Daily Papers

Neural posterior estimation has emerged as a powerful tool for amortized inference, with growing adoption across scientific and applied domains. In many of these applications, the conditioning variable is a set of observations whose elements depend not only on the target but also on unknown factors shared across the set. Optimal inference therefore requires treating the set jointly, which in turn requires training the estimator at the deployment set size -- a regime where memory and compute quickly become prohibitive. We introduce a simple, theoretically grounded strategy that decouples representation learning from posterior modeling. Our method trains a mean-pool Deep Set on sets of size at most two, producing an encoder that generalizes to arbitrary set sizes. The inference head is then finetuned on pre-aggregated embeddings, making training cost essentially independent of the deployment set size N. Across scalar, image, multi-view 3D, molecular, and high-dimensional conditional generation benchmarks with N in the thousands, our approach matches or outperforms standard baselines at a fraction of the compute.

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.

Existing methods for capturing datasets of 3D heads in dense semantic correspondence are slow, and commonly address the problem in two separate steps; multi-view stereo (MVS) reconstruction followed by non-rigid registration. To simplify this process, we introduce TEMPEH (Towards Estimation of 3D Meshes from Performances of Expressive Heads) to directly infer 3D heads in dense correspondence from calibrated multi-view images. Registering datasets of 3D scans typically requires manual parameter tuning to find the right balance between accurately fitting the scans surfaces and being robust to scanning noise and outliers. Instead, we propose to jointly register a 3D head dataset while training TEMPEH. Specifically, during training we minimize a geometric loss commonly used for surface registration, effectively leveraging TEMPEH as a regularizer. Our multi-view head inference builds on a volumetric feature representation that samples and fuses features from each view using camera calibration information. To account for partial occlusions and a large capture volume that enables head movements, we use view- and surface-aware feature fusion, and a spatial transformer-based head localization module, respectively. We use raw MVS scans as supervision during training, but, once trained, TEMPEH directly predicts 3D heads in dense correspondence without requiring scans. Predicting one head takes about 0.3 seconds with a median reconstruction error of 0.26 mm, 64% lower than the current state-of-the-art. This enables the efficient capture of large datasets containing multiple people and diverse facial motions. Code, model, and data are publicly available at https://tempeh.is.tue.mpg.de.

Transformer models rely on self-attention to capture token dependencies but face challenges in effectively integrating positional information while allowing multi-head attention (MHA) flexibility. Prior methods often model semantic and positional differences disparately or apply uniform positional adjustments across heads, potentially limiting representational capacity. This paper introduces ComplexFormer, featuring Complex Multi-Head Attention-CMHA. CMHA empowers each head to independently model semantic and positional differences unified within the complex plane, representing interactions as rotations and scaling. ComplexFormer incorporates two key improvements: (1) a per-head Euler transformation, converting real-valued query/key projections into polar-form complex vectors for head-specific complex subspace operation; and (2) a per-head adaptive differential rotation mechanism, exp[i(Adapt(ASmn,i) + Delta(Pmn),i)], allowing each head to learn distinct strategies for integrating semantic angle differences (ASmn,i) with relative positional encodings (Delta(Pmn),i). Extensive experiments on language modeling, text generation, code generation, and mathematical reasoning show ComplexFormer achieves superior performance, significantly lower generation perplexity , and improved long-context coherence compared to strong baselines like RoPE-Transformers. ComplexFormer demonstrates strong parameter efficiency, offering a more expressive, adaptable attention mechanism.

The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sharing undermines model performance by neglecting the functional diversity of attention heads. To address this, we propose LycheeDecode, an efficient decoding method centered on a fine-grained hybrid-head attention mechanism that employs a hardware-efficient top-k selection strategy. Specifically, the novel HardKuma-based mechanism partitions attention heads into a small subset of retrieval heads that dynamically identify crucial tokens and a majority of sparse heads that reuse them for efficient computation. Through extensive experiments on leading models like Llama3 and Qwen3 across diverse benchmarks for long-context understanding (e.g., LongBench, RULER) and complex reasoning (e.g., AIME24, OlympiadBench), we demonstrate that LycheeDecode achieves generative quality comparable to, and at times surpassing even the full-attention baseline. Crucially, this is accomplished with up to a 2.7x speedup at a 128K context length. By preserving the functional diversity of attention heads, our fine-grained strategy overcomes the performance bottlenecks of existing methods, providing a powerful and validated pathway to both efficient and high-quality long-context LLM inference.

Large language models (LLMs) have shown remarkable advances in supporting long-context comprehension and processing tasks. However, scaling the generation inference of LLMs to such long contexts incurs significant additional computation load, and demands a substantial GPU memory footprint to maintain the key-value (KV) cache of transformer-based LLMs. Existing KV cache compression methods, such as quantization, face memory bottlenecks as context length increases, while static-sized caches, such as eviction, suffer from inefficient policies. These limitations restrict deployment on consumer-grade devices like a single Nvidia 4090 GPU. To overcome this, we propose Locret, a framework for long-context LLM inference that introduces retaining heads to evaluate the causal importance of KV cache units, allowing for more accurate eviction within a fixed cache size. Locret is fine-tuned on top of the frozen backbone LLM using a minimal amount of data from standard long-context SFT datasets. During inference, we evict low-importance cache units along with a chunked prefill pattern, significantly reducing peak GPU memory usage. We conduct an extensive empirical study to evaluate Locret, where the experimental results show that Locret outperforms the recent competitive approaches, including InfLLM, Quantization, SirLLM, and MInference, in terms of memory efficiency and the quality of generated contents -- Locret achieves over a 20x and 8x KV cache compression ratio compared to the full KV cache for Phi-3-mini-128K and Llama-3.1-8B-instruct. Additionally, Locret can be combined with other methods, such as quantization and token merging. To our knowledge, Locret is the first framework capable of deploying Llama-3.1-8B or similar models on a single Nvidia 4090 GPU, enabling 128K long-context inference without compromising generation quality, and requiring little additional system optimizations.

Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to high latency from excessive token generation, or unstable reasoning that alternates between underthinking (shallow, inconsistent steps) and overthinking (repetitive, verbose reasoning). In this work, we study the structure of reasoning trajectories and uncover specialized attention heads that correlate with distinct cognitive behaviors such as verification and backtracking. By lightly intervening on these heads at inference time, we can steer the model away from inefficient modes. Building on this insight, we propose CREST, a training-free method for Cognitive REasoning Steering at Test-time. CREST has two components: (1) an offline calibration step that identifies cognitive heads and derives head-specific steering vectors, and (2) an inference-time procedure that rotates hidden representations to suppress components along those vectors. CREST adaptively suppresses unproductive reasoning behaviors, yielding both higher accuracy and lower computational cost. Across diverse reasoning benchmarks and models, CREST improves accuracy by up to 17.5% while reducing token usage by 37.6%, offering a simple and effective pathway to faster, more reliable LLM reasoning.

The adoption of Vision Transformers (ViTs) in resource-constrained applications necessitates improvements in inference throughput. To this end several token pruning and merging approaches have been proposed that improve efficiency by successively reducing the number of tokens. However, it remains an open problem to design a token reduction method that is fast, maintains high performance, and is applicable to various vision tasks. In this work, we present a token pruner that uses auxiliary prediction heads that learn to select tokens end-to-end based on task relevance. These auxiliary heads can be removed after training, leading to throughput close to that of a random pruner. We evaluate our method on image classification, semantic segmentation, object detection, and instance segmentation, and show speedups of 1.5 to 4x with small drops in performance. As a best case, on the ADE20k semantic segmentation benchmark, we observe a 2x speedup relative to the no-pruning baseline, with a negligible performance penalty of 0.1 median mIoU across 5 seeds.

Identifying the frames of news is important to understand the articles' vision, intention, message to be conveyed, and which aspects of the news are emphasized. Framing is a widely studied concept in journalism, and has emerged as a new topic in computing, with the potential to automate processes and facilitate the work of journalism professionals. In this paper, we study this issue with articles related to the Covid-19 anti-vaccine movement. First, to understand the perspectives used to treat this theme, we developed a protocol for human labeling of frames for 1786 headlines of No-Vax movement articles of European newspapers from 5 countries. Headlines are key units in the written press, and worth of analysis as many people only read headlines (or use them to guide their decision for further reading.) Second, considering advances in Natural Language Processing (NLP) with large language models, we investigated two approaches for frame inference of news headlines: first with a GPT-3.5 fine-tuning approach, and second with GPT-3.5 prompt-engineering. Our work contributes to the study and analysis of the performance that these models have to facilitate journalistic tasks like classification of frames, while understanding whether the models are able to replicate human perception in the identification of these frames.

We set out to train behavioral dispositions (self-verification, uncertainty acknowledgment, feedback integration) into small language models (0.6B to 2.3B effective parameters) through a four-stage all-MIT distillation pipeline, with follow-on experiments on inference-time attention-head interventions and a frozen-base confidence-gated sidecar. An internal draft reported +33.9-point MCAS and +15.3-point HumanEval gains on a Qwen3-0.6B student; a second-pass sanity check falsified both numbers before publication. The HumanEval delta was a truncation artifact (n_predict=512) that inverted to -8.0 points at n_predict=1024; the MCAS gain disappeared under apples-to-apples scoring. That falsification triggered three subsequent arcs. Across (1) SFT/DPO LoRA on three model families and two domains, (2) inference-time attention-head tempering on o_proj, and (3) a training-free frozen-base sidecar reading the final-token hidden state h_last, we find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry. The failure is consistent across five models (Qwen3-0.6B, Qwen3-1.7B, Qwen3.5-0.8B, Gemma 4 E2B, and SmolLM2-1.7B-Instruct). A within-distribution cross-validation pass (AUC=0.683) collapsed to chance on fresh prompts (AUC=0.516). We contribute a three-arc negative result with mechanism, a two-failure-mode taxonomy for linear h_last probes, and an honest falsification pipeline that converts the class of false positives we ourselves produced into publishable negatives. As an independent finding, Gemma 4 E2B exhibits near-complete confidence-correctness decoupling on the Chef domain (assertion asymmetry -0.009; the model asserts at 91% regardless of correctness).

Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.

Large Language Models (LLMs) with hundreds of billions of parameters have transformed the field of machine learning. However, serving these models at inference time is both compute and memory intensive, where a single request can require multiple GPUs and tens of Gigabytes of memory. Multi-Head Attention is one of the key components of LLMs, which can account for over 50% of LLMs memory and compute requirement. We observe that there is a high amount of redundancy across heads on which tokens they pay attention to. Based on this insight, we propose Clustered Head Attention (CHAI). CHAI combines heads with a high amount of correlation for self-attention at runtime, thus reducing both memory and compute. In our experiments, we show that CHAI is able to reduce the memory requirements for storing K,V cache by up to 21.4% and inference time latency by up to 1.73x without any fine-tuning required. CHAI achieves this with a maximum 3.2% deviation in accuracy across 3 different models (i.e. OPT-66B, LLAMA-7B, LLAMA-33B) and 5 different evaluation datasets.

Large Language Models (LLMs) often struggle with complex reasoning scenarios. While preference optimization methods enhance reasoning performance through training, they often lack transparency in why one reasoning outcome is preferred over another. Verbal reflection techniques improve explainability but are limited in LLMs' critique and refinement capacity. To address these challenges, we introduce a contrastive reflection synthesis pipeline that enhances the accuracy and depth of LLM-generated reflections. We further propose a dual-model reasoning framework within a verbal reinforcement learning paradigm, decoupling inference-time self-reflection into specialized, trained models for reasoning critique and refinement. Extensive experiments show that our framework outperforms traditional preference optimization methods across all evaluation metrics. Our findings also show that "two heads are better than one", demonstrating that a collaborative Reasoner-Critic model achieves superior reasoning performance and transparency, compared to single-model approaches.

The inference process in Large Language Models (LLMs) is often limited due to the absence of parallelism in the auto-regressive decoding process, resulting in most operations being restricted by the memory bandwidth of accelerators. While methods such as speculative decoding have been suggested to address this issue, their implementation is impeded by the challenges associated with acquiring and maintaining a separate draft model. In this paper, we present Medusa, an efficient method that augments LLM inference by adding extra decoding heads to predict multiple subsequent tokens in parallel. Using a tree-based attention mechanism, Medusa constructs multiple candidate continuations and verifies them simultaneously in each decoding step. By leveraging parallel processing, Medusa introduces only minimal overhead in terms of single-step latency while substantially reducing the number of decoding steps required. We present two levels of fine-tuning procedures for Medusa to meet the needs of different use cases: Medusa-1: Medusa is directly fine-tuned on top of a frozen backbone LLM, enabling lossless inference acceleration. Medusa-2: Medusa is fine-tuned together with the backbone LLM, enabling better prediction accuracy of Medusa heads and higher speedup but needing a special training recipe that preserves the backbone model's capabilities. Moreover, we propose several extensions that improve or expand the utility of Medusa, including a self-distillation to handle situations where no training data is available and a typical acceptance scheme to boost the acceptance rate while maintaining generation quality. We evaluate Medusa on models of various sizes and training procedures. Our experiments demonstrate that Medusa-1 can achieve over 2.2x speedup without compromising generation quality, while Medusa-2 further improves the speedup to 2.3-3.6x.

Although applications involving long-context inputs are crucial for the effective utilization of large language models (LLMs), they also result in increased computational costs and reduced performance. To address this challenge, we propose an efficient, training-free prompt compression method that retains key information within compressed prompts. We identify specific attention heads in transformer-based LLMs, which we designate as evaluator heads, that are capable of selecting tokens in long inputs that are most significant for inference. Building on this discovery, we develop EHPC, an Evaluator Head-based Prompt Compression method, which enables LLMs to rapidly "skim through" input prompts by leveraging only the first few layers with evaluator heads during the pre-filling stage, subsequently passing only the important tokens to the model for inference. EHPC achieves state-of-the-art results across two mainstream benchmarks: prompt compression and long-context inference acceleration. Consequently, it effectively reduces the complexity and costs associated with commercial API calls. We further demonstrate that EHPC attains competitive results compared to key-value cache-based acceleration methods, thereby highlighting its potential to enhance the efficiency of LLMs for long-context tasks.

Pre-trained Language Models (LMs) exhibit strong zero-shot and in-context learning capabilities; however, their behaviors are often difficult to control. By utilizing Reinforcement Learning from Human Feedback (RLHF), it is possible to fine-tune unsupervised LMs to follow instructions and produce outputs that reflect human preferences. Despite its benefits, RLHF has been shown to potentially harm a language model's reasoning capabilities and introduce artifacts such as hallucinations where the model may fabricate facts. To address this issue we introduce Direct Preference Heads (DPH), a fine-tuning framework that enables LMs to learn human preference signals through an auxiliary reward head without directly affecting the output distribution of the language modeling head. We perform a theoretical analysis of our objective function and find strong ties to Conservative Direct Preference Optimization (cDPO). Finally we evaluate our models on GLUE, RACE, and the GPT4All evaluation suite and demonstrate that our method produces models which achieve higher scores than those fine-tuned with Supervised Fine-Tuning (SFT) or Direct Preference Optimization (DPO) alone.

Large Language Models (LLMs) inherently use autoregressive decoding, which lacks parallelism in inference and results in significantly slow inference speeds, especially when hardware parallel accelerators and memory bandwidth are not fully utilized. In this work, we propose Amphista, a speculative decoding algorithm that adheres to a non-autoregressive decoding paradigm. Owing to the increased parallelism, our method demonstrates higher efficiency in inference compared to autoregressive methods. Specifically, Amphista models an Auto-embedding Block capable of parallel inference, incorporating bi-directional attention to enable interaction between different drafting heads. Additionally, Amphista implements Staged Adaptation Layers to facilitate the transition of semantic information from the base model's autoregressive inference to the drafting heads' non-autoregressive speculation, thereby achieving paradigm transformation and feature fusion. We conduct a series of experiments on a suite of Vicuna models using MT-Bench and Spec-Bench. For the Vicuna 33B model, Amphista achieves up to 2.75times and 1.40times wall-clock acceleration compared to vanilla autoregressive decoding and Medusa, respectively, while preserving lossless generation quality.

Language models are increasingly adopting smaller architectures optimized for consumer devices. In this setting, inference efficiency is the primary constraint. Meanwhile, vocabulary sizes continue to grow rapidly, making the classification head a critical bottleneck that accounts for up to 60\% of model parameters, and 50\% of inference compute. We introduce FlashHead, the first efficient drop-in replacement for the dense classification head that is training-free and hardware-friendly. FlashHead builds on principles from information retrieval, reframing that computation at the output head as a retrieval problem rather than a dense classification over the full vocabulary. FlashHead introduces four key innovations: (1) a balanced clustering scheme that structures vocabulary partitions into compact hardware-efficient tensors, (2) extending multiprobe retrieval to language model heads, enabling thousands of clusters to be scored in parallel, (3) a novel inference-time sampling mechanism that extends retrieval beyond top tokens, enabling probabilistic sampling across the full vocabulary, and (4) selective quantization, enabling effective low-bit computation in the head. Experiments on Llama-3.2, Gemma-3, and Qwen-3 show that FlashHead delivers model-level inference speedups of up to 1.75x which maintaining output accuracy compared to the original head. By overcoming the classification head bottleneck, FlashHead establishes a new benchmark for efficient inference and removes a key barrier to developing smaller, capable models for consumer hardware.

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.

The quadratic complexity of attention mechanisms poses a critical bottleneck for large language models processing long contexts. While dynamic sparse attention methods offer input-adaptive efficiency, they face fundamental trade-offs: requiring preprocessing, lacking global evaluation, violating query independence, or incurring high computational overhead. We present RRAttention, a novel dynamic sparse attention method that simultaneously achieves all desirable properties through a head round-robin (RR) sampling strategy. By rotating query sampling positions across attention heads within each stride, RRAttention maintains query independence while enabling efficient global pattern discovery with stride-level aggregation. Our method reduces complexity from O(L^2) to O(L^2/S^2) and employs adaptive Top-τ selection for optimal sparsity. Extensive experiments on natural language understanding (HELMET) and multimodal video comprehension (Video-MME) demonstrate that RRAttention recovers over 99\% of full attention performance while computing only half of the attention blocks, achieving 2.4times speedup at 128K context length and outperforming existing dynamic sparse attention methods.

Long-context inference in large language models is bottlenecked by Key--Value (KV) cache loading during the decoding stage, where the sequential nature of generation requires repeatedly transferring the KV cache from off-chip High-Bandwidth Memory (HBM) to on-chip Static Random-Access Memory (SRAM) at each step. While Multi-Head Latent Attention (MLA) significantly reduces the total KV cache size, it suffers from a sharding bottleneck during distributed decoding via Tensor Parallelism (TP). Since its single latent head cannot be partitioned, each device is forced to redundantly load the complete KV cache for every token, consuming excessive memory traffic and diminishing TP benefits like weight sharding. In this work, we propose Multi-Head Low-Rank Attention (MLRA), which enables partitionable latent states for efficient 4-way TP decoding. Extensive experiments show that MLRA achieves state-of-the-art perplexity and downstream task performance, while also delivering a 2.8times decoding speedup over MLA. Code is available at https://github.com/SongtaoLiu0823/MLRA. Pretrained weights, along with the training and evaluation data, are available at https://huggingface.co/Soughing/MLRA.

Visual Geometry Grounded Transformer (VGGT) has advanced 3D vision, yet its global attention layers suffer from quadratic computational costs that hinder scalability. Several sparsification-based acceleration techniques have been proposed to alleviate this issue, but they often suffer from substantial accuracy degradation. We hypothesize that the accuracy degradation stems from the heterogeneity in head-wise sparsification sensitivity, as the existing methods apply a uniform sparsity pattern across all heads. Motivated by this hypothesis, we present a two-stage sparsification pipeline that effectively quantifies and exploits headwise sparsification sensitivity. In the first stage, we measure head-wise sparsification sensitivity using a novel metric, the Head Sensitivity Score (HeSS), which approximates the Hessian with respect to two distinct error terms on a small calibration set. In the inference stage, we perform HeSS-Guided Sparsification, leveraging the pre-computed HeSS to reallocate the total attention budget-assigning denser attention to sensitive heads and sparser attention to more robust ones. We demonstrate that HeSS effectively captures head-wise sparsification sensitivity and empirically confirm that attention heads in the global attention layers exhibit heterogeneous sensitivity characteristics. Extensive experiments further show that our method effectively mitigates performance degradation under high sparsity, demonstrating strong robustness across varying sparsification levels. Code is available at https://github.com/libary753/HeSS.

This paper presents a novel technique for accelerating inference in large, pre-trained language models (LLMs) by introducing early exits during inference. The computational demands of these models, used across a wide range of applications, can be substantial. By capitalizing on the inherent variability in token complexity, our approach enables selective acceleration of the inference process. Specifically, we propose the integration of early exit ''heads'' atop existing transformer layers, which facilitate conditional terminations based on a confidence metric. These heads are trained in a self-supervised manner using the model's own predictions as training data, thereby eliminating the need for additional annotated data. The confidence metric, established using a calibration set, ensures a desired level of accuracy while enabling early termination when confidence exceeds a predetermined threshold. Notably, our method preserves the original accuracy and reduces computational time on certain tasks, leveraging the existing knowledge of pre-trained LLMs without requiring extensive retraining. This lightweight, modular modification has the potential to greatly enhance the practical usability of LLMs, particularly in applications like real-time language processing in resource-constrained environments.

We study the problem of efficient generative inference for Transformer models, in one of its most challenging settings: large deep models, with tight latency targets and long sequence lengths. Better understanding of the engineering tradeoffs for inference for large Transformer-based models is important as use cases of these models are growing rapidly throughout application areas. We develop a simple analytical model for inference efficiency to select the best multi-dimensional partitioning techniques optimized for TPU v4 slices based on the application requirements. We combine these with a suite of low-level optimizations to achieve a new Pareto frontier on the latency and model FLOPS utilization (MFU) tradeoffs on 500B+ parameter models that outperforms the FasterTransformer suite of benchmarks. We further show that with appropriate partitioning, the lower memory requirements of multiquery attention (i.e. multiple query heads share single key/value head) enables scaling up to 32x larger context lengths. Finally, we achieve a low-batch-size latency of 29ms per token during generation (using int8 weight quantization) and a 76% MFU during large-batch-size processing of input tokens, while supporting a long 2048-token context length on the PaLM 540B parameter model.

With the development of large language models (LLMs), the ability to handle longer contexts has become a key capability for Web applications such as cross-document understanding and LLM-powered search systems. However, this progress faces two major challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues hinder the application of LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a model-agnostic, training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using Query-Key dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a small number of critical KV cache tokens in the attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we designed the Selection Cache based on observations of consecutive Query similarity and implemented efficient dot product kernel, significantly reducing the overhead of token selection. A comprehensive evaluation of TokenSelect demonstrates up to 23.84x speedup in attention computation and up to 2.28x acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

Multi-Head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value states into a low-rank latent vector, caching only this vector to reduce memory. In tensor parallelism (TP), however, attention heads are computed across multiple devices, and each device must load the full cache, eroding the advantage of MLA over Grouped Query Attention (GQA). We propose Tensor-Parallel Latent Attention (TPLA): a scheme that partitions both the latent representation and each head's input dimension across devices, performs attention independently per shard, and then combines results with an all-reduce. TPLA preserves the benefits of a compressed KV cache while unlocking TP efficiency. Unlike Grouped Latent Attention (GLA), every head in TPLA still leverages the full latent representation, maintaining stronger representational capacity. TPLA is drop-in compatible with models pre-trained using MLA: it supports MLA-style prefilling and enables efficient tensor-parallel decoding without retraining. Applying simple orthogonal transforms -- e.g., the Hadamard transform or PCA -- before TP slicing further mitigates cross-shard interference, yielding minimal accuracy degradation. By reducing the per-device KV cache for DeepSeek-V3 and Kimi-K2, we achieve 1.79x and 1.93x speedups, respectively, at a 32K-token context length while maintaining performance on commonsense and LongBench benchmarks. TPLA can be implemented with FlashAttention-3, enabling practical end-to-end acceleration.

The generation of stylistic 3D facial animations driven by speech poses a significant challenge as it requires learning a many-to-many mapping between speech, style, and the corresponding natural facial motion. However, existing methods either employ a deterministic model for speech-to-motion mapping or encode the style using a one-hot encoding scheme. Notably, the one-hot encoding approach fails to capture the complexity of the style and thus limits generalization ability. In this paper, we propose DiffPoseTalk, a generative framework based on the diffusion model combined with a style encoder that extracts style embeddings from short reference videos. During inference, we employ classifier-free guidance to guide the generation process based on the speech and style. We extend this to include the generation of head poses, thereby enhancing user perception. Additionally, we address the shortage of scanned 3D talking face data by training our model on reconstructed 3DMM parameters from a high-quality, in-the-wild audio-visual dataset. Our extensive experiments and user study demonstrate that our approach outperforms state-of-the-art methods. The code and dataset will be made publicly available.

Auto-regressive inference of transformers benefit greatly from Key-Value (KV) caching, but can lead to major memory bottlenecks as model size, batch size, and sequence length grow at scale. We introduce Multi-Layer Key-Value (MLKV) sharing, a novel approach extending KV sharing across transformer layers to reduce memory usage beyond what was possible with Multi-Query Attention (MQA) and Grouped-Query Attention (GQA). Evaluations on various NLP benchmarks and inference metrics using uptrained Pythia-160M variants demonstrate that MLKV significantly reduces memory usage with minimal performance loss, reducing KV cache size down to a factor of 6x compared to MQA. These results highlight MLKV's potential for efficient deployment of transformer models at scale. We provide code at https://github.com/zaydzuhri/pythia-mlkv

Transformer-based large language models (LLMs) excel in modeling complex language patterns but face significant computational costs during inference, especially with long inputs due to the attention mechanism's memory overhead. We observe that certain attention heads exhibit a distribution where the attention weights concentrate on tokens near the query token, termed as recency aware, which focuses on local and short-range dependencies. Leveraging this insight, we propose RecurFormer, a novel architecture that replaces these attention heads with linear recurrent neural networks (RNNs), specifically the Mamba architecture. This replacement reduces the cache size without evicting tokens, thus maintaining generation quality. RecurFormer retains the ability to model long-range dependencies through the remaining attention heads and allows for reusing pre-trained Transformer-based LLMs weights with continual training. Experiments demonstrate that RecurFormer matches the original model's performance while significantly enhancing inference efficiency. Our approach provides a practical solution to the computational challenges of Transformer-based LLMs inference, making it highly attractive for tasks involving long inputs.

Steering the behavior of Large Language Models (LLMs) remains a challenge, particularly in engineering applications where precision and reliability are critical. While fine-tuning and prompting methods can modify model behavior, they lack the dynamic and exact control necessary for engineering applications. Inference-time intervention techniques provide a promising alternative, allowing targeted adjustments to LLM outputs. In this work, we demonstrate how interventions enable fine-grained control for automating the usually time-intensive requirement verification process in Model-Based Systems Engineering (MBSE). Using two early-stage Capella SysML models of space missions with associated requirements, we apply the intervened LLMs to reason over a graph representation of the model to determine whether a requirement is fulfilled. Our method achieves robust and reliable outputs, significantly improving over both a baseline model and a fine-tuning approach. By identifying and modifying as few as one to three specialised attention heads, we can significantly change the model's behavior. When combined with self-consistency, this allows us to achieve perfect precision on our holdout test set.

The size of the key-value (KV) cache plays a critical role in determining both the maximum context length and the number of concurrent requests supported during inference in modern language models. The KV cache size grows proportionally with the number of attention heads and the tokens processed, leading to increased memory consumption and slower inference for long inputs. In this work, we explore the use of MixAttention, a model architecture modification closely related to a blog published by Character.AI. MixAttention combines sliding window attention, where only a small subset of recent tokens is stored in the KV cache, with KV cache sharing across layers. Our experiments demonstrate that MixAttention significantly reduces memory usage and improves inference speed without sacrificing model performance in both short and long-context tasks. We also explore various configurations of this architecture, identifying those that maintain quality across evaluation metrics while optimizing resource efficiency.

We introduce Self-Monitored Inference-Time INtervention (SMITIN), an approach for controlling an autoregressive generative music transformer using classifier probes. These simple logistic regression probes are trained on the output of each attention head in the transformer using a small dataset of audio examples both exhibiting and missing a specific musical trait (e.g., the presence/absence of drums, or real/synthetic music). We then steer the attention heads in the probe direction, ensuring the generative model output captures the desired musical trait. Additionally, we monitor the probe output to avoid adding an excessive amount of intervention into the autoregressive generation, which could lead to temporally incoherent music. We validate our results objectively and subjectively for both audio continuation and text-to-music applications, demonstrating the ability to add controls to large generative models for which retraining or even fine-tuning is impractical for most musicians. Audio samples of the proposed intervention approach are available on our demo page http://tinyurl.com/smitin .

The quadratic complexity and weak length extrapolation of Transformers limits their ability to scale to long sequences, and while sub-quadratic solutions like linear attention and state space models exist, they empirically underperform Transformers in pretraining efficiency and downstream task accuracy. We introduce Megalodon, a neural architecture for efficient sequence modeling with unlimited context length. Megalodon inherits the architecture of Mega (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability and stability, including complex exponential moving average (CEMA), timestep normalization layer, normalized attention mechanism and pre-norm with two-hop residual configuration. In a controlled head-to-head comparison with Llama2, Megalodon achieves better efficiency than Transformer in the scale of 7 billion parameters and 2 trillion training tokens. Megalodon reaches a training loss of 1.70, landing mid-way between Llama2-7B (1.75) and 13B (1.67). Code: https://github.com/XuezheMax/megalodon

Sparse Mixtures of Experts (SMoE) scales model capacity without significant increases in training and inference costs, but exhibits the following two issues: (1) Low expert activation, where only a small subset of experts are activated for optimization. (2) Lacking fine-grained analytical capabilities for multiple semantic concepts within individual tokens. We propose Multi-Head Mixture-of-Experts (MH-MoE), which employs a multi-head mechanism to split each token into multiple sub-tokens. These sub-tokens are then assigned to and processed by a diverse set of experts in parallel, and seamlessly reintegrated into the original token form. The multi-head mechanism enables the model to collectively attend to information from various representation spaces within different experts, while significantly enhances expert activation, thus deepens context understanding and alleviate overfitting. Moreover, our MH-MoE is straightforward to implement and decouples from other SMoE optimization methods, making it easy to integrate with other SMoE models for enhanced performance. Extensive experimental results across three tasks: English-focused language modeling, Multi-lingual language modeling and Masked multi-modality modeling tasks, demonstrate the effectiveness of MH-MoE.

Multi-head attention layers, as used in the Transformer neural sequence model, are a powerful alternative to RNNs for moving information across and between sequences. While training these layers is generally fast and simple, due to parallelizability across the length of the sequence, incremental inference (where such paralleization is impossible) is often slow, due to the memory-bandwidth cost of repeatedly loading the large "keys" and "values" tensors. We propose a variant called multi-query attention, where the keys and values are shared across all of the different attention "heads", greatly reducing the size of these tensors and hence the memory bandwidth requirements of incremental decoding. We verify experimentally that the resulting models can indeed be much faster to decode, and incur only minor quality degradation from the baseline.

Recent advances in diffusion models have revolutionized audio-driven talking head synthesis. Beyond precise lip synchronization, diffusion-based methods excel in generating subtle expressions and natural head movements that are well-aligned with the audio signal. However, these methods are confronted by slow inference speed, insufficient fine-grained control over facial motions, and occasional visual artifacts largely due to an implicit latent space derived from Variational Auto-Encoders (VAE), which prevent their adoption in realtime interaction applications. To address these issues, we introduce Ditto, a diffusion-based framework that enables controllable realtime talking head synthesis. Our key innovation lies in bridging motion generation and photorealistic neural rendering through an explicit identity-agnostic motion space, replacing conventional VAE representations. This design substantially reduces the complexity of diffusion learning while enabling precise control over the synthesized talking heads. We further propose an inference strategy that jointly optimizes three key components: audio feature extraction, motion generation, and video synthesis. This optimization enables streaming processing, realtime inference, and low first-frame delay, which are the functionalities crucial for interactive applications such as AI assistants. Extensive experimental results demonstrate that Ditto generates compelling talking head videos and substantially outperforms existing methods in both motion control and realtime performance.

Large Language Models have revolutionized natural language processing, yet serving them efficiently in data centers remains challenging due to mixed workloads comprising latency-sensitive (LS) and best-effort (BE) jobs. Existing inference systems employ iteration-level first-come-first-served scheduling, causing head-of-line blocking when BE jobs delay LS jobs. We introduce QLLM, a novel inference system designed for Mixture of Experts (MoE) models, featuring a fine-grained, priority-aware preemptive scheduler. QLLM enables expert-level preemption, deferring BE job execution while minimizing LS time-to-first-token (TTFT). Our approach removes iteration-level scheduling constraints, enabling the scheduler to preempt jobs at any layer based on priority. Evaluations on an Nvidia A100 GPU show that QLLM significantly improves performance. It reduces LS TTFT by an average of 65.5times and meets the SLO at up to 7 requests/sec, whereas the baseline fails to do so under the tested workload. Additionally, it cuts LS turnaround time by up to 12.8times without impacting throughput. QLLM is modular, extensible, and seamlessly integrates with Hugging Face MoE models.

Large language models (LLMs) frequently generate toxic content, posing significant risks for safe deployment. Current mitigation strategies often degrade generation quality or require costly human annotation. We propose CAUSALDETOX, a framework that identifies and intervenes on the specific attention heads causally responsible for toxic generation. Using the Probability of Necessity and Sufficiency (PNS), we isolate a minimal set of heads that are necessary and sufficient for toxicity. We utilize these components via two complementary strategies: (1) Local Inference-Time Intervention, which constructs dynamic, input-specific steering vectors for context-aware detoxification, and (2) PNS-Guided Fine-Tuning, which permanently unlearns toxic representations. We also introduce PARATOX, a novel benchmark of aligned toxic/non-toxic sentence pairs enabling controlled counterfactual evaluation. Experiments on ToxiGen, ImplicitHate, and ParaDetox show that CAUSALDETOX achieves up to 5.34% greater toxicity reduction compared to baselines while preserving linguistic fluency, and offers a 7x speedup in head selection.

Transformer encoders are widely deployed in large-scale web services for natural language understanding tasks such as text classification, semantic retrieval, and content ranking. However, their high inference latency and memory consumption pose significant challenges for real-time serving and scalability. These limitations stem largely from architectural redundancy, particularly in the attention module. The inherent parameter redundancy of the attention mechanism, coupled with the fact that its attention heads operate with a degree of independence, makes it particularly amenable to structured model compression. In this paper, we propose SHRP (Specialized Head Routing and Pruning), a novel structured pruning framework that automatically identifies and removes redundant attention heads while preserving most of the model's accuracy and compatibility. SHRP introduces Expert Attention, a modular design that treats each attention head as an independent expert, followed by a lightweight shared expander feed-forward network that refines their outputs. The framework employs a unified Top-1 usage-driven mechanism to jointly perform dynamic routing during training and deterministic pruning at deployment. Experimental results on the GLUE benchmark using a BERT-base encoder show that SHRP achieves 93% of the original model accuracy while reducing parameters by 48 percent. Under an extreme compression scenario where 11/12 of the layers are pruned, the model still maintains 84% accuracy and delivers a 4.2x throughput gain while reducing computation to as low as 11.5 percent of the original FLOPs, demonstrating its practical utility for large-scale and latency-sensitive web deployments.

As large language models (LLMs) become increasingly powerful, the sequential nature of autoregressive generation creates a fundamental throughput bottleneck that limits the practical deployment. While Multi-Token Prediction (MTP) has demonstrated remarkable benefits for model training efficiency and performance, its inherent potential for inference acceleration remains largely unexplored. This paper introduces FastMTP, a simple yet effective method that improves multi-step draft quality by aligning MTP training with its inference pattern, significantly enhancing speculative decoding performance. Our approach fine-tunes a single MTP head with position-shared weights on self-distilled data, enabling it to capture dependencies among consecutive future tokens and maintain high acceptance rates across multiple recursive draft steps. By integrating language-aware dynamic vocabulary compression into the MTP head, we further reduce computational overhead in the drafting process. Experimental results across seven diverse benchmarks demonstrate that FastMTP achieves an average of 2.03x speedup compared to standard next token prediction with lossless output quality, outperforming vanilla MTP by 82%. FastMTP requires only lightweight training and seamlessly integrates with existing inference frameworks, offering a practical and rapidly deployable solution for accelerating LLM inference.

Recent advances in motion planning for autonomous driving have led to models capable of generating high-quality trajectories. However, most existing planners tend to fix their policy after supervised training, leading to consistent but rigid driving behaviors. This limits their ability to reflect human preferences or adapt to dynamic, instruction-driven demands. In this work, we propose a diffusion-based multi-head trajectory planner(M-diffusion planner). During the early training stage, all output heads share weights to learn to generate high-quality trajectories. Leveraging the probabilistic nature of diffusion models, we then apply Group Relative Policy Optimization (GRPO) to fine-tune the pre-trained model for diverse policy-specific behaviors. At inference time, we incorporate a large language model (LLM) to guide strategy selection, enabling dynamic, instruction-aware planning without switching models. Closed-loop simulation demonstrates that our post-trained planner retains strong planning capability while achieving state-of-the-art (SOTA) performance on the nuPlan val14 benchmark. Open-loop results further show that the generated trajectories exhibit clear diversity, effectively satisfying multi-modal driving behavior requirements. The code and related experiments will be released upon acceptance of the paper.

Existing LLM training and inference frameworks struggle in boosting efficiency with sparsity while maintaining the integrity of context and model architecture. Inspired by the sharding concept in database and the fact that attention parallelizes over heads on accelerators, we propose Sparsely-Sharded (S2) Attention, an attention algorithm that allocates heterogeneous context partitions for different attention heads to divide and conquer. S2-Attention enforces each attention head to only attend to a partition of contexts following a strided sparsity pattern, while the full context is preserved as the union of all the shards. As attention heads are processed in separate thread blocks, the context reduction for each head can thus produce end-to-end speed-up and memory reduction. At inference, LLMs trained with S2-Attention can then take the KV cache reduction as free meals with guaranteed model quality preserve. In experiments, we show S2-Attentioncan provide as much as (1) 25.3X wall-clock attention speed-up over FlashAttention-2, resulting in 6X reduction in end-to-end training time and 10X inference latency, (2) on-par model training quality compared to default attention, (3)perfect needle retrieval accuracy over 32K context window. On top of the algorithm, we build DKernel, an LLM training and inference kernel library that allows users to customize sparsity patterns for their own models. We open-sourced DKerneland make it compatible with Megatron, Pytorch, and vLLM.

We introduce Inference-Time Intervention (ITI), a technique designed to enhance the truthfulness of large language models (LLMs). ITI operates by shifting model activations during inference, following a set of directions across a limited number of attention heads. This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark. On an instruction-finetuned LLaMA called Alpaca, ITI improves its truthfulness from 32.5% to 65.1%. We identify a tradeoff between truthfulness and helpfulness and demonstrate how to balance it by tuning the intervention strength. ITI is minimally invasive and computationally inexpensive. Moreover, the technique is data efficient: while approaches like RLHF require extensive annotations, ITI locates truthful directions using only few hundred examples. Our findings suggest that LLMs may have an internal representation of the likelihood of something being true, even as they produce falsehoods on the surface.

Modern large language models (LLMs) often encounter communication bottlenecks on current hardware, rather than purely computational constraints. Multi-head Latent Attention (MLA) tackles this challenge by using low-rank matrices in the key-value (KV) layers, thereby allowing compressed latent KV states to be cached. This approach significantly reduces the KV cache size relative to traditional multi-head attention, leading to faster inference. Moreover, MLA employs an up-projection matrix to increase expressiveness, trading additional computation for reduced communication overhead. Although MLA has demonstrated efficiency and effectiveness in Deepseek V2/V3/R1, many major model providers still rely on Group Query Attention (GQA) and have not announced any plans to adopt MLA. In this paper, we show that GQA can always be represented by MLA while maintaining the same KV cache overhead, but the converse does not hold. To encourage broader use of MLA, we introduce **TransMLA**, a post-training method that converts widely used GQA-based pre-trained models (e.g., LLaMA, Qwen, Mixtral) into MLA-based models. After conversion, the model can undergo additional training to boost expressiveness without increasing the KV cache size. Furthermore, we plan to develop MLA-specific inference acceleration techniques to preserve low latency in transformed models, thus enabling more efficient distillation of Deepseek R1.

Speculative decoding (SD) accelerates large language model inference by using a smaller draft model to predict multiple tokens, which are then verified in parallel by the larger target model. However, the limited capacity of the draft model often necessitates tree-based sampling to improve prediction accuracy, where multiple candidates are generated at each step. We identify a key limitation in this approach: the candidates at the same step are derived from the same representation, limiting diversity and reducing overall effectiveness. To address this, we propose Jakiro, leveraging Mixture of Experts (MoE), where independent experts generate diverse predictions, effectively decoupling correlations among candidates. Furthermore, we introduce a hybrid inference strategy, combining autoregressive decoding for initial tokens with parallel decoding for subsequent stages, and enhance the latter with contrastive mechanism in features to improve accuracy. Our method significantly boosts prediction accuracy and achieves higher inference speedups. Extensive experiments across diverse models validate the effectiveness and robustness of our approach, establishing a new SOTA in speculative decoding. Our codes are available at https://github.com/haiduo/Jakiro.

By equipping the most recent 3D Gaussian Splatting representation with head 3D morphable models (3DMM), existing methods manage to create head avatars with high fidelity. However, most existing methods only reconstruct a head without the body, substantially limiting their application scenarios. We found that naively applying Gaussians to model the clothed chest and shoulders tends to result in blurry reconstruction and noisy floaters under novel poses. This is because of the fundamental limitation of Gaussians and point clouds -- each Gaussian or point can only have a single directional radiance without spatial variance, therefore an unnecessarily large number of them is required to represent complicated spatially varying texture, even for simple geometry. In contrast, we propose to model the body part with a neural texture that consists of coarse and pose-dependent fine colors. To properly render the body texture for each view and pose without accurate geometry nor UV mapping, we optimize another sparse set of Gaussians as anchors that constrain the neural warping field that maps image plane coordinates to the texture space. We demonstrate that Gaussian Head & Shoulders can fit the high-frequency details on the clothed upper body with high fidelity and potentially improve the accuracy and fidelity of the head region. We evaluate our method with casual phone-captured and internet videos and show our method archives superior reconstruction quality and robustness in both self and cross reenactment tasks. To fully utilize the efficient rendering speed of Gaussian splatting, we additionally propose an accelerated inference method of our trained model without Multi-Layer Perceptron (MLP) queries and reach a stable rendering speed of around 130 FPS for any subjects.

We introduce GraviBERT, a novel deep learning framework for gravitational wave inference, built on a multi-scale feature extractor with a transformer encoder and a suitable regression head. A key novelty of GraviBERT is its staged training: a BERT-style self-supervised pretraining phase to learn transferable representations, followed by supervised fine-tuning on labeled data. GraviBERT demonstrates consistent transfer learning across detector configurations and waveform models. On in-domain data, pretraining reduces the MAE by up to 31% and accelerates convergence by sim 6.6 times, with mean relative precision for point estimates reaching the few-percent level and MAE in effective spin of sim 10^{-3} at SNR = 10. For domain adaptation to new detector noise profiles, the pretrained model converges up to 15times faster on small target datasets and reduces estimation errors by up to sim 47%, demonstrating detector-agnostic learning. Cross-waveform approximant transfer achieves up to 44% MAE reductions and up to 15times training speedups, with R^2 scores consistently exceeding 0.9 for mass parameters at SNR = 10 compared to 0.74 - 0.87 when training from scratch. GraviBERT works directly with noisy waveforms, and in its current form quantifies predictive uncertainty through MC dropouts. After pretraining, the regression head could be adapted to multiple downstream inference tasks in gravitational-wave astronomy.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across a range of multimodal tasks. However, their inference efficiency is constrained by the large number of visual tokens processed during decoding. To address this challenge, we propose Per-Layer Per-Head Vision Token Pruning (PLPHP), a two-level fine-grained pruning method including Layer-Level Retention Rate Allocation and Head-Level Vision Token Pruning. Motivated by the Vision Token Re-attention phenomenon across decoder layers, we dynamically adjust token retention rates layer by layer. Layers that exhibit stronger attention to visual information preserve more vision tokens, while layers with lower vision attention are aggressively pruned. Furthermore, PLPHP applies pruning at the attention head level, enabling different heads within the same layer to independently retain critical context. Experiments on multiple benchmarks demonstrate that PLPHP delivers an 18% faster decoding speed and reduces the Key-Value Cache (KV Cache) size by over 50%, all at the cost of 0.46% average performance drop, while also achieving notable performance improvements in multi-image tasks. These results highlight the effectiveness of fine-grained token pruning and contribute to advancing the efficiency and scalability of LVLMs. Our source code will be made publicly available.

Large language models (LLMs) now support extremely long context windows, but the quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. Existing approaches to address this complexity require additional pretraining or finetuning, and often sacrifice model accuracy. In this paper, we first provide both theoretical and empirical foundations for near-lossless sparse attention. We find dynamically capturing head-specific sparse patterns at runtime with low overhead is crucial. To address this, we propose SampleAttention, an adaptive structured and near-lossless sparse attention. Leveraging observed significant sparse patterns, SampleAttention attends to a fixed percentage of adjacent tokens to capture local window patterns, and employs a two-stage query-guided key-value filtering approach, which adaptively select a minimum set of key-values with low overhead, to capture column stripe patterns. Comprehensive evaluations show that SampleAttention can seamlessly replace vanilla attention in off-the-shelf LLMs with nearly no accuracy loss, and reduces TTFT by up to 2.42times compared with FlashAttention.

3D head animation has seen major quality and runtime improvements over the last few years, particularly empowered by the advances in differentiable rendering and neural radiance fields. Real-time rendering is a highly desirable goal for real-world applications. We propose HeadGaS, a model that uses 3D Gaussian Splats (3DGS) for 3D head reconstruction and animation. In this paper we introduce a hybrid model that extends the explicit 3DGS representation with a base of learnable latent features, which can be linearly blended with low-dimensional parameters from parametric head models to obtain expression-dependent color and opacity values. We demonstrate that HeadGaS delivers state-of-the-art results in real-time inference frame rates, surpassing baselines by up to 2dB, while accelerating rendering speed by over x10.

DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.

To combat the memory bandwidth-bound nature of autoregressive LLM inference, previous research has proposed the speculative decoding framework. To perform speculative decoding, a small draft model proposes candidate continuations of the input sequence, that are then verified in parallel by the base model. One way to specify the draft model, as used in the recent Medusa decoding framework, is as a collection of light-weight heads, called draft heads, that operate on the base model's hidden states. To date, all existing draft heads have been sequentially independent, meaning that they speculate tokens in the candidate continuation independently of any preceding tokens in the candidate continuation. In this work, we propose Hydra heads, a sequentially dependent, drop-in replacement for standard draft heads that significantly improves speculation accuracy. Decoding with Hydra heads improves throughput compared to Medusa decoding with standard draft heads. We further explore the design space of Hydra head training objectives and architectures, and propose a carefully-tuned Hydra head recipe, which we call Hydra++, that improves decoding throughput by 1.31x and 2.71x compared to Medusa decoding and autoregressive decoding, respectively. Overall, Hydra heads are a simple intervention on standard draft heads that significantly improve the end-to-end speed of draft head based speculative decoding.

Practitioners have consistently observed three puzzling phenomena in transformer-based large language models (LLMs): attention sinks, value-state drains, and residual-state peaks, collectively referred to as extreme-token phenomena. These phenomena are characterized by certain so-called "sink tokens" receiving disproportionately high attention weights, exhibiting significantly smaller value states, and having much larger residual-state norms than those of other tokens. These extreme tokens give rise to various challenges in LLM inference, quantization, and interpretability. We elucidate the mechanisms behind extreme-token phenomena. First, we show that these phenomena arise in very simple architectures -- transformers with one to three layers -- trained on a toy model, the Bigram-Backcopy (BB) task. In this setting, we identify an active-dormant mechanism, where attention heads become sinks for specific input domains while remaining non-sinks for others. Our theoretical analysis of the training dynamics reveals that these phenomena are driven by a mutual reinforcement mechanism. Building on these insights, we propose strategies to mitigate extreme-token phenomena during pretraining, including replacing softmax with ReLU and Adam with SGD. Next, we extend our analysis to pretrained LLMs, including Llama and OLMo, showing that many attention heads exhibit a similar active-dormant mechanism as in the BB task, and that the mutual reinforcement mechanism also governs the emergence of extreme-token phenomena during LLM pretraining. Our results reveal that many of the static and dynamic properties of extreme-token phenomena predicted by the BB task align with observations in pretrained LLMs.

In-context Learning (ICL) is an emerging few-shot learning paradigm on Language Models (LMs) with inner mechanisms un-explored. There are already existing works describing the inner processing of ICL, while they struggle to capture all the inference phenomena in large language models. Therefore, this paper proposes a comprehensive circuit to model the inference dynamics and try to explain the observed phenomena of ICL. In detail, we divide ICL inference into 3 major operations: (1) Summarize: LMs encode every input text (demonstrations and queries) into linear representation in the hidden states with sufficient information to solve ICL tasks. (2) Semantics Merge: LMs merge the encoded representations of demonstrations with their corresponding label tokens to produce joint representations of labels and demonstrations. (3) Feature Retrieval and Copy: LMs search the joint representations similar to the query representation on a task subspace, and copy the searched representations into the query. Then, language model heads capture these copied label representations to a certain extent and decode them into predicted labels. The proposed inference circuit successfully captured many phenomena observed during the ICL process, making it a comprehensive and practical explanation of the ICL inference process. Moreover, ablation analysis by disabling the proposed steps seriously damages the ICL performance, suggesting the proposed inference circuit is a dominating mechanism. Additionally, we confirm and list some bypass mechanisms that solve ICL tasks in parallel with the proposed circuit.

Accurate detection and classification of nuclei in histopathology images are critical for diagnostic and research applications. We present KongNet, a multi-headed deep learning architecture featuring a shared encoder and parallel, cell-type-specialised decoders. Through multi-task learning, each decoder jointly predicts nuclei centroids, segmentation masks, and contours, aided by Spatial and Channel Squeeze-and-Excitation (SCSE) attention modules and a composite loss function. We validate KongNet in three Grand Challenges. The proposed model achieved first place on track 1 and second place on track 2 during the MONKEY Challenge. Its lightweight variant (KongNet-Det) secured first place in the 2025 MIDOG Challenge. KongNet pre-trained on the MONKEY dataset and fine-tuned on the PUMA dataset ranked among the top three in the PUMA Challenge without further optimisation. Furthermore, KongNet established state-of-the-art performance on the publicly available PanNuke and CoNIC datasets. Our results demonstrate that the specialised multi-decoder design is highly effective for nuclei detection and classification across diverse tissue and stain types. The pre-trained model weights along with the inference code have been publicly released to support future research.

Deep learning drives a new wave in computing systems and triggers the automation of increasingly complex problems. In particular, Large Language Models (LLMs) have significantly advanced cognitive tasks, often matching or even surpassing human-level performance. However, their extensive parameters result in high computational costs and slow inference, posing challenges for deployment in resource-limited settings. Among the strategies to overcome the aforementioned challenges, pruning emerges as a successful mechanism since it reduces model size while maintaining predictive ability. In this paper, we introduce AMP: Attention Heads and MLP Pruning, a novel structured pruning method that efficiently compresses LLMs by removing less critical structures within Multi-Head Attention (MHA) and Multilayer Perceptron (MLP). By projecting the input data onto weights, AMP assesses structural importance and overcomes the limitations of existing techniques, which often fall short in flexibility or efficiency. In particular, AMP surpasses the current state-of-the-art on commonsense reasoning tasks by up to 1.49 percentage points, achieving a 30% pruning ratio with minimal impact on zero-shot task performance. Moreover, AMP also improves inference speeds, making it well-suited for deployment in resource-constrained environments. We confirm the flexibility of AMP on different families of LLMs, including LLaMA and Phi.

Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than 5%) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers 1.38x real-time acceleration and 52% memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8timesA800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to 2.8times and 2.0times in the prefill and decode stages.

While Transformer self-attention offers strong parallelism, the Key-Value (KV) cache grows linearly with sequence length and becomes a bottleneck for inference efficiency. Multi-head latent attention was recently developed to compress the KV cache into a low-rank latent space. This paper proposes Multi-head Temporal Latent Attention (MTLA), which further reduces the KV cache size along the temporal dimension, greatly lowering the memory footprint of self-attention inference. MTLA employs a hyper-network to dynamically merge temporally adjacent KV cache vectors. To address the mismatch between the compressed KV cache and processed sequence lengths, a stride-aware causal mask is proposed to ensure efficient parallel training and consistency with inference behaviour. Experiments across tasks, including speech translation, speech recognition, speech understanding and text summarisation, demonstrate that MTLA achieves competitive performance compared to standard Multi-Head Attention (MHA), while greatly improving inference speed and GPU memory usage. For example, on a English-German speech translation task, MTLA achieves a 5.3x speedup and a reduction in GPU memory usage by a factor of 8.3 compared to MHA, while maintaining translation quality.

As vision-language models (VLMs) tackle increasingly complex and multimodal tasks, the rapid growth of Key-Value (KV) cache imposes significant memory and computational bottlenecks during inference. While Multi-Head Latent Attention (MLA) offers an effective means to compress the KV cache and accelerate inference, adapting existing VLMs to the MLA architecture without costly pretraining remains largely unexplored. In this work, we present MHA2MLA-VLM, a parameter-efficient and multimodal-aware framework for converting off-the-shelf VLMs to MLA. Our approach features two core techniques: (1) a modality-adaptive partial-RoPE strategy that supports both traditional and multimodal settings by selectively masking nonessential dimensions, and (2) a modality-decoupled low-rank approximation method that independently compresses the visual and textual KV spaces. Furthermore, we introduce parameter-efficient fine-tuning to minimize adaptation cost and demonstrate that minimizing output activation error, rather than parameter distance, substantially reduces performance loss. Extensive experiments on three representative VLMs show that MHA2MLA-VLM restores original model performance with minimal supervised data, significantly reduces KV cache footprint, and integrates seamlessly with KV quantization.

We introduce HEAD-QA v2, an expanded and updated version of a Spanish/English healthcare multiple-choice reasoning dataset originally released by Vilares and Gómez-Rodríguez (2019). The update responds to the growing need for high-quality datasets that capture the linguistic and conceptual complexity of healthcare reasoning. We extend the dataset to over 12,000 questions from ten years of Spanish professional exams, benchmark several open-source LLMs using prompting, RAG, and probability-based answer selection, and provide additional multilingual versions to support future work. Results indicate that performance is mainly driven by model scale and intrinsic reasoning ability, with complex inference strategies obtaining limited gains. Together, these results establish HEAD-QA v2 as a reliable resource for advancing research on biomedical reasoning and model improvement.

Recent advancements in video diffusion models have significantly enhanced audio-driven portrait animation. However, current methods still suffer from flickering, identity drift, and poor audio-visual synchronization. These issues primarily stem from entangled appearance-motion representations and unstable inference strategies. In this paper, we introduce ConsistTalk, a novel intensity-controllable and temporally consistent talking head generation framework with diffusion noise search inference. First, we propose an optical flow-guided temporal module (OFT) that decouples motion features from static appearance by leveraging facial optical flow, thereby reducing visual flicker and improving temporal consistency. Second, we present an Audio-to-Intensity (A2I) model obtained through multimodal teacher-student knowledge distillation. By transforming audio and facial velocity features into a frame-wise intensity sequence, the A2I model enables joint modeling of audio and visual motion, resulting in more natural dynamics. This further enables fine-grained, frame-wise control of motion dynamics while maintaining tight audio-visual synchronization. Third, we introduce a diffusion noise initialization strategy (IC-Init). By enforcing explicit constraints on background coherence and motion continuity during inference-time noise search, we achieve better identity preservation and refine motion dynamics compared to the current autoregressive strategy. Extensive experiments demonstrate that ConsistTalk significantly outperforms prior methods in reducing flicker, preserving identity, and delivering temporally stable, high-fidelity talking head videos.

Diffusion-based talking head models generate high-quality, photorealistic videos but suffer from slow inference, limiting practical applications. Existing acceleration methods for general diffusion models fail to exploit the temporal and spatial redundancies unique to talking head generation. In this paper, we propose a task-specific framework addressing these inefficiencies through two key innovations. First, we introduce Lightning-fast Caching-based Parallel denoising prediction (LightningCP), caching static features to bypass most model layers in inference time. We also enable parallel prediction using cached features and estimated noisy latents as inputs, efficiently bypassing sequential sampling. Second, we propose Decoupled Foreground Attention (DFA) to further accelerate attention computations, exploiting the spatial decoupling in talking head videos to restrict attention to dynamic foreground regions. Additionally, we remove reference features in certain layers to bring extra speedup. Extensive experiments demonstrate that our framework significantly improves inference speed while preserving video quality.

As the rapid progression of practical applications based on Large Language Models continues, the importance of extrapolating performance has grown exponentially in the research domain. In our study, we identified an anomalous behavior in Transformer models that had been previously overlooked, leading to a chaos around closest tokens which carried the most important information. We've coined this discovery the "headache of Transformers". To address this at its core, we introduced a novel self-attention structure named Collinear Constrained Attention (CoCA). This structure can be seamlessly integrated with existing extrapolation, interpolation methods, and other optimization strategies designed for traditional Transformer models. We have achieved excellent extrapolating performance even for 16 times to 24 times of sequence lengths during inference without any fine-tuning on our model. We have also enhanced CoCA's computational and spatial efficiency to ensure its practicality. We plan to open-source CoCA shortly. In the meantime, we've made our code available in the appendix for reappearing experiments.

Attention heads are one of the building blocks of large language models (LLMs). Prior work on investigating their operation mostly focused on analyzing their behavior during inference for specific circuits or tasks. In this work, we seek a comprehensive mapping of the operations they implement in a model. We propose MAPS (Mapping Attention head ParameterS), an efficient framework that infers the functionality of attention heads from their parameters, without any model training or inference. We showcase the utility of MAPS for answering two types of questions: (a) given a predefined operation, mapping how strongly heads across the model implement it, and (b) given an attention head, inferring its salient functionality. Evaluating MAPS on 20 operations across 6 popular LLMs shows its estimations correlate with the head's outputs during inference and are causally linked to the model's predictions. Moreover, its mappings reveal attention heads of certain operations that were overlooked in previous studies, and valuable insights on function universality and architecture biases in LLMs. Next, we present an automatic pipeline and analysis that leverage MAPS to characterize the salient operations of a given head. Our pipeline produces plausible operation descriptions for most heads, as assessed by human judgment, while revealing diverse operations.

Mixture of experts has emerged as the primary mechanism for making Large Language Models (LLMs) computationally efficient. However, in distributed settings, communicating token embeddings between experts is a significant bottleneck. We present the novel Federation of Experts (FoE) architecture. FoE restructures the MoE block of a transformer layer into multiple MoE clusters. Each cluster is responsible for only one of the KV heads and expert parallelism is applied between those experts. Between clusters, a sum synchronizes the post-attention residuals, which then drives routing and dispatch for the next MoE block. In a single-node setting, FoE completely eliminates all-to-all communication as all experts within a group are contained on the same GPU. In multi-node settings, FoE confines all-to-all communication to the intra-node fabric, thus significantly reducing communication overhead. An implementation of FoE finds that on LongBench, FoE significantly improves inference throughput and latency in both single-node and multi-node settings, reducing the end-to-end forward-pass latency by up to 5.2x, TTFT by 3.62x, and TBT by 1.95x. It does so while achieving comparable generation quality to a mixture of experts model of the same size and training configuration.

Despite the remarkable strides made by autoregressive language models, their potential is often hampered by the slow inference speeds inherent in sequential token generation. Blockwise parallel decoding (BPD) was proposed by Stern et al. (2018) as a way to improve inference speed of language models. In this paper, we make two contributions to understanding and improving BPD drafts. We first offer an analysis of the token distributions produced by the BPD prediction heads. Secondly, we use this analysis to inform algorithms to improve BPD inference speed by refining the BPD drafts using small n-gram or neural language models. We empirically show that these refined BPD drafts yield a higher average verified prefix length across tasks.

Answering multi-hop reasoning questions requires retrieving and synthesizing information from diverse sources. Large Language Models (LLMs) struggle to perform such reasoning consistently. Here we propose an approach to pinpoint and rectify multi-hop reasoning failures through targeted memory injections on LLM attention heads. First, we analyze the per-layer activations of GPT-2 models in response to single and multi-hop prompts. We then propose a mechanism that allows users to inject pertinent prompt-specific information, which we refer to as "memories," at critical LLM locations during inference. By thus enabling the LLM to incorporate additional relevant information during inference, we enhance the quality of multi-hop prompt completions. We show empirically that a simple, efficient, and targeted memory injection into a key attention layer can often increase the probability of the desired next token in multi-hop tasks, by up to 424%.

We present PIRATR, an end-to-end 3D object detection framework for robotic use cases in point clouds. Extending PI3DETR, our method streamlines parametric 3D object detection by jointly estimating multi-class 6-DoF poses and class-specific parametric attributes directly from occlusion-affected point cloud data. This formulation enables not only geometric localization but also the estimation of task-relevant properties for parametric objects, such as a gripper's opening, where the 3D model is adjusted according to simple, predefined rules. The architecture employs modular, class-specific heads, making it straightforward to extend to novel object types without re-designing the pipeline. We validate PIRATR on an automated forklift platform, focusing on three structurally and functionally diverse categories: crane grippers, loading platforms, and pallets. Trained entirely in a synthetic environment, PIRATR generalizes effectively to real outdoor LiDAR scans, achieving a detection mAP of 0.919 without additional fine-tuning. PIRATR establishes a new paradigm of pose-aware, parameterized perception. This bridges the gap between low-level geometric reasoning and actionable world models, paving the way for scalable, simulation-trained perception systems that can be deployed in dynamic robotic environments. Code available at https://github.com/swingaxe/piratr.

Vision Transformers deliver state-of-the-art performance, yet their fixed computational budget prevents scalable deployment across heterogeneous hardware. Recent nested Transformer architectures mitigate this by embedding nested subnetworks within a single model to enable scalable inference. However, these models allocate the same amount of compute to all inputs, regardless of their complexity, which leads to inefficiencies. To address this, we introduce ThinkingViT, a nested ViT architecture that employs progressive thinking stages to dynamically adjust inference computation based on input difficulty. ThinkingViT initiates inference by activating a small subset of the most important attention heads and terminates early if predictions reach sufficient certainty. Otherwise, it activates additional attention heads and re-evaluates the input. At the core of ThinkingViT is our Token Recycling mechanism, which conditions each subsequent inference stage on the embeddings from the previous stage, enabling progressive improvement. Due to its backbone-preserving design, ThinkingViT also serves as a plugin upgrade for vanilla ViT. Experiments show that ThinkingViT surpasses nested baselines by up to 2.0 percentage points (p.p.) in accuracy at the same throughput and by up to 2.9 p.p. at equal GMACs on ImageNet-1K. The source code is available at https://github.com/ds-kiel/ThinkingViT.

Large Language Models (LLMs) exhibit high inference latency due to their autoregressive decoding nature. While the draft head in speculative decoding mitigates this issue, its full potential remains unexplored. In this paper, we introduce KOALA (K-layer Optimized Adversarial Learning Architecture), an orthogonal approach to the draft head. By transforming the conventional single-layer draft head into a multi-layer architecture and incorporating adversarial learning into the traditional supervised training, KOALA significantly improves the accuracy of the draft head in predicting subsequent tokens, thus more closely mirroring the functionality of LLMs. Although this improvement comes at the cost of slightly increased drafting overhead, KOALA substantially unlocks the draft head's potential, greatly enhancing speculative decoding. We conducted comprehensive evaluations of KOALA, including both autoregressive and non-autoregressive draft heads across various tasks, demonstrating a latency speedup ratio improvement of 0.24x-0.41x, which is 10.57%-14.09% faster than the original draft heads.

In this work, we upgrade the multi-head attention mechanism, the core of the Transformer model, to improve efficiency while maintaining or surpassing the previous accuracy level. We show that multi-head attention can be expressed in the summation form. Drawing on the insight that not all attention heads hold equal significance, we propose Mixture-of-Head attention (MoH), a new architecture that treats attention heads as experts in the Mixture-of-Experts (MoE) mechanism. MoH has two significant advantages: First, MoH enables each token to select the appropriate attention heads, enhancing inference efficiency without compromising accuracy or increasing the number of parameters. Second, MoH replaces the standard summation in multi-head attention with a weighted summation, introducing flexibility to the attention mechanism and unlocking extra performance potential. Extensive experiments on ViT, DiT, and LLMs demonstrate that MoH outperforms multi-head attention by using only 50%-90% of the attention heads. Moreover, we demonstrate that pre-trained multi-head attention models, such as LLaMA3-8B, can be further continue-tuned into our MoH models. Notably, MoH-LLaMA3-8B achieves an average accuracy of 64.0% across 14 benchmarks, outperforming LLaMA3-8B by 2.4% by utilizing only 75% of the attention heads. We believe the proposed MoH is a promising alternative to multi-head attention and provides a strong foundation for developing advanced and efficient attention-based models.

Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

The quadratic computational complexity of the attention mechanism in current Large Language Models (LLMs) renders inference with long contexts prohibitively expensive. To address this challenge, various approaches aim to retain critical portions of the context to optimally approximate Full Attention (FA) through Key-Value (KV) compression or Sparse Attention (SA), enabling the processing of virtually unlimited text lengths in a streaming manner. However, these methods struggle to achieve performance levels comparable to FA, particularly in retrieval tasks. In this paper, our analysis of attention head patterns reveals that LLMs' attention distributions show strong local correlations, naturally reflecting a chunking mechanism for input context. We propose Ltri-LLM framework, which divides KVs into spans, stores them in an offline index, and retrieves the relevant KVs into memory for various queries. Experimental results on popular long text benchmarks show that Ltri-LLM can achieve performance close to FA while maintaining efficient, streaming-based inference.

Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold. FlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference.

The recent surge in large-scale foundation models has spurred the development of efficient methods for adapting these models to various downstream tasks. Low-rank adaptation methods, such as LoRA, have gained significant attention due to their outstanding parameter efficiency and no additional inference latency. This paper investigates a more general form of adapter module based on the analysis that parallel and sequential adaptation branches learn novel and general features during fine-tuning, respectively. The proposed method, named Hydra, due to its multi-head computational branches, combines parallel and sequential branch to integrate capabilities, which is more expressive than existing single branch methods and enables the exploration of a broader range of optimal points in the fine-tuning process. In addition, the proposed adaptation method explicitly leverages the pre-trained weights by performing a linear combination of the pre-trained features. It allows the learned features to have better generalization performance across diverse downstream tasks. Furthermore, we perform a comprehensive analysis of the characteristics of each adaptation branch with empirical evidence. Through an extensive range of experiments, encompassing comparisons and ablation studies, we substantiate the efficiency and demonstrate the superior performance of Hydra. This comprehensive evaluation underscores the potential impact and effectiveness of Hydra in a variety of applications. Our code is available on https://github.com/extremebird/Hydra

Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1times and 6.4times, respectively.

We present a cross-architecture evaluation of production LLM inference on AMD Instinct MI325X GPUs, benchmarking four models spanning 235B to 1 trillion parameters across three architectural families (MoE+MLA, Dense+GQA, MoE+GQA) on an 8-GPU cluster with 2TB aggregate HBM3e using vLLM v0.14.1. Our results demonstrate that architecture-aware optimization is essential: MLA models require block size 1 and cannot use KV cache offloading, while GQA models benefit from both. The AMD AITER runtime is required for competitive MLA inference throughput and must be selectively disabled for architectures with incompatible attention head configurations. A controlled AITER ablation on Llama-3.1-405B (n=5 per condition) reveals a modest 3-5% throughput benefit at high concurrency but 2-16x higher measurement variability, confirming that AITER's large speedups target MoE/MLA kernels specifically. Under text-only workloads, Llama-405B and DeepSeek V3.2 achieve comparable peak throughput (15,944 and 15,343 tok/s) despite an order-of-magnitude difference in active parameters. Under vision workloads, Qwen3-VL-235B reaches 47,873 tok/s, 6.5x higher than Kimi-K2.5 (7,327 tok/s). Active parameter count per token is associated with inference throughput, though confounded by differences in quantization, AITER acceleration, and tensor parallelism. All four models exhibit a common throughput saturation point consistent with a memory-bandwidth bottleneck (~500 concurrent for short sequences, ~100-200 for longer sequences). All models maintain 100% HTTP-level success rates through 1,000 concurrent users, processing 18.9 million tokens across 17,406 requests without failures.

Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG. We propose Kascade, a training-free sparse attention method that leverages known observations such as 1) post-softmax attention is intrinsically sparse, and 2) the identity of high-weight keys is stable across nearby layers. Kascade computes exact Top-k indices in a small set of anchor layers, then reuses those indices in intermediate reuse layers. The anchor layers are selected algorithmically, via a dynamic-programming objective that maximizes cross-layer similarity over a development set, allowing easy deployment across models. The method incorporates efficient implementation constraints (e.g. tile-level operations), across both prefill and decode attention. The Top-k selection and reuse in Kascade is head-aware and we show in our experiments that this is critical for high accuracy. Kascade achieves up to 4.1x speedup in decode attention and 2.2x speedup in prefill attention over FlashAttention-3 baseline on H100 GPUs while closely matching dense attention accuracy on long-context benchmarks such as LongBench and AIME-24.

Large language models (LLMs) exhibit remarkable reasoning capabilities across diverse downstream tasks. However, their autoregressive nature leads to substantial inference latency, posing challenges for real-time applications. Speculative sampling mitigates this issue by introducing a drafting phase followed by a parallel validation phase, enabling faster token generation and verification. Existing approaches, however, overlook the inherent coherence in text generation, limiting their efficiency. To address this gap, we propose a Speculative Sampling with Syntactic and Semantic Coherence (S^4C) framework, which extends speculative sampling by leveraging multi-head drafting for rapid token generation and a continuous verification tree for efficient candidate validation and feature reuse. Experimental results demonstrate that S^4C surpasses baseline methods across mainstream tasks, offering enhanced efficiency, parallelism, and the ability to generate more valid tokens with fewer computational resources. On Spec-bench benchmarks, S^4C achieves an acceleration ratio of 2.26x-2.60x, outperforming state-of-the-art methods.

Large language models have revolutionized natural language processing but face significant challenges of high storage and runtime costs, due to the transformer architecture's reliance on self-attention, particularly the large Key-Value (KV) cache for long-sequence inference. Recent efforts to reduce KV cache size by pruning less critical entries based on attention weights remain empirical and lack formal grounding. This paper presents a formal study on identifying critical KV cache entries by analyzing attention output perturbation. Our analysis reveals that, beyond attention weights, the value states within KV entries and pretrained parameter matrices are also crucial. Based on this, we propose a perturbation-constrained selection algorithm that optimizes the worst-case output perturbation to identify critical entries. Evaluations on the Needle-in-a-Haystack test and Longbench benchmark show our algorithm enhances state-of-the-art cache eviction methods. Further empirical analysis confirms that our algorithm achieves lower output perturbations in over 92% attention heads in Llama model, thereby providing a significant improvement over existing methods.

Talking head generation with arbitrary identities and speech audio remains a crucial problem in the realm of the virtual metaverse. Recently, diffusion models have become a popular generative technique in this field with their strong generation capabilities. However, several challenges remain for diffusion-based methods: 1) inefficient inference and visual artifacts caused by the implicit latent space of Variational Auto-Encoders (VAE), which complicates the diffusion process; 2) a lack of authentic facial expressions and head movements due to inadequate multi-modal information fusion. In this paper, MoDA handles these challenges by: 1) defining a joint parameter space that bridges motion generation and neural rendering, and leveraging flow matching to simplify diffusion learning; 2) introducing a multi-modal diffusion architecture to model the interaction among noisy motion, audio, and auxiliary conditions, enhancing overall facial expressiveness. In addition, a coarse-to-fine fusion strategy is employed to progressively integrate different modalities, ensuring effective feature fusion. Experimental results demonstrate that MoDA improves video diversity, realism, and efficiency, making it suitable for real-world applications. Project Page: https://lixinyyang.github.io/MoDA.github.io/

The advent of pre-trained large language models (LLMs) has revolutionized various natural language processing tasks. These models predominantly employ an auto-regressive decoding mechanism that utilizes Key-Value (KV) caches to eliminate redundant calculations for previous tokens. Nevertheless, as context lengths and batch sizes increase, the linear expansion in memory footprint of KV caches becomes a key bottleneck of LLM deployment, which decreases generation speeds significantly. To mitigate this issue, previous techniques like multi-query attention (MQA) and grouped-query attention (GQA) have been developed, in order to reduce KV heads to accelerate inference with comparable accuracy to multi-head attention (MHA). Despite their effectiveness, existing strategies for compressing MHA often overlook the intrinsic properties of the KV caches. In this work, we explore the low-rank characteristics of the KV caches and propose a novel approach for compressing KV heads. In particular, we carefully optimize the MHA-to-GQA transformation to minimize compression error, and to remain compatible with rotary position embeddings (RoPE), we also introduce specialized strategies for key caches with RoPE. We demonstrate that our method can compress half or even three-quarters of KV heads while maintaining performance comparable to the original LLMs, which presents a promising direction for more efficient LLM deployment in resource-constrained environments.

Automated fetal head segmentation in ultrasound images is critical for accurate biometric measurements in prenatal care. While existing deep learning approaches have achieved a reasonable performance, they struggle with issues like low contrast, noise, and complex anatomical boundaries which are inherent to ultrasound imaging. This paper presents Attention-ResUNet. It is a novel architecture that synergistically combines residual learning with multi-scale attention mechanisms in order to achieve enhanced fetal head segmentation. Our approach integrates attention gates at four decoder levels to focus selectively on anatomically relevant regions while suppressing the background noise, and complemented by residual connections which facilitates gradient flow and feature reuse. Extensive evaluation on the HC18 Challenge dataset where n = 200 demonstrates that Attention ResUNet achieves a superior performance with a mean Dice score of 99.30 +/- 0.14% against similar architectures. It significantly outperforms five baseline architectures including ResUNet (99.26%), Attention U-Net (98.79%), Swin U-Net (98.60%), Standard U-Net (98.58%), and U-Net++ (97.46%). Through statistical analysis we confirm highly significant improvements (p < 0.001) with effect sizes that range from 0.230 to 13.159 (Cohen's d). Using Saliency map analysis, we reveal that our architecture produces highly concentrated, anatomically consistent activation patterns, which demonstrate an enhanced interpretability which is crucial for clinical deployment. The proposed method establishes a new state of the art performance for automated fetal head segmentation whilst maintaining computational efficiency with 14.7M parameters and a 45 GFLOPs inference cost. Code repository: https://github.com/Ammar-ss

Scaling inference for large language models (LLMs) is increasingly constrained by limited GPU memory, especially due to growing key-value (KV) caches required for long-context generation. While existing approaches offload KV caches to CPU memory or apply sparse attention to reduce GPU load, they often underutilize CPU compute resources and compromise accuracy. We present HGCA, a hybrid CPU-GPU attention mechanism that enables scalable, high-throughput LLM inference with near-full attention quality. HGCA performs dense attention on recently generated KV entries retained in GPU memory and parallel sparse attention on selected, salient KV entries in CPU memory. The attention outputs are efficiently merged using log-sum-exp fusion, minimizing PCIe transfer overhead. HGCA also introduces a finegrained, per-head sparsification strategy optimized for CPU execution, preserving contextual relevance while reducing computation. Our implementation seamlessly integrates into existing LLM frameworks without requiring model retraining. Experiments across diverse models and workloads show that HGCA achieves superior scalability, supports longer sequences and larger batch sizes, and outperforms existing sparse attention baselines in both performance and accuracy -- all on commodity GPU hardware.

Key-value (KV) caching has emerged as a crucial optimization technique for accelerating inference in large language models (LLMs). By allowing the attention operation to scale linearly rather than quadratically with the total sequence length, KV caching significantly enhances generation throughput. However, due to large context lengths in the modern LLMs, the memory footprint of the KV is a huge bottleneck for model deployment directly impacting the model's batch size, hindering its ability to deliver high-throughput. Existing research addresses this challenge using several techniques, such as discarding low-attention tokens, quantization, and matrix approximation which typically lead to a negative impact on the model accuracy. In this paper, We propose KVCrush technology which can be combined with many KV compression technologies to improve the model accuracy at a much smaller memory. KVCrush provides an alternate representation scheme for key-value states, along with a low-overhead token pruning algorithm that accounts for the token distribution in the KV cache, which in turn allows for a a smaller footprint while maintaining the accuracy of the model. Based on our results, KVCrush reduces LongBench KV Cache size by 4x with less than 1% accuracy drop and achieves state-of-the-art average accuracy with minimal overhead, incurring less than 0.5% total inference latency. KVCrush not only outperforms the accuracy of state-of-the-art importance-based token retention schemes but is also compatible with typical practical LLM deployments using KV cache paging schemes such as vLLM and mixed precision quantization.

Multimodal large language models can exhibit text dominance, over-relying on linguistic priors instead of grounding predictions in non-text inputs. One example is large audio-language models (LALMs) where decisive audio evidence can be under-utilized even when it contains important information. To address this issue we use mechanistic interpretability to identify a small set of audio-specialist attention heads whose audio attention yields a ``listening'' signal. We show that this signal increases when audio evidence affects the model's output, providing an indicator of audio engagement under standard prompting. Leveraging this localization, we construct an audio--silence steering direction and apply an inference-time activation intervention to the final representation, amplifying the model's audio effect. To demonstrate the utility of this intervention, we show on MMAU that this improves accuracy by up to +8.0 percentage points on two Qwen-based LALMs, without any parameter updates.

Achieving a balance between high-fidelity visual quality and low-latency streaming remains a formidable challenge in audio-driven portrait generation. Existing large-scale models often suffer from prohibitive computational costs, while lightweight alternatives typically compromise on holistic facial representations and temporal stability. In this paper, we propose SoulX-FlashHead, a unified 1.3B-parameter framework designed for real-time, infinite-length, and high-fidelity streaming video generation. To address the instability of audio features in streaming scenarios, we introduce Streaming-Aware Spatiotemporal Pre-training equipped with a Temporal Audio Context Cache mechanism, which ensures robust feature extraction from short audio fragments. Furthermore, to mitigate the error accumulation and identity drift inherent in long-sequence autoregressive generation, we propose Oracle-Guided Bidirectional Distillation, leveraging ground-truth motion priors to provide precise physical guidance. We also present VividHead, a large-scale, high-quality dataset containing 782 hours of strictly aligned footage to support robust training. Extensive experiments demonstrate that SoulX-FlashHead achieves state-of-the-art performance on HDTF and VFHQ benchmarks. Notably, our Lite variant achieves an inference speed of 96 FPS on a single NVIDIA RTX 4090, facilitating ultra-fast interaction without sacrificing visual coherence.

Transformers have emerged as the backbone of large language models (LLMs). However, generation remains inefficient due to the need to store in memory a cache of key-value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solution, we propose Dynamic Memory Compression (DMC), a method for on-line key-value cache compression at inference time. Most importantly, the model learns to apply different compression rates in different heads and layers. We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achieving up to ~3.7x throughput increase in auto-regressive inference on a NVIDIA H100 GPU. DMC is applied via continued pre-training on a negligible percentage of the original data without adding any extra parameters. We find that DMC preserves the original downstream performance with up to 4x cache compression, outperforming up-trained grouped-query attention (GQA). GQA and DMC can be even combined to obtain compounded gains. As a result DMC fits longer contexts and larger batches within any given memory budget.

Large language models (LLMs) have demonstrated remarkable capabilities across various tasks. However, their widespread application is hindered by the resource-intensive decoding process. To address this challenge, current approaches have incorporated additional decoding heads to enable parallel prediction of multiple subsequent tokens, thereby achieving inference acceleration. Nevertheless, the accuracy of these decoding heads falls short of the auto-regressive decoding approach. In light of these limitations, we propose Chimera, a novel framework specifically designed for speculative sampling. Within this framework, we introduce a lightweight draft model that effectively utilizes previously generated tokens to predict subsequent words. To ensure both accuracy and efficiency, we present two strategies within the lightweight draft model. Firstly, we focus on capturing short-range dependencies at the bottom layer. Secondly, we leverage the readily available representations from the original LLM.Through empirical evaluation on the Vicuna and LlaMA-2 series, Chimera demonstrates impressive results, achieving an average latency speedup ratio of 2.7x compared to the vanilla auto-regressive decoding approach. This highlights the potential of our proposed framework in significantly improving the efficiency of large language models during the decoding process.

Semantic representation and inference is essential for Natural Language Processing (NLP). The state of the art for semantic representation and inference is deep learning, and particularly Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and transformer Self-Attention models. This thesis investigates the use of deep learning for novel semantic representation and inference, and makes contributions in the following three areas: creating training data, improving semantic representations and extending inference learning. In terms of creating training data, we contribute the largest publicly available dataset of real-life factual claims for the purpose of automatic claim verification (MultiFC), and we present a novel inference model composed of multi-scale CNNs with different kernel sizes that learn from external sources to infer fact checking labels. In terms of improving semantic representations, we contribute a novel model that captures non-compositional semantic indicators. By definition, the meaning of a non-compositional phrase cannot be inferred from the individual meanings of its composing words (e.g., hot dog). Motivated by this, we operationalize the compositionality of a phrase contextually by enriching the phrase representation with external word embeddings and knowledge graphs. Finally, in terms of inference learning, we propose a series of novel deep learning architectures that improve inference by using syntactic dependencies, by ensembling role guided attention heads, incorporating gating layers, and concatenating multiple heads in novel and effective ways. This thesis consists of seven publications (five published and two under review).

The quadratic complexity of attention remains the central bottleneck in long-context inference for large language models. Prior acceleration methods either sparsify the attention map with structured patterns or permanently evict tokens at specific layers, which can retain irrelevant tokens or rely on irreversible early decisions despite the layer-/head-wise dynamics of token importance. In this paper, we propose Token Sparse Attention, a lightweight and dynamic token-level sparsification mechanism that compresses per-head Q, K, V to a reduced token set during attention and then decompresses the output back to the original sequence, enabling token information to be reconsidered in subsequent layers. Furthermore, Token Sparse Attention exposes a new design point at the intersection of token selection and sparse attention. Our approach is fully compatible with dense attention implementations, including Flash Attention, and can be seamlessly composed with existing sparse attention kernels. Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to times3.23 attention speedup at 128K context with less than 1% accuracy degradation. These results demonstrate that dynamic and interleaved token-level sparsification is a complementary and effective strategy for scalable long-context inference.

Recent years have witnessed remarkable advances in talking head generation, owing to its potential to revolutionize the human-AI interaction from text interfaces into realistic video chats. However, research on text-driven talking heads remains underexplored, with existing methods predominantly adopting a cascaded pipeline that combines TTS systems with audio-driven talking head models. This conventional pipeline not only introduces system complexity and latency overhead but also fundamentally suffers from asynchronous audiovisual output and stylistic discrepancies between generated speech and visual expressions. To address these limitations, we introduce OmniTalker, an end-to-end unified framework that simultaneously generates synchronized speech and talking head videos from text and reference video in real-time zero-shot scenarios, while preserving both speech style and facial styles. The framework employs a dual-branch diffusion transformer architecture: the audio branch synthesizes mel-spectrograms from text, while the visual branch predicts fine-grained head poses and facial dynamics. To bridge modalities, we introduce a novel audio-visual fusion module that integrates cross-modal information to ensure temporal synchronization and stylistic coherence between audio and visual outputs. Furthermore, our in-context reference learning module effectively captures both speech and facial style characteristics from a single reference video without introducing an extra style extracting module. To the best of our knowledge, OmniTalker presents the first unified framework that jointly models speech style and facial style in a zero-shot setting, achieving real-time inference speed of 25 FPS. Extensive experiments demonstrate that our method surpasses existing approaches in generation quality, particularly excelling in style preservation and audio-video synchronization.

End-to-end transformer-based detectors (DETRs) have shown exceptional performance in both closed-set and open-vocabulary object detection (OVD) tasks through the integration of language modalities. However, their demanding computational requirements have hindered their practical application in real-time object detection (OD) scenarios. In this paper, we scrutinize the limitations of two leading models in the OVDEval benchmark, OmDet and Grounding-DINO, and introduce OmDet-Turbo. This novel transformer-based real-time OVD model features an innovative Efficient Fusion Head (EFH) module designed to alleviate the bottlenecks observed in OmDet and Grounding-DINO. Notably, OmDet-Turbo-Base achieves a 100.2 frames per second (FPS) with TensorRT and language cache techniques applied. Notably, in zero-shot scenarios on COCO and LVIS datasets, OmDet-Turbo achieves performance levels nearly on par with current state-of-the-art supervised models. Furthermore, it establishes new state-of-the-art benchmarks on ODinW and OVDEval, boasting an AP of 30.1 and an NMS-AP of 26.86, respectively. The practicality of OmDet-Turbo in industrial applications is underscored by its exceptional performance on benchmark datasets and superior inference speed, positioning it as a compelling choice for real-time object detection tasks. Code: https://github.com/om-ai-lab/OmDet

Large Language Model (LLM) serving faces a fundamental tension between stringent latency Service Level Objectives (SLOs) and limited GPU memory capacity. When high request rates exhaust the KV cache budget, existing LLM inference systems often suffer severe head-of-line (HOL) blocking. While prior work explored PCIe-based offloading, these approaches cannot sustain responsiveness under high request rates, often failing to meet tight Time-To-First-Token (TTFT) and Time-Between-Tokens (TBT) SLOs. We present SuperInfer, a high-performance LLM inference system designed for emerging Superchips (e.g., NVIDIA GH200) with tightly coupled GPU-CPU architecture via NVLink-C2C. SuperInfer introduces RotaSched, the first proactive, SLO-aware rotary scheduler that rotates requests to maintain responsiveness on Superchips, and DuplexKV, an optimized rotation engine that enables full-duplex transfer over NVLink-C2C. Evaluations on GH200 using various models and datasets show that SuperInfer improves TTFT SLO attainment rates by up to 74.7% while maintaining comparable TBT and throughput compared to state-of-the-art systems, demonstrating that SLO-aware scheduling and memory co-design unlocks the full potential of Superchips for responsive LLM serving.

This paper presents EGSTalker, a real-time audio-driven talking head generation framework based on 3D Gaussian Splatting (3DGS). Designed to enhance both speed and visual fidelity, EGSTalker requires only 3-5 minutes of training video to synthesize high-quality facial animations. The framework comprises two key stages: static Gaussian initialization and audio-driven deformation. In the first stage, a multi-resolution hash triplane and a Kolmogorov-Arnold Network (KAN) are used to extract spatial features and construct a compact 3D Gaussian representation. In the second stage, we propose an Efficient Spatial-Audio Attention (ESAA) module to fuse audio and spatial cues, while KAN predicts the corresponding Gaussian deformations. Extensive experiments demonstrate that EGSTalker achieves rendering quality and lip-sync accuracy comparable to state-of-the-art methods, while significantly outperforming them in inference speed. These results highlight EGSTalker's potential for real-time multimedia applications.

Mixed-precision inference techniques reduce the memory and computational demands of Large Language Models (LLMs) by applying hybrid precision formats to model weights, activations, and KV caches. This work introduces mixed-precision LLM inference techniques that encompass (i) systematic memory and compute optimization across hierarchical storage and tensor core architectures, and (ii) comprehensive end-to-end mixed-precision optimization across diverse precision formats and hardware configurations. Our approach features two novel mixed-precision pipelines designed for optimal hardware utilization: a General Matrix Multiply (GEMM) pipeline that optimizes matrix operations through offline weight packing and online acceleration, and an attention pipeline that enables efficient attention computation with arbitrary Query, Key, and Value precision combinations. The key implementation of the pipelines includes (i) hardware-aware weight packing for automatic format optimization, (ii) adaptive head alignment for efficient attention computation, (iii) instruction-level parallelism for memory hierarchy exploitation, and (iv) KV memory loading pipeline for enhanced inference efficiency. We conduct comprehensive evaluations across 16 popular LLMs and 4 representative GPU architectures. Results demonstrate that our approach achieves up to 61% lower serving latency (30% on average) and up to 156% higher throughput (58% on average) in mixed-precision workloads compared to existing mixed-precision frameworks, establishing consistent performance improvements across all tested configurations and hardware types. This work is integrated into TurboMind, a high-performance inference engine of the LMDeploy project, which is open-sourced and publicly available at https://github.com/InternLM/lmdeploy.

We present LAM, an innovative Large Avatar Model for animatable Gaussian head reconstruction from a single image. Unlike previous methods that require extensive training on captured video sequences or rely on auxiliary neural networks for animation and rendering during inference, our approach generates Gaussian heads that are immediately animatable and renderable. Specifically, LAM creates an animatable Gaussian head in a single forward pass, enabling reenactment and rendering without additional networks or post-processing steps. This capability allows for seamless integration into existing rendering pipelines, ensuring real-time animation and rendering across a wide range of platforms, including mobile phones. The centerpiece of our framework is the canonical Gaussian attributes generator, which utilizes FLAME canonical points as queries. These points interact with multi-scale image features through a Transformer to accurately predict Gaussian attributes in the canonical space. The reconstructed canonical Gaussian avatar can then be animated utilizing standard linear blend skinning (LBS) with corrective blendshapes as the FLAME model did and rendered in real-time on various platforms. Our experimental results demonstrate that LAM outperforms state-of-the-art methods on existing benchmarks.

Audio-driven portrait animation has made significant advances with diffusion-based models, improving video quality and lipsync accuracy. However, the increasing complexity of these models has led to inefficiencies in training and inference, as well as constraints on video length and inter-frame continuity. In this paper, we propose JoyVASA, a diffusion-based method for generating facial dynamics and head motion in audio-driven facial animation. Specifically, in the first stage, we introduce a decoupled facial representation framework that separates dynamic facial expressions from static 3D facial representations. This decoupling allows the system to generate longer videos by combining any static 3D facial representation with dynamic motion sequences. Then, in the second stage, a diffusion transformer is trained to generate motion sequences directly from audio cues, independent of character identity. Finally, a generator trained in the first stage uses the 3D facial representation and the generated motion sequences as inputs to render high-quality animations. With the decoupled facial representation and the identity-independent motion generation process, JoyVASA extends beyond human portraits to animate animal faces seamlessly. The model is trained on a hybrid dataset of private Chinese and public English data, enabling multilingual support. Experimental results validate the effectiveness of our approach. Future work will focus on improving real-time performance and refining expression control, further expanding the applications in portrait animation. The code is available at: https://github.com/jdh-algo/JoyVASA.

Large transformer-based models have significant potential for speech transcription and translation. Their self-attention mechanisms and parallel processing enable them to capture complex patterns and dependencies in audio sequences. However, this potential comes with challenges, as these large and computationally intensive models lead to slow inference speeds. Various optimization strategies have been proposed to improve performance, including efficient hardware utilization and algorithmic enhancements. In this paper, we introduce Whisper-Medusa, a novel approach designed to enhance processing speed with minimal impact on Word Error Rate (WER). The proposed model extends the OpenAI's Whisper architecture by predicting multiple tokens per iteration, resulting in a 50% reduction in latency. We showcase the effectiveness of Whisper-Medusa across different learning setups and datasets.

Existing neural head avatars methods have achieved significant progress in the image quality and motion range of portrait animation. However, these methods neglect the computational overhead, and to the best of our knowledge, none is designed to run on mobile devices. This paper presents MobilePortrait, a lightweight one-shot neural head avatars method that reduces learning complexity by integrating external knowledge into both the motion modeling and image synthesis, enabling real-time inference on mobile devices. Specifically, we introduce a mixed representation of explicit and implicit keypoints for precise motion modeling and precomputed visual features for enhanced foreground and background synthesis. With these two key designs and using simple U-Nets as backbones, our method achieves state-of-the-art performance with less than one-tenth the computational demand. It has been validated to reach speeds of over 100 FPS on mobile devices and support both video and audio-driven inputs.

Large Language Models have excelled in various domains but face efficiency challenges due to the growing Key-Value (KV) cache required for long-sequence inference. Recent efforts aim to reduce KV cache size by evicting vast non-critical cache elements during runtime while preserving generation quality. However, these methods typically allocate compression budgets uniformly across all attention heads, ignoring the unique attention patterns of each head. In this paper, we establish a theoretical loss upper bound between pre- and post-eviction attention output, explaining the optimization target of prior cache eviction methods, while guiding the optimization of adaptive budget allocation. Base on this, we propose {\it Ada-KV}, the first head-wise adaptive budget allocation strategy. It offers plug-and-play benefits, enabling seamless integration with prior cache eviction methods. Extensive evaluations on 13 datasets from Ruler and 16 datasets from LongBench, all conducted under both question-aware and question-agnostic scenarios, demonstrate substantial quality improvements over existing methods. Our code is available at https://github.com/FFY0/AdaKV.

The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be needed for future queries. In this paper, we propose a novel compression technique for KV cache that preserves all token information. Our investigation reveals that: i) Most attention heads primarily focus on the local context; ii) Only a few heads, denoted as retrieval heads, can essentially pay attention to all input tokens. These key observations motivate us to use separate caching strategy for attention heads. Therefore, we propose RazorAttention, a training-free KV cache compression algorithm, which maintains a full cache for these crucial retrieval heads and discards the remote tokens in non-retrieval heads. Furthermore, we introduce a novel mechanism involving a "compensation token" to further recover the information in the dropped tokens. Extensive evaluations across a diverse set of large language models (LLMs) demonstrate that RazorAttention achieves a reduction in KV cache size by over 70% without noticeable impacts on performance. Additionally, RazorAttention is compatible with FlashAttention, rendering it an efficient and plug-and-play solution that enhances LLM inference efficiency without overhead or retraining of the original model.

Current auto-regressive models can generate high-quality, topologically precise meshes; however, they necessitate thousands-or even tens of thousands-of next-token predictions during inference, resulting in substantial latency. We introduce XSpecMesh, a quality-preserving acceleration method for auto-regressive mesh generation models. XSpecMesh employs a lightweight, multi-head speculative decoding scheme to predict multiple tokens in parallel within a single forward pass, thereby accelerating inference. We further propose a verification and resampling strategy: the backbone model verifies each predicted token and resamples any tokens that do not meet the quality criteria. In addition, we propose a distillation strategy that trains the lightweight decoding heads by distilling from the backbone model, encouraging their prediction distributions to align and improving the success rate of speculative predictions. Extensive experiments demonstrate that our method achieves a 1.7x speedup without sacrificing generation quality. Our code will be released.

The generation of high-fidelity, animatable 3D human avatars remains a core challenge in computer graphics and vision, with applications in VR, telepresence, and entertainment. Existing approaches based on implicit representations like NeRFs suffer from slow rendering and dynamic inconsistencies, while 3D Gaussian Splatting (3DGS) methods are typically limited to static head generation, lacking dynamic control. We bridge this gap by introducing AGORA, a novel framework that extends 3DGS within a generative adversarial network to produce animatable avatars. Our key contribution is a lightweight, FLAME-conditioned deformation branch that predicts per-Gaussian residuals, enabling identity-preserving, fine-grained expression control while allowing real-time inference. Expression fidelity is enforced via a dual-discriminator training scheme leveraging synthetic renderings of the parametric mesh. AGORA generates avatars that are not only visually realistic but also precisely controllable. Quantitatively, we outperform state-of-the-art NeRF-based methods on expression accuracy while rendering at 250+ FPS on a single GPU, and, notably, at sim9 FPS under CPU-only inference - representing, to our knowledge, the first demonstration of practical CPU-only animatable 3DGS avatar synthesis. This work represents a significant step toward practical, high-performance digital humans. Project website: https://ramazan793.github.io/AGORA/

Efficient long-context inference is critical as large language models (LLMs) adopt context windows of ranging from 128K to 1M tokens. However, the growing key-value (KV) cache and the high computational complexity of attention create significant bottlenecks in memory usage and latency. In this paper, we find that attention in diverse long-context tasks exhibits sparsity, and LLMs implicitly "know" which tokens can be dropped or evicted at the head level after the pre-filling stage. Based on this insight, we propose Self-Attention Guided Eviction~(SAGE-KV), a simple and effective KV eviction cache method for long-context inference. After prefilling, our method performs a one-time top-k selection at both the token and head levels to compress the KV cache, enabling efficient inference with the reduced cache. Evaluations on LongBench and three long-context LLMs (Llama3.1-8B-Instruct-128k, Llama3-8B-Prolong-512k-Instruct, and Qwen2.5-7B-Instruct-128k) show that SAGE-KV maintains accuracy comparable to full attention while significantly improving efficiency. Specifically, SAGE-KV achieves 4x higher memory efficiency with improved accuracy over the static KV cache selection method StreamLLM, and 2x higher memory efficiency with better accuracy than the dynamic KV cache selection method Quest.

The Transformer architecture, underpinned by the Multi-Head Attention (MHA) mechanism, has become the de facto standard for state-of-the-art models in artificial intelligence. However, the quadratic computational complexity of MHA with respect to sequence length presents a significant barrier to scaling, particularly for applications involving long contexts. Prevailing solutions, such as Multi-Query Attention (MQA) and Grouped-Query Attention (GQA), have effectively addressed the memory bandwidth bottleneck that dominates autoregressive inference latency by sharing Key and Value projections. While highly successful, these methods do not reduce the fundamental number of floating-point operations (FLOPs) required for the attention score computation, which remains a critical bottleneck for training and full-sequence processing. This paper introduces Sparse Query Attention (SQA), a novel attention architecture that pursues an alternative and complementary optimization path. Instead of reducing Key/Value heads, SQA reduces the number of Query heads. This architectural modification directly decreases the computational complexity of the attention mechanism by a factor proportional to the reduction in query heads, thereby lowering the overall FLOPs. This work presents the theoretical foundation of SQA, its mathematical formulation, and a family of architectural variants. Empirical benchmarks on long sequences (32k-200k tokens) demonstrate that SQA can achieve significant throughput improvements of up to 3x in computation-bound scenarios such as model pre-training, fine-tuning, and encoder-based tasks, with only a minimal impact on model quality in preliminary smallscale experiments. SQA was discovered serendipitously during the development of the upcoming Reactive Transformer architecture, suggesting its potential as a powerful tool for building more efficient and scalable models

Large language models (LLMs) with hundreds of billions of parameters have sparked a new wave of exciting AI applications. However, they are computationally expensive at inference time. Sparsity is a natural approach to reduce this cost, but existing methods either require costly retraining, have to forgo LLM's in-context learning ability, or do not yield wall-clock time speedup on modern hardware. We hypothesize that contextual sparsity, which are small, input-dependent sets of attention heads and MLP parameters that yield approximately the same output as the dense model for a given input, can address these issues. We show that contextual sparsity exists, that it can be accurately predicted, and that we can exploit it to speed up LLM inference in wall-clock time without compromising LLM's quality or in-context learning ability. Based on these insights, we propose DejaVu, a system that uses a low-cost algorithm to predict contextual sparsity on the fly given inputs to each layer, along with an asynchronous and hardware-aware implementation that speeds up LLM inference. We validate that DejaVu can reduce the inference latency of OPT-175B by over 2X compared to the state-of-the-art FasterTransformer, and over 6X compared to the widely used Hugging Face implementation, without compromising model quality. The code is available at https://github.com/FMInference/DejaVu.

Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: https://github.com/HanzhiZhang-Ulrica/DAM.

In this work, we present QuickSRNet, an efficient super-resolution architecture for real-time applications on mobile platforms. Super-resolution clarifies, sharpens, and upscales an image to higher resolution. Applications such as gaming and video playback along with the ever-improving display capabilities of TVs, smartphones, and VR headsets are driving the need for efficient upscaling solutions. While existing deep learning-based super-resolution approaches achieve impressive results in terms of visual quality, enabling real-time DL-based super-resolution on mobile devices with compute, thermal, and power constraints is challenging. To address these challenges, we propose QuickSRNet, a simple yet effective architecture that provides better accuracy-to-latency trade-offs than existing neural architectures for single-image super resolution. We present training tricks to speed up existing residual-based super-resolution architectures while maintaining robustness to quantization. Our proposed architecture produces 1080p outputs via 2x upscaling in 2.2 ms on a modern smartphone, making it ideal for high-fps real-time applications.

Structured dilated attention has an appealing inference-time efficiency knob: it reduces the FLOPs of the attention and the KV cache size by a factor of the dilation size D, while preserving long-range connectivity. However, we find a persistent failure mode of them -sparsifying a pretrained attention model to a dilated pattern leads to severe accuracy degradation. We introduce RAT+, a dense-pretraining architecture that augments attention with full-sequence recurrence and active recurrence learning. A single RAT+ model is pretrained densely once, then flexibly switched at inference time to dilated attention (optionally with local windows) or hybrid layer/head compositions, requiring only a short 1B-token resolution adaptation rather than retraining separate sparse models. At 1.5B parameters trained on 100B tokens, RAT+ closely matches dense accuracy at D=16 and drops by about 2-3 points at D=64 on commonsense reasoning and LongBench tasks, respectively. Moreover, RAT+ outperforms attention when sparsifying to the top-k block attention. We further scale to 2.6B parameters and 200B tokens and observe the same trend. Code is available at https://github.com/wimh966/rat-plus.

Today's most advanced multimodal models remain proprietary. The strongest open-weight models rely heavily on synthetic data from proprietary VLMs to achieve good performance, effectively distilling these closed models into open ones. As a result, the community is still missing foundational knowledge about how to build performant VLMs from scratch. We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. Our key innovation is a novel, highly detailed image caption dataset collected entirely from human annotators using speech-based descriptions. To enable a wide array of user interactions, we also introduce a diverse dataset mixture for fine-tuning that includes in-the-wild Q&A and innovative 2D pointing data. The success of our approach relies on careful choices for the model architecture details, a well-tuned training pipeline, and, most critically, the quality of our newly collected datasets, all of which will be released. The best-in-class 72B model within the Molmo family not only outperforms others in the class of open weight and data models but also compares favorably against proprietary systems like GPT-4o, Claude 3.5, and Gemini 1.5 on both academic benchmarks and human evaluation. We will be releasing all of our model weights, captioning and fine-tuning data, and source code in the near future. Select model weights, inference code, and demo are available at https://molmo.allenai.org.

In this paper we undertake a systematic study of privacy attacks against open source Large Language Models (LLMs), where an adversary has access to either the model weights, gradients, or losses, and tries to exploit them to learn something about the underlying training data. Our headline results are the first membership inference attacks (MIAs) against pre-trained LLMs that are able to simultaneously achieve high TPRs and low FPRs, and a pipeline showing that over 50% (!) of the fine-tuning dataset can be extracted from a fine-tuned LLM in natural settings. We consider varying degrees of access to the underlying model, customization of the language model, and resources available to the attacker. In the pre-trained setting, we propose three new white-box MIAs: an attack based on the gradient norm, a supervised neural network classifier, and a single step loss ratio attack. All outperform existing black-box baselines, and our supervised attack closes the gap between MIA attack success against LLMs and other types of models. In fine-tuning, we find that given access to the loss of the fine-tuned and base models, a fine-tuned loss ratio attack FLoRA is able to achieve near perfect MIA peformance. We then leverage these MIAs to extract fine-tuning data from fine-tuned language models. We find that the pipeline of generating from fine-tuned models prompted with a small snippet of the prefix of each training example, followed by using FLoRa to select the most likely training sample, succeeds the majority of the fine-tuning dataset after only 3 epochs of fine-tuning. Taken together, these findings show that highly effective MIAs are available in almost all LLM training settings, and highlight that great care must be taken before LLMs are fine-tuned on highly sensitive data and then deployed.

In this paper, we provide the observation that too few queries assigned as positive samples in DETR with one-to-one set matching leads to sparse supervision on the encoder's output which considerably hurt the discriminative feature learning of the encoder and vice visa for attention learning in the decoder. To alleviate this, we present a novel collaborative hybrid assignments training scheme, namely Co-DETR, to learn more efficient and effective DETR-based detectors from versatile label assignment manners. This new training scheme can easily enhance the encoder's learning ability in end-to-end detectors by training the multiple parallel auxiliary heads supervised by one-to-many label assignments such as ATSS and Faster RCNN. In addition, we conduct extra customized positive queries by extracting the positive coordinates from these auxiliary heads to improve the training efficiency of positive samples in the decoder. In inference, these auxiliary heads are discarded and thus our method introduces no additional parameters and computational cost to the original detector while requiring no hand-crafted non-maximum suppression (NMS). We conduct extensive experiments to evaluate the effectiveness of the proposed approach on DETR variants, including DAB-DETR, Deformable-DETR, and DINO-Deformable-DETR. The state-of-the-art DINO-Deformable-DETR with Swin-L can be improved from 58.5% to 59.5% AP on COCO val. Surprisingly, incorporated with ViT-L backbone, we achieve 66.0% AP on COCO test-dev and 67.9% AP on LVIS val, outperforming previous methods by clear margins with much fewer model sizes. Codes are available at https://github.com/Sense-X/Co-DETR.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, methods such as GRPO and DAPO suffer from substantial computational cost, since they rely on sampling many rollouts for each prompt. Moreover, in RLVR the relative advantage is often sparse: many samples become nearly all-correct or all-incorrect, yielding low within-group reward variance and thus weak learning signals. In this paper, we introduce arrol (Accelerating RLVR via online Rollout Pruning), an online rollout pruning method that prunes rollouts during generation while explicitly steering the surviving ones more correctness-balanced to enhance learning signals. Specifically, arrol trains a lightweight quality head on-the-fly to predict the success probability of partial rollouts and uses it to make early pruning decisions. The learned quality head can further weigh candidates to improve inference accuracy during test-time scaling. To improve efficiency, we present a system design that prunes rollouts inside the inference engine and re-batches the remaining ones for log-probability computation and policy updates. Across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models (1B-8B), arrol improves average accuracy by +2.30 to +2.99 while achieving up to 1.7x training speedup, and yielding up to +8.33 additional gains in average accuracy in test-time scaling. The code is available at https://github.com/Hsu1023/ARRoL.

We introduce Long-VITA, a simple yet effective large multi-modal model for long-context visual-language understanding tasks. It is adept at concurrently processing and analyzing modalities of image, video, and text over 4K frames or 1M tokens while delivering advanced performances on short-context multi-modal tasks. We propose an effective multi-modal training schema that starts with large language models and proceeds through vision-language alignment, general knowledge learning, and two sequential stages of long-sequence fine-tuning. We further implement context-parallelism distributed inference and logits-masked language modeling head to scale Long-VITA to infinitely long inputs of images and texts during model inference. Regarding training data, Long-VITA is built on a mix of 17M samples from public datasets only and demonstrates the state-of-the-art performance on various multi-modal benchmarks, compared against recent cutting-edge models with internal data. Long-VITA is fully reproducible and supports both NPU and GPU platforms for training and testing. By leveraging our inference designs, Long-VITA models achieve a remarkable 2x prefill speedup and 4x context length extension in single node with 8 GPUs. We hope Long-VITA can serve as a competitive baseline and offer valuable insights for the open-source community in advancing long-context multi-modal understanding.

Inspired by the Many-Worlds Interpretation (MWI), this work introduces a novel neural network architecture that splits the same input signal into parallel branches at each layer, utilizing a Hyper Rectified Activation, referred to as ANDHRA. The branched layers do not merge and form separate network paths, leading to multiple network heads for output prediction. For a network with a branching factor of 2 at three levels, the total number of heads is 2^3 = 8 . The individual heads are jointly trained by combining their respective loss values. However, the proposed architecture requires additional parameters and memory during training due to the additional branches. During inference, the experimental results on CIFAR-10/100 demonstrate that there exists one individual head that outperforms the baseline accuracy, achieving statistically significant improvement with equal parameters and computational cost.

This paper tackles the problem of depth estimation from a single image. Existing work either focuses on generalization performance disregarding metric scale, i.e. relative depth estimation, or state-of-the-art results on specific datasets, i.e. metric depth estimation. We propose the first approach that combines both worlds, leading to a model with excellent generalization performance while maintaining metric scale. Our flagship model, ZoeD-M12-NK, is pre-trained on 12 datasets using relative depth and fine-tuned on two datasets using metric depth. We use a lightweight head with a novel bin adjustment design called metric bins module for each domain. During inference, each input image is automatically routed to the appropriate head using a latent classifier. Our framework admits multiple configurations depending on the datasets used for relative depth pre-training and metric fine-tuning. Without pre-training, we can already significantly improve the state of the art (SOTA) on the NYU Depth v2 indoor dataset. Pre-training on twelve datasets and fine-tuning on the NYU Depth v2 indoor dataset, we can further improve SOTA for a total of 21% in terms of relative absolute error (REL). Finally, ZoeD-M12-NK is the first model that can jointly train on multiple datasets (NYU Depth v2 and KITTI) without a significant drop in performance and achieve unprecedented zero-shot generalization performance to eight unseen datasets from both indoor and outdoor domains. The code and pre-trained models are publicly available at https://github.com/isl-org/ZoeDepth .

Transformer model with multi-head attention requires caching intermediate results for efficient inference in generation tasks. However, cache brings new memory-related costs and prevents leveraging larger batch size for faster speed. We propose memory-efficient lossless attention (called EL-attention) to address this issue. It avoids heavy operations for building multi-head keys and values, cache for them is not needed. EL-attention constructs an ensemble of attention results by expanding query while keeping key and value shared. It produces the same result as multi-head attention with less GPU memory and faster inference speed. We conduct extensive experiments on Transformer, BART, and GPT-2 for summarization and question generation tasks. The results show EL-attention speeds up existing models by 1.6x to 5.3x without accuracy loss.

Large Language Models (LLMs) have shown impressive capabilities but still suffer from the issue of hallucinations. A significant type of this issue is the false premise hallucination, which we define as the phenomenon when LLMs generate hallucinated text when confronted with false premise questions. In this paper, we perform a comprehensive analysis of the false premise hallucination and elucidate its internal working mechanism: a small subset of attention heads (which we designate as false premise heads) disturb the knowledge extraction process, leading to the occurrence of false premise hallucination. Based on our analysis, we propose FAITH (False premise Attention head constraIining for miTigating Hallucinations), a novel and effective method to mitigate false premise hallucinations. It constrains the false premise attention heads during the model inference process. Impressively, extensive experiments demonstrate that constraining only approximately 1% of the attention heads in the model yields a notable increase of nearly 20% of model performance.