention outputs in more flexible and memory-efficient ways. For decoding, KVBuffer enables chunkwise computation, which reduces average memory access and decoding latency by deferring state updates and applying them in batch. For speculative decoding, KVBuffer verifies draft tokens in parallel and avoids storing temporary states. For short contexts, KVBuffer computes attention outputs directly from buffered keys and values, without creating or updating the linear attention state. We implement KVBuffer in SGLang for Qwen3-Next. Our evaluations show that KVBuffer can reduce linear attention decoding latency by up to 45.17% and increase the maximum number of serving requests by 5x for speculative decoding when verifying four draft tokens.

Request scheduling is a key challenge in large-scale clusters serving agentic large language model (LLM) workloads. An effective scheduler must preserve key-value cache (KVC) reuse across long, shared prefixes, meet token-level latency service-level objectives (SLOs), and minimize GPU resource footprint. Existing schedulers struggle to reconcile these requirements: request consolidation can sacrifice cache locality and increase prefill/decode interference, compromising both SLO attainment and resource efficiency. We present PackServe, a scheduler designed to reduce resource costs while meeting latency SLOs for agentic LLM serving. PackServe uses compact white-box models to predict latency under prefill/decode interference. Guided by these predictions, it packs requests onto fewer serving instances while preserving KVC reuse and SLO constraints, trading available latency headroom for improved per-instance throughput. Evaluation on 64 NVIDIA H20 GPUs shows that PackServe uses up to 16.8% and 24.6% fewer GPU-hours than state-of-the-art schedulers under 30-ms and 50-ms TPOT targets, respectively, while meeting the target TPOT objectives. PackServe has also been deployed in our production cluster comprising over 1000 GPUs, where it reduces the resource footprint by 34.7% compared with the original production scheduler.

During language model decoding, it is known that using higher temperature sampling gives more creative responses, while lower temperatures are more factually accurate. However, such models are commonly applied to general instruction following, which involves both creative and fact seeking tasks, using a single fixed temperature across all examples and tokens. In this work, we introduce Adaptive Decoding, a layer added to the model to select the sampling temperature dynamically at inference time, at either the token or example level, in order to optimize performance. To learn its parameters we introduce Latent Preference Optimization (LPO) a general approach to train discrete latent variables such as choices of temperature. Our method outperforms all fixed decoding temperatures across a range of tasks that require different temperatures, including UltraFeedback, Creative Story Writing, and GSM8K.

Two widely adopted techniques for LLM inference serving systems today are hybrid batching and disaggregated serving. A hybrid batch combines prefill and decode tokens of different requests in the same batch to improve resource utilization and throughput at the cost of increased latency per token. In contrast, disaggregated serving decouples compute-bound prefill and bandwidth-bound decode phases to optimize for service level objectives (SLOs) at the cost of resource under-utilization and KV-cache transfer overheads. To address the limitations of these techniques, we propose RAPID-Serve: a technique to concurrently execute prefill and decode on the same GPU(s) to meet latency SLOs while maintaining high throughput and efficient resource utilization. Furthermore, we propose Adaptive Resource Management for runtime compute resource allocation, optionally leveraging CU masking (a fine-grained Compute Unit partitioning feature on AMD InstinctTM GPUs). RAPID-Serve provides up to 4.1x (average 1.7x) unconstrained throughput improvement and 32x and higher (average 4.9x) throughput improvement under SLO constraints, showing it as an effective strategy compared to the state-of-the-art approaches, particularly in resource-constrained environments.

Video Large Language Models (VLLMs) unlock world-knowledge-aware video understanding through pretraining on internet-scale data and have already shown promise on tasks such as movie analysis and video question answering. However, deploying VLLMs for downstream tasks such as video recommendation remains challenging, since real systems require multi-video inputs, lightweight backbones, low-latency sequential inference, and rapid response. In practice, (1) decode-only generation yields high latency for sequential inference, (2) typical interfaces do not support multi-video inputs, and (3) constraining outputs to language discards fine-grained visual details that matter for downstream vision tasks. We argue that these limitations stem from the absence of a representation that preserves pixel-level detail while leveraging world knowledge. We present LinkedOut, a representation that extracts VLLM world knowledge directly from video to enable fast inference, supports multi-video histories, and removes the language bottleneck. LinkedOut extracts semantically grounded, knowledge-aware tokens from raw frames using VLLMs, guided by promptable queries and optional auxiliary modalities. We introduce a cross-layer knowledge fusion MoE that selects the appropriate level of abstraction from the rich VLLM features, enabling personalized, interpretable, and low-latency recommendation. To our knowledge, LinkedOut is the first VLLM-based video recommendation method that operates on raw frames without handcrafted labels, achieving state-of-the-art results on standard benchmarks. Interpretability studies and ablations confirm the benefits of layer diversity and layer-wise fusion, pointing to a practical path that fully leverages VLLM world-knowledge priors and visual reasoning for downstream vision tasks such as recommendation.

At each step, decoding one sequence with a large language model rereads the projection weights, whose traffic is fixed, and the key-value (KV) cache, whose traffic grows with context. Activation sparsity trims the first term and KV-cache sparsity the second, yet their reported speedups are hard to compare because each depends on context length and on the dense attention kernel it is measured against. We derive a byte crossover, the context length at which the two savings are equal, together with ideal speedup bounds for each branch and for their composition, from model dimensions and keep ratios alone. We then time both branches and their composition from 2K to 128K tokens on two GPUs after a dense prefill of real text, with dense and sparse modes reading the cache through the same split-K attention kernel. The projection branch leads at short context and the KV branch at long context, with speedups that follow their byte bounds up to fixed kernel costs. Adding these costs, measured in separate sweeps, lets the byte account predict the measured crossings of three keep-ratio pairs, a second model, and a second GPU to within 4.1K tokens. Timing the dense baseline with masked instead of split-K attention inflates the apparent speedup of the same KV policy about fivefold. An attention-scored KV selection answers the same passkey and multi-key placements as dense decoding up to 127K tokens, whereas a KV window misses most of them. Under matched perplexity budgets, activation sparsity composed with this selection decodes 14 to 26% faster than the best single branch on both GPUs. Code is available at https://github.com/js-lee-AI/ByteCross.

Latent diffusion models (LDMs) power state-of-the-art high-resolution generative image models. LDMs learn the data distribution in the latent space of an autoencoder (AE) and produce images by mapping the generated latents into RGB image space using the AE decoder. While this approach allows for efficient model training and sampling, it induces a disconnect between the training of the diffusion model and the decoder, resulting in a loss of detail in the generated images. To remediate this disconnect, we propose to leverage the internal features of the decoder to define a latent perceptual loss (LPL). This loss encourages the models to create sharper and more realistic images. Our loss can be seamlessly integrated with common autoencoders used in latent diffusion models, and can be applied to different generative modeling paradigms such as DDPM with epsilon and velocity prediction, as well as flow matching. Extensive experiments with models trained on three datasets at 256 and 512 resolution show improved quantitative -- with boosts between 6% and 20% in FID -- and qualitative results when using our perceptual loss.

Disaggregated inference architectures physically separate prefill and decode phases onto distinct GPU pools, creating competing "agents" that share a fixed hardware budget. We provide, to our knowledge, the first formal game-theoretic analysis of this architecture, using NVIDIA Dynamo as a concrete case study. We model disaggregated serving as three coupled games: a two-player resource game between prefill and decode pools, a selfish caching game over the hierarchical KV cache, and a congestion game with positive externalities for request routing. We empirically validate the latter two; the P/D resource game is treated analytically (Section 9.2). We characterize how GPU saturation induces regime transitions that shift the game's payoff structure: below saturation, selfish behavior has bounded Price of Anarchy (PoA); at saturation, superlinear latency and cache externalities drive our empirical estimator PoA-hat (defined in Section 6.4) upward. Based on this analysis, we design an adaptive controller that detects saturation transitions in real time and adjusts routing parameters accordingly, shifting from cache-affinity exploitation to load-balanced congestion avoidance. We instantiate our framework on a 3-node NVIDIA B200 cluster running Dynamo with two models, Nemotron-4-340B (TP=8, full-node workers with cross-InfiniBand KV transfers) and Llama-3.1-70B (TP=4), and find the same three-regime PoA-hat structure with the same first post-knee grid point (C=128) on both models. Adaptive routing shifts each model to a better operating point. Our strongest result is on the 70B 1P/5D topology, where PoA-hat drops 3.1x (66.4 to 21.5) in the saturated phase at a 13% throughput cost. On the 70B 1P/2D, PoA-hat drops 2.2x and TTFT P99 drops 7.6x (see Section 8.5).

Connectionist Temporal Classification (CTC) suffers from the latency problem when applied to streaming models. We argue that in CTC lattice, the alignments that can access more future context are preferred during training, thereby leading to higher symbol delay. In this work we propose the delay-penalized CTC which is augmented with latency penalty regularization. We devise a flexible and efficient implementation based on the differentiable Finite State Transducer (FST). Specifically, by attaching a binary attribute to CTC topology, we can locate the frames that firstly emit non-blank tokens on the resulting CTC lattice, and add the frame offsets to the log-probabilities. Experimental results demonstrate the effectiveness of our proposed delay-penalized CTC, which is able to balance the delay-accuracy trade-off. Furthermore, combining the delay-penalized transducer enables the CTC model to achieve better performance and lower latency. Our work is open-sourced and publicly available https://github.com/k2-fsa/k2.

Speculative decoding is a technique that accelerates the inference speed of autoregressive large language models (LLMs) by sampling from a faster, lower quality draft model. Traditionally, a fundamental characteristic of speculative decoding is that it is lossless with respect to the verifier distribution, guaranteeing speed without sacrificing quality. This differs from recently developed diffusion LLMs (dLLMs), which are notably able to achieve substantial speedups by degrading the quality of their outputs with less diffusion effort. Because diffusion models are able to sample faster than autoregressive models by generating tokens in parallel, they are considered strong candidates for draft models in speculative diffusion. While it is widely understood that diffusion models can sacrifice quality for speed, speculative decoding itself can also be generalized to generate lossy samples faster using more lenient acceptance criteria with respect to the verifier distribution. Presented with the two parameters of diffusion effort and speculative decoding leniency, this paper explores the full achievable Pareto frontier via joint variations of these parameters. Our experiments show that jointly varying diffusion effort and acceptance leniency can extend the speed-quality Pareto frontier over single step diffusion drafting.

Large Language Models (LLMs) have demonstrated remarkable capabilities in leveraging extensive external knowledge to enhance responses in multi-turn and agentic applications, such as retrieval-augmented generation (RAG). However, processing long-context inputs introduces significant system latency and demands substantial memory for the key-value cache, resulting in reduced throughput and a fundamental trade-off between knowledge enrichment and system efficiency. While minimizing latency for long-context inputs is a primary objective for LLMs, we contend that RAG require specialized consideration. In RAG, much of the LLM context consists of concatenated passages from retrieval, with only a small subset directly relevant to the query. These passages often exhibit low semantic similarity due to diversity or deduplication during re-ranking, leading to block-diagonal attention patterns that differ from those in standard LLM generation tasks. Based on this observation, we argue that most computations over the RAG context during decoding are unnecessary and can be eliminated with minimal impact on performance. To this end, we propose REFRAG, an efficient decoding framework that compresses, senses, and expands to improve latency in RAG applications. By exploiting the sparsity structure, we demonstrate a 30.85 the time-to-first-token acceleration (3.75 improvement to previous work) without loss in perplexity. In addition, our optimization framework for large context enables REFRAG to extend the context size of LLMs by 16. We provide rigorous validation of REFRAG across diverse long-context tasks, including RAG, multi-turn conversations, and long document summarization, spanning a wide range of datasets. Experimental results confirm that REFRAG delivers substantial speedup with no loss in accuracy compared to LLaMA models and other state-of-the-art baselines across various context sizes.

We present LFM2, a family of Liquid Foundation Models designed for efficient on-device deployment and strong task capabilities. Using hardware-in-the-loop architecture search under edge latency and memory constraints, we obtain a compact hybrid backbone that combines gated short convolutions with a small number of grouped query attention blocks, delivering up to 2x faster prefill and decode on CPUs compared to similarly sized models. The LFM2 family covers 350M-8.3B parameters, including dense models (350M, 700M, 1.2B, 2.6B) and a mixture-of-experts variant (8.3B total, 1.5B active), all with 32K context length. LFM2's training pipeline includes a tempered, decoupled Top-K knowledge distillation objective that avoids support mismatch; curriculum learning with difficulty-ordered data; and a three-stage post-training recipe of supervised fine-tuning, length-normalized preference optimization, and model merging. Pre-trained on 10-12T tokens, LFM2 models achieve strong results across diverse benchmarks; for example, LFM2-2.6B reaches 79.56% on IFEval and 82.41% on GSM8K. We further build multimodal and retrieval variants: LFM2-VL for vision-language tasks, LFM2-Audio for speech, and LFM2-ColBERT for retrieval. LFM2-VL supports tunable accuracy-latency tradeoffs via token-efficient visual processing, while LFM2-Audio separates audio input and output pathways to enable real-time speech-to-speech interaction competitive with models 3x larger. LFM2-ColBERT provides a low-latency encoder for queries and documents, enabling high-performance retrieval across multiple languages. All models are released with open weights and deployment packages for ExecuTorch, llama.cpp, and vLLM, making LFM2 a practical base for edge applications that need fast, memory-efficient inference and strong task capabilities.

The rapid scaling of Protein Language Models (PLMs) has unlocked unprecedented accuracy in protein structure prediction and design, but the quadratic memory growth of the Key-Value (KV) cache during inference remains a prohibitive barrier for single-GPU deployment and high-throughput generation. While 8-bit quantization is now standard, 3-bit quantization remains elusive due to severe numerical outliers in activations. This paper presents TurboESM, an adaptation of Google's TurboQuant to the PLM domain. We solve the fundamental incompatibility between Rotary Position Embeddings (RoPE) and orthogonal transformations by deriving a RoPE-first rotation pipeline. We introduce a head-wise SVD calibration method tailored to the amino acid activation manifold, a dual look-up table (LUT) strategy for asymmetric K/V distributions, and a 1-bit Quantized Johnson-Lindenstrauss (QJL) residual correction. All experiments are conducted on ESM-2 650M, where our implementation achieves a 7.1x memory reduction (330 MB to 47 MB) while maintaining cosine similarity > 0.96 in autoregressive decoding across diverse protein families, including short peptides, transmembrane helices, enzyme active site fragments, and intrinsically disordered regions. We further implement a Triton-based fused decode attention kernel that eliminates intermediate dequantization memory allocations, achieving a 1.96x speedup over the PyTorch two-step path for the KV fetch operation alone; however, TurboESM incurs a prefill overhead of 21-27 ms relative to the original model due to KV quantization and packing, making it most suitable for memory-bound scenarios rather than latency-critical short-sequence workloads. Analysis reveals that PLMs exhibit sharper outlier profiles than large language models (LLMs) due to amino acid vocabulary sparsity, and our method effectively addresses these distributions.

Prefill-decode (PD) disaggregation has become the standard architecture for large-scale LLM serving, but in practice its deployment boundary is still determined by KVCache transfer. In conventional dense-attention models, prefill generates huge KVCache traffics that keep prefill and decode tightly coupled within a single high-bandwidth network domain, limiting heterogeneous deployment and resource elasticity. Recent hybrid-attention architectures substantially reduce KVCache size, making cross-cluster KVCache transport increasingly plausible. However, smaller KVCache alone does not make heterogeneous cross-datacenter PD serving practical: real workloads remain bursty, request lengths are highly skewed, prefix caches are unevenly distributed, and inter-cluster bandwidth fluctuates. A naive design that fully externalizes prefill can therefore still suffer from congestion, unstable queueing, and poor utilization. We present Prefill-as-a-Service (PrfaaS), a cross-datacenter serving architecture that selectively offloads long-context prefill to standalone, compute-dense prefill clusters and transfers the resulting KVCache over commodity Ethernet to local PD clusters for decode. Rather than treating reduced KVCache as sufficient, PrfaaS combines model-side KV efficiency with system-side selective offloading, bandwidth-aware scheduling, and cache-aware request placement. This design removes the requirement that heterogeneous accelerators share the same low-latency RDMA fabric, enabling independent scaling of prefill and decode capacity across loosely coupled clusters. In a case study using an internal 1T-parameter hybrid model, a PrfaaS-augmented heterogeneous deployment achieves 54% higher serving throughput and 64% lower P90 TTFT than a homogeneous PD baseline, with approximately 15% throughput gain at equal cost, while consuming only modest cross-datacenter bandwidth.

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.

Simultaneous machine translation consists in starting output generation before the entire input sequence is available. Wait-k decoders offer a simple but efficient approach for this problem. They first read k source tokens, after which they alternate between producing a target token and reading another source token. We investigate the behavior of wait-k decoding in low resource settings for spoken corpora using IWSLT datasets. We improve training of these models using unidirectional encoders, and training across multiple values of k. Experiments with Transformer and 2D-convolutional architectures show that our wait-k models generalize well across a wide range of latency levels. We also show that the 2D-convolution architecture is competitive with Transformers for simultaneous translation of spoken language.

Simultaneous machine translation (SiMT) has traditionally relied on offline machine translation models coupled with human-engineered heuristics or learned policies. We propose Hikari, a policy-free, fully end-to-end model that performs simultaneous speech-to-text translation and streaming transcription by encoding READ/WRITE decisions into a probabilistic WAIT token mechanism. We also introduce Decoder Time Dilation, a mechanism that reduces autoregressive overhead and ensures a balanced training distribution. Additionally, we present a supervised fine-tuning strategy that trains the model to recover from delays, significantly improving the quality-latency trade-off. Evaluated on English-to-Japanese, German, and Russian, Hikari achieves new state-of-the-art BLEU scores in both low- and high-latency regimes, outperforming recent baselines.

Although significant efforts have been dedicated to aligning large language models (LLMs), red-teaming reports suggest that these carefully aligned LLMs could still be jailbroken through adversarial prompts, tuning, or decoding. Upon examining the jailbreaking vulnerability of aligned LLMs, we observe that the decoding distributions of jailbroken and aligned models differ only in the initial generations. This observation motivates us to propose the weak-to-strong jailbreaking attack, where adversaries can utilize smaller unsafe/aligned LLMs (e.g., 7B) to guide jailbreaking against significantly larger aligned LLMs (e.g., 70B). To jailbreak, one only needs to additionally decode two smaller LLMs once, which involves minimal computation and latency compared to decoding the larger LLMs. The efficacy of this attack is demonstrated through experiments conducted on five models from three different organizations. Our study reveals a previously unnoticed yet efficient way of jailbreaking, exposing an urgent safety issue that needs to be considered when aligning LLMs. As an initial attempt, we propose a defense strategy to protect against such attacks, but creating more advanced defenses remains challenging. The code for replicating the method is available at https://github.com/XuandongZhao/weak-to-strong

The performance of multi-turn, agentic LLM inference is increasingly dominated by KV-Cache storage I/O rather than computation. In prevalent disaggregated architectures, loading the massive KV-Cache from external storage creates a fundamental imbalance: storage NICs on prefill engines become bandwidth-saturated, while those on decoding engines remain idle. This asymmetry severely constrains overall system throughput. We present DualPath, an inference system that breaks this bottleneck by introducing dual-path KV-Cache loading. Beyond the traditional storage-to-prefill path, DualPath enables a novel storage-to-decode path, in which the KV-Cache is loaded into decoding engines and then efficiently transferred to prefill engines via RDMA over the compute network. DualPath combines this optimized data path -- which inherently avoids network congestion and avoids interference with latency-critical model execution communications -- with a global scheduler that dynamically balances load across prefill and decode engines. Our evaluation on three models with production agentic workloads demonstrates that DualPath improves offline inference throughput by up to 1.87times on our in-house inference system. It can also improve online serving throughput by an average factor of 1.96times without violating SLO.

LAPS identifies and disaggregates requests with different prompt lengths in LLM serving to reduce TTFT latency. While recent systems have decoupled the prefill and decode stages to improve throughput, they still rely on unified scheduling policies that fail to adapt to heterogeneous workload characteristics. We observe that prompt-length variations lead to distinct performance bottlenecks, motivating an adaptive scheduling strategy. LAPS disaggregates multi-turn long-prefill requests from short-prefill ones and introduces a length-aware smart batching mechanism for short-prefill workloads. It adopts a dual-queue design that supports temporal disaggregation on a single prefill instance or spatial disaggregation across multiple instances. For short-prefill batches, a batch waiting window and CUDA Graph-based clustering mitigate interference from heterogeneous computation, reducing batching delay and lowering average latency. In real multi-turn workloads, LAPS reduces prefill latency by over 30\% compared to vanilla SGLang under prefill-decode disaggregation, and further decreases SLO violations by 28\% in multi-instance deployments with vanilla data-parallel configuration. Compared to the SGLang router with load balancing, it further lowers SLO violations by 12\% in multi-GPU settings. Under high concurrency and mixed-request scenarios, LAPS improves request throughput by 35\% serving Qwen2.5-32B model for prefill instance, demonstrating its effectiveness in optimizing heterogeneous LLM serving workloads.

Large Language Model (LLM) workloads have distinct prefill and decode phases with different compute and memory requirements which should ideally be accounted for when scheduling input queries across different LLM instances in a cluster. However existing scheduling algorithms treat LLM workloads as monolithic jobs without considering the distinct characteristics of the two phases in each workload. This leads to sub-optimal scheduling and increased response latency. In this work, we propose a heuristic-guided reinforcement learning-based intelligent router for data-driven and workload-aware scheduling. Our router leverages a trainable response-length predictor, and a novel formulation for estimating the impact of mixing different workloads to schedule queries across LLM instances and achieve over 11\% lower end-to-end latency than existing approaches.

As large language models (LLMs) continue to grow in size, distributed inference has become increasingly important. Model-parallel strategies must now efficiently scale not only across multiple GPUs but also across multiple nodes. In this work, we present a detailed performance study of multi-node distributed inference using LLMs on GPU-based supercomputers. We conduct experiments with several state-of-the-art inference engines alongside YALIS, a research-oriented prototype engine designed for controlled experimentation. We analyze the strong-scaling behavior of different model-parallel schemes and identify key bottlenecks. Since all-reduce operations are a common performance bottleneck, we develop NVRAR, a hierarchical all-reduce algorithm based on recursive doubling with NVSHMEM. NVRAR achieves up to 1.9x-3.6x lower latency than NCCL for message sizes between 128 KB and 2 MB on HPE Slingshot and InfiniBand interconnects. Integrated into YALIS, NVRAR achieves up to a 1.72x reduction in end-to-end batch latency for the Llama 3.1 405B model in multi-node decode-heavy workloads using tensor parallelism.

We present a single-GPU (H100, bf16) evaluation of GPT-OSS-20B (Mixture-of-Experts; 20.9B total, approx. 3.61B active) against dense baselines Qwen3-32B and Yi-34B across multiple dimensions. We measure true time-to-first-token (TTFT), full-decode throughput (TPOT), end-to-end latency percentiles, peak VRAM with past key values (PKV) held, and energy via a consistent nvidia-smi-based sampler. At a 2048-token context with 64-token decode, GPT-OSS-20B delivers higher decode throughput and tokens per Joule than dense baselines Qwen3-32B and Yi-34B, while substantially reducing peak VRAM and energy per 1000 generated tokens; its TTFT is higher due to MoE routing overhead. With only 17.3% of parameters active (3.61B of 20.9B), GPT-OSS-20B provides about 31.8% higher decode throughput and 25.8% lower energy per 1000 generated tokens than Qwen3-32B at 2048/64, while using 31.7% less peak VRAM. Normalized by active parameters, GPT-OSS-20B shows markedly stronger per-active-parameter efficiency (APE), underscoring MoE's deployment advantages. We do not evaluate accuracy; this is a deployment-focused study. We release code and consolidated results to enable replication and extension.

LLM agents today communicate via text, which incurs considerable latency and information loss due to the need to autoregressively decode the sharer model's state and encode at the receiver model. Recent work such as Cache-to-Cache (C2C; Fu et al., 2026) seeks to exchange KV caches by learning adapters that translate sharer KV matrices to the receiver model. However, the adapters are large and expensive to train, and translate individual tokens, which requires the target context to be identical. This is unsuitable for agent communication, where the LLMs have differing context. We introduce Latent Cache Flow (LCF). To address efficiency, we observe that keys and values can be jointly translated and compressed, reducing the adapter to about 4% of C2C's size. To address differing context, we design the adapter to transmit a summary of new information that the target model does not have. Our early experiments show that a pruned 13 MB LCF adapter can be more accurate than C2C at 956 MB in shared-context settings; for different contexts, LCF improves F1 by 7.5% and Exact Match by 23% while 8.5 times faster than text-based communication.

Visual dubbing must synchronize mouth motion with replacement speech while preserving identity, appearance, and temporal consistency. Although X-Dub provides a strong mask-free video-editing baseline, its application to livestream and generated-video content reveals limitations in production-domain robustness, temporal and motion stability, identity and oral-detail preservation, and inference efficiency. We present TBDub, a production-oriented extension of X-Dub that combines task-adaptive post-training with task-aware few-step distillation. Post-training adapts the video DiT using production-domain data, production-specific conditioning and filtering, and enhanced audio features to obtain a 30-step Teacher. Distillation adapts DMD/DMD2 to conditional video editing and compresses the Teacher into a two-step Student. On 38 TalkVid clips, the Teacher improves all eight reported reconstruction, perceptual, identity, and synchronization metrics over X-Dub. In the MOS evaluation, it improves lip-sync consistency, identity consistency, and visual quality over X-Dub by 0.14, 0.95, and 0.90 points, while the Student achieves the highest lip-sync and visual-quality scores and remains close to the Teacher in identity consistency. In paired end-to-end generation timing from the first VAE encode through the final VAE decode on a single NVIDIA H20 GPU at 512times512, the Student reaches 7.13 effective FPS and reduces total latency by 13.93times; the DiT stage alone is accelerated by 42.49times. The Student largely retains the Teacher's generation quality and audiovisual synchronization. The code is available on GitHub at \https://github.com/TaoLiveAIGC/TBDub, and the 30-step Teacher and two-step Student weights are available on Hugging Face at https://huggingface.co/TaoLiveAIGC/TBDub.

Agentic services alternate between model inference on clustered GPUs and external tools on distributed CPUs. In this scenario, existing systems (1) overlook tool-return readiness in service evaluation, wasting limited resources, and (2) use heuristic strategies to decide admission and placement on GPUs, resulting in suboptimal scheduling decisions. To bridge this gap, we present TAO, a tool-aware joint scheduling framework for LLM-based agentic serving. Its value interface combines near-term tool readiness, node-specific cache credit with time-aware decay, and token footprint; a dynamic program jointly selects and places waiting sessions under capacity constraints. Across the 12 reported model-workload-concurrency settings, TAO reduces latency by up to 26.8%, reduces mean time to first token (TTFT) by up to 36.3%, and increases decode throughput by up to 29.5% relative to the best evaluated state-of-the-art baseline for each metric.

As large language models (LLMs) evolve to handle increasingly longer contexts, serving inference requests for context lengths in the range of millions of tokens presents unique challenges. While existing techniques are effective for training, they fail to address the unique challenges of inference, such as varying prefill and decode phases and their associated latency constraints - like Time to First Token (TTFT) and Time Between Tokens (TBT). Furthermore, there are no long context inference solutions that allow batching requests to increase the hardware utilization today. In this paper, we propose three key innovations for efficient interactive long context LLM inference, without resorting to any approximation: adaptive chunking to reduce prefill overheads in mixed batching, Sequence Pipeline Parallelism (SPP) to lower TTFT, and KV Cache Parallelism (KVP) to minimize TBT. These contributions are combined into a 3D parallelism strategy, enabling Mnemosyne to scale interactive inference to context lengths at least up to 10 million tokens with high throughput enabled with batching. To our knowledge, Mnemosyne is the first to be able to achieve support for 10 million long context inference efficiently, while satisfying production-grade SLOs on TBT (30ms) on contexts up to and including 10 million.

Deploying large language models (LLMs) for online inference is often constrained by limited GPU memory, particularly due to the growing KV cache during auto-regressive decoding. Hybrid GPU-CPU execution has emerged as a promising solution by offloading KV cache management and parts of attention computation to the CPU. However, a key bottleneck remains: existing schedulers fail to effectively overlap CPU-offloaded tasks with GPU execution during the latency-critical, bandwidth-bound decode phase. This particularly penalizes real-time, decode-heavy applications (e.g., chat, Chain-of-Thought reasoning) which are currently underserved by existing systems, especially under memory pressure typical of edge or low-cost deployments. We present APEX, a novel, profiling-informed scheduling strategy that maximizes CPU-GPU parallelism during hybrid LLM inference. Unlike systems relying on static rules or purely heuristic approaches, APEX dynamically dispatches compute across heterogeneous resources by predicting execution times of CPU and GPU subtasks to maximize overlap while avoiding scheduling overheads. We evaluate APEX on diverse workloads and GPU architectures (NVIDIA T4, A10), using LLaMa-2-7B and LLaMa-3.1-8B models. Compared to GPU-only schedulers like VLLM, APEX improves throughput by 84% - 96% on T4 and 11% - 89% on A10 GPUs, while preserving latency. Against the best existing hybrid schedulers, it delivers up to 49% (T4) and 37% (A10) higher throughput in long-output settings. APEX significantly advances hybrid LLM inference efficiency on such memory-constrained hardware and provides a blueprint for scheduling in heterogeneous AI systems, filling a critical gap for efficient real-time LLM applications.

The key-value (KV) cache dominates memory bandwidth and footprint in long-context autoregressive inference. Recent rotation-preconditioned codecs (TurboQuant, PolarQuant) show that a structured random rotation followed by a per-coordinate scalar quantizer matched to an analytically tractable marginal is a near-optimal recipe for KV compression. OCTOPUS advances this paradigm through joint quantization of rotated coordinate triplets. Each triplet's direction is mapped to a square via an octahedral parameterization, and the two resulting coordinates and the triplet norm are Lloyd-Max quantized against implementation-matched marginals. Optimizing the per-triplet squared error gives a strictly non-uniform bit allocation depending only on the total dimensionality of the keys. We find the finite-dimensional quality optimum with sweeps to be constant on every real decoder we test. The codec is data-oblivious, online, and deterministic given a seed. Across text, video, and audio, OCTOPUS matches or beats every prior rotation codec at every reported bit width and metric, with a lead that grows as bits drop for extreme compression. Furthermore, a fused Triton implementation reconstructs keys on the fly without materializing the uncompressed key, so the codec adds no decode-time bandwidth or latency over the existing dequantization. Project Page: https://octopus-quant.github.io/

Concurrent local LLM serving on unified-memory desktops must preserve memory headroom and output fidelity, which speed-only rankings overlook. We introduce SiliconBench, which evaluates nine Apple Silicon serving engines through three lenses: speed, memory, and fidelity. We evaluate chat and agent serving on Qwen3, Qwen3.5, and Gemma 4. We use a classification task to check for quality regressions against an NVIDIA reference. DGX Spark provides a complementary serving-performance reference. Three desiderata guide interpretation: serving architecture readiness, memory discipline, and multi-node scaling. On Qwen3-0.6B, vllm-metal alone more than doubles throughput on both workloads from concurrency 1 to 16. CUDA vLLM and SGLang show stronger concurrency scaling on the same prompts. Explicit memory budgets do not guarantee memory headroom: two stacks complete every request while memory use approaches physical capacity and throughput declines. The newer model architectures have narrower engine support. Their evaluated implementations match the fidelity reference. Only three stacks satisfy the completion, fidelity, and model-coverage gates. Comparisons on larger dense and MoE models reinforce the importance of scheduling prompt processing alongside ongoing generation: vllm-metal's packed prefill-decode path maintains lower first-token latency than omlx under concurrent load. In the tested two-machine configurations, tensor parallelism over Thunderbolt RDMA scales while pipeline parallelism over TCP regresses. We release benchmark code, per-run results, and maintenance journals, supported by a workflow combining bounded agent fixes with human review.

State--prediction separation (SPS) relieves a language model's hidden state of two competing burdens---summarizing the context and predicting the next token---by splitting the forward pass into a state stream and a prediction stream. The separation works, but it is expensive: the prediction stream is a second pass over the whole backbone, costing sim1.9times the pretraining FLOPs, and even more in terms of wall-clock time when using a flexible attention mask. This paper makes state--prediction separation almost free. We take the separation to its limit with a free pause token: a prediction stream that writes no keys or values at all and so rides the sequence's existing positions. It improves next-token prediction of a standard Transformer by 2-3 centinats in practice on a 1B parameter model, and because it adds no position it costs nothing at inference---no added context length, no KV cache, no decode steps, and essentially no latency, with the growth in inference flops typically irrelevant as it is not the active bottleneck on throughput. The cost is therefore entirely in training where we use four mechanisms to drive it down: a two-pass split that keeps FlashAttention kernels viable, the w{=}0 prediction window, a shared gated FFN that evaluates one FFN per position rather than one per stream, and phasing the separation onto the tail of the run. Together these bring the overhead versus an optimized pretraining pipeline to 1.33times wall-clock while recovering ~94% of the gain compared to SPS, and to as low as 1.09times along a graceful quality/compute tradeoff. Furthermore, the FFN optimization reduces the raw flops required at inference time. The result is an isoflop, isoparameter, and isotoken improvement over standard next token trained transformers.

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 2.39times 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.