Inference Acceleration

Model merging offers an effective way to integrate the capabilities of multiple fine-tuned models. However, the performance degradation of the merged model remains a challenge, particularly when none or few data are available. This paper first highlights the necessity of domain-specific data for model merging by proving that data-agnostic algorithms can have arbitrarily bad worst-case performances. Building on this theoretical insight, we explore the relationship between model merging and distillation, introducing a novel few-shot merging algorithm, ProDistill (Progressive Layer-wise Distillation). Unlike common belief that layer-wise training hurts performance, we show that layer-wise teacher-student distillation not only enhances the scalability but also improves model merging performance. We conduct extensive experiments to show that compared to existing few-shot merging methods,ProDistill achieves state-of-the-art performance, with up to 6.14\% and 6.61\% improvements in vision and NLU tasks. Furthermore, we extend the experiments to models with over 10B parameters, showcasing the exceptional scalability of ProDistill.

Neural network ensembles have proven effective in improving performance across a range of tasks; however, their high computational cost limits their applicability in resource-constrained environments or for large models. Ensemble distillation, the process of transferring knowledge from an ensemble teacher to a smaller student model, offers a promising solution to this challenge. The key is to ensure that the student model is both cost-efficient and achieves performance comparable to the ensemble teacher. With this in mind, we propose a novel ensemble distribution distillation method, which leverages flow matching to effectively transfer the diversity from the ensemble teacher to the student model. Our extensive experiments demonstrate the effectiveness of our proposed method compared to existing ensemble distillation approaches.

In this work we present the BDIA-transformer, which is an exact bit-level reversible transformer that uses an unchanged standard architecture for inference. The basic idea is to first treat each transformer block as the Euler integration approximation for solving an ordinary differential equation (ODE) and then incorporate the technique of bidirectional integration approximation (BDIA) (originally designed for diffusion inversion) into the neural architecture, together with activation quantization to make it exactly bit-level reversible. In the training process, we let a hyper-parameter $\gamma$ in BDIA-transformer randomly take one of the two values $\{0.5, -0.5\}$ per training sample per transformer block for averaging every two consecutive integration approximations. As a result, BDIA-transformer can be viewed as training an ensemble of ODE solvers parameterized by a set of binary random variables, which regularizes the model and results in improved validation accuracy. Lightweight side information is required to be stored in the forward process to account for binary quantization loss to enable exact bit-level reversibility. In the inference procedure, the expectation $\mathbb{E}(\gamma)=0$ is taken to make the resulting architecture identical to transformer up to activation quantization. Our experiments in natural language generation, image classification, and language translation show that BDIA-transformers outperform their conventional counterparts significantly in terms of validation performance while also requiring considerably less training memory. Thanks to the regularizing effect of the ensemble, the BDIA-transformer is particularly suitable for fine-tuning with limited data. Source-code can be found via \href{https://github.com/guoqiang-zhang-x/BDIA-Transformer}{this link}.

The growing computational demands of training large language models (LLMs) necessitate more efficient methods. Quantized training presents a promising solution by enabling low-bit arithmetic operations to reduce these costs. While FP8 precision has demonstrated feasibility, leveraging FP4 remains a challenge due to significant quantization errors and limited representational capacity. This work introduces the first FP4 training framework for LLMs, addressing these challenges with two key innovations: a differentiable quantization estimator for precise weight updates and an outlier clamping and compensation strategy to prevent activation collapse. To ensure stability, the framework integrates a mixed-precision training scheme and vector-wise quantization. Experimental results demonstrate that our FP4 framework achieves accuracy comparable to BF16 and FP8, with minimal degradation, scaling effectively to 13B-parameter LLMs trained on up to 100B tokens. With the emergence of next-generation hardware supporting FP4, our framework sets a foundation for efficient ultra-low precision training.

We present any4, a learned 4-bit weight quantization solution for large language models (LLMs) providing arbitrary numeric representations without requiring pre-processing of weights or activations. any4 yields higher accuracy compared to other related 4-bit numeric representation types: int4, fp4 and nf4, as evaluated on a range of model sizes, generations and families (Llama 2, Llama 3, Mistral and Mixtral). While any4 does not require preprocessing of weights or activations, it is also competitive with orthogonal techniques that require such preprocessing (e.g., AWQ and GPTQ). We also experiment with any3 and any2 and show competitiveness at lower bits. Additionally, we show that we can calibrate using a single curated diverse sample rather than hundreds of samples from a dataset as done in most quantization approaches. We also open source tinygemm, a latency optimized GPU matrix multiplication library for LLMs, that implements any4 using a GPU-efficient lookup table strategy along with other common quantization methods. We open source our code at https://github.com/facebookresearch/any4.

Existing analysis of Local (Stochastic) Gradient Descent for heterogeneous objectives requires stepsizes $\eta \leq 1/K$ where $K$ is the communication interval, which ensures monotonic decrease of the objective. In contrast, we analyze Local Gradient Descent for logistic regression with separable, heterogeneous data using any stepsize $\eta > 0$. With $R$ communication rounds and $M$ clients, we show convergence at a rate $\mathcal{O}(1/\eta K R)$ after an initial unstable phase lasting for $\widetilde{\mathcal{O}}(\eta K M)$ rounds. This improves upon the existing $\mathcal{O}(1/R)$ rate for general smooth, convex objectives. Our analysis parallels the single machine analysis of Wu et al. (2024) in which instability is caused by extremely large stepsizes, but in our setting another source of instability is large local updates with heterogeneous objectives.

Quantizing model weights is critical for reducing the communication and inference costs of large models. However, quantizing models – especially to low precisions like int4 or int2 – requires a trade-off in model quality; int2, in particular, is known to severely degrade model quality. Consequently, practitioners are often forced to maintain multiple models with different quantization levels or serve a single model that best satisfies the quality-latency trade-off. On the other hand, integer data types, such as int8, inherently possess a nested (Matryoshka) structure where smaller bit-width integers, like int4 or int2, are nested within the most significant bits. Leveraging this insight, in this paper, we propose Matryoshka Quantization (MatQuant), a novel multi-scale quantization technique that alleviates the aforementioned challenge. This technique allows us to train and maintain a single quantized model but serve it with the precision demanded by the deployment. Furthermore, leveraging MatQuant’s co-training and co-distillation, int2 precision models extracted by MatQuant outperform standard int2 quantization by up to 4% and 7% with OmniQuant and QAT as base algorithms respectively. Finally, we demonstrate that by using an extra bit to represent outliers, a model with an effective precision of 2.05-bit improves further by 6% with OmniQuant as the base algorithm.

Although quantization for linear layers has been widely used, its application to accelerate the attention process remains limited. To further enhance the efficiency of attention computation compared to SageAttention while maintaining precision, we propose SageAttention2, which utilizes significantly faster 4-bit matrix multiplication (Matmul) alongside additional precision-enhancing techniques. First, we propose to quantize matrixes $(Q, K)$ to INT4 in a hardware-friendly thread-level granularity and quantize matrixes $(\widetilde P, V)$ to FP8. Second, we propose a method to smooth $Q$, enhancing the accuracy of INT4 $QK^\top$. Third, we propose a two-level accumulation strategy for $\widetilde PV$ to enhance the accuracy of FP8 $\widetilde PV$. The operations per second (OPS) of SageAttention2 surpass FlashAttention2 and xformers by about **3x** and **4.5x**, respectively. Moreover, SageAttention2 matches the speed of FlashAttention3(fp8) on the Hopper GPUs, while delivering much higher accuracy. Comprehensive experiments confirm that our approach incurs negligible end-to-end metrics loss across diverse models, including those for language, image, and video generation.

We reveal that feedforward network (FFN) layers, rather than attention layers, are the primary contributors to Vision Transformer (ViT) inference latency, with their impact signifying as model size increases. This finding highlights a critical opportunity for optimizing the efficiency of large-scale ViTs by focusing on FFN layers. In this work, we propose a novel channel idle mechanism that facilitates post-training structural reparameterization for efficient FFN layers during testing. Specifically, a set of feature channels remains idle and bypasses the nonlinear activation function in each FFN layer, thereby forming a linear pathway that enables structural reparameterization during inference. This mechanism results in a family of **RePa**rameterizable **Vi**sion **T**ransformers (RePaViTs), which achieve remarkable latency reductions with acceptable sacrifices (sometimes gains) in accuracy across various ViTs. The benefits of our method scale consistently with model sizes, demonstrating greater speed improvements and progressively narrowing accuracy gaps or even higher accuracies on larger models. In particular, RePa-ViT-Large and RePa-ViT-Huge enjoy **66.8%** and **68.7%** speed-ups with **+1.7%** and **+1.1%** higher top-1 accuracies under the same training strategy, respectively. RePaViT is the first to employ structural reparameterization on FFN layers to expedite ViTs to our best knowledge, and we believe that it represents an auspicious direction for efficient ViTs. Source code is available at https://github.com/Ackesnal/RePaViT.

Speculative decoding has emerged as a popular method to accelerate the inference of Large Language Models (LLMs) while retaining their superior text generation performance. Previous methods either adopt a fixed speculative decoding configuration regardless of the prefix tokens, or train draft models in an offline or online manner to align them with the context. This paper proposes a training-free online learning framework to adaptively choose the configuration of the hyperparameters for speculative decoding as text is being generated. We first formulate this hyperparameter selection problem as a Multi-Armed Bandit problem and provide a general speculative decoding framework BanditSpec. Furthermore, two bandit-based hyperparameter selection algorithms, UCBSpec and EXP3Spec, are designed and analyzed in terms of a novel quantity, the stopping time regret. We upper bound this regret under both stochastic and adversarial reward settings. By deriving an information-theoretic impossibility result, it is shown that the regret performance of UCBSpec is optimal up to universal constants. Finally, extensive empirical experiments with LLaMA3 and Qwen2 demonstrate that our algorithms are effective compared to existing methods, and the throughput is close to the oracle best hyperparameter in simulated real-life LLM serving scenarios with diverse input prompts.

Large multimodal models excel in multimodal tasks but face significant computational challenges due to excessive visual tokens. Unlike token reduction methods that focus on token-level redundancy, we identify and study the computation-level redundancy on vision tokens to ensure no information loss. Our key insight is that vision tokens from the pretrained vision encoder do not necessarily require all the heavy operations (e.g., self-attention, FFNs) in decoder-only LMMs and could be processed more lightly with proper designs. We designed a series of experiments to discover and progressively squeeze out the vision-related computation redundancy. Based on our findings, we propose ProxyV, a novel approach that utilizes proxy vision tokens to alleviate the computational burden on original vision tokens. ProxyV enhances efficiency without compromising performance and can even yield notable performance gains in scenarios with more moderate efficiency improvements. Furthermore, the flexibility of ProxyV is demonstrated through its combination with token reduction methods to boost efficiency further.

State Space Models (SSMs) are gaining attention as an efficient alternative to Transformers due to their constant memory complexity and comparable performance. Yet, deploying large-scale SSMs on cloud-based services or resource-constrained devices faces challenges. To address this, quantizing SSMs using low bit-width data types is proposed to reduce model size and leverage hardware acceleration. Given that SSMs are sensitive to quantization errors, recent advancements focus on quantizing a specific model or bit-width to improve their efficiency while maintaining performance. However, different bit-width configurations, such as W4A8 for cloud service throughput and W4A16 for improving question-answering on personal devices, are necessary for specific scenarios. To this end, we present Quamba2, compatible with \textbf{W8A8}, \textbf{W4A8}, and \textbf{W4A16} for both \textbf{Mamba} and \textbf{Mamba2}, addressing the rising demand for SSM deployment across various platforms. We propose an offline approach to quantize inputs of a linear recurrence in 8-bit by sorting and clustering for $x$, combined with a per-state-group quantization for $B$ and $C$. To ensure compute-invariance in the SSM output, we offline rearrange weights according to the clustering sequence. The experiments show Quamba2-8B outperforms several state-of-the-art SSMs quantization methods and delivers 1.3$\times$ and 3$\times$ speedup in the pre-filling and generation stages and 4$\times$ memory reduction with only a $1.6$% accuracy drop on average. The code and quantized models will be released at:

Decoding with autoregressive language models traditionally occurs sequentially, generating one token after another. Recent attempts to introduce parallelism require a pre-determined structure in the generated content to implement parallel generation, such as by pattern-matching on bullet points. In this work, we present a new technique to automate parallel generation by dynamically exploiting the semantic independence of generation outputs to implement asynchronous decoding. We introduce an annotation language Pasta-Lang for language models to initiate asynchronous decoding at inference time. We also develop an accompanying Pasta-Lang interpreter that performs on-the-fly asynchronous decoding, effectively implementing parallel generation and speeding up inference. We present an instruction-finetuning dataset with Pasta-Lang-annotated responses for teaching LLMs to annotate semantic independence with Pasta-Lang as well as the methodology for creating the dataset. Our evaluation shows using the interpreter with a Pasta-Lang-equipped model achieves significant speedup while maintaining the same generation quality.

Dynamic data selection aims to accelerate training with lossless performances. However, reducing training data inherently limits data diversity, potentially hindering generalization. While data augmentation is widely used to enhance diversity, it is typically not optimized in conjunction with selection. As a result, directly combining these techniques fails to fully exploit their synergies. To tackle the challenge, we propose a novel online data training framework that, for the first time, unifies dynamic data selection and augmentation, achieving both training efficiency and enhanced performance. Our method estimates each sample's joint distribution of local density and multimodal semantic consistency, allowing for the targeted selection of augmentation-suitable samples while suppressing the inclusion of noisy or ambiguous data. This enables a more significant reduction in dataset size without sacrificing model generalization. Experimental results demonstrate that our method outperforms existing state-of-the-art approaches on various benchmark datasets and architectures, e.g., reducing 50% training costs on ImageNet-1k with lossless performance. Furthermore, our approach enhances noise resistance and improves model robustness, reinforcing its practical utility in real-world scenarios.

Diffusion models have achieved remarkable success in generative tasks but suffer from high computational costs due to their iterative sampling process and quadratic‐attention costs. Existing training-free acceleration strategies that reduce per-step computation cost, while effectively reducing sampling time, demonstrate low faithfulness compared to the original baseline. We hypothesize that this fidelity gap arises because (a) different prompts correspond to varying denoising trajectory, and (b) such methods do not consider the underlying ODE formulation and its numerical solution. In this paper, we propose **Stability-guided Adaptive Diffusion Acceleration (SADA)**, a novel paradigm that unifies step-wise and token-wise sparsity decisions via a single stability criterion to accelerate sampling of ODE-based generative models (Diffusion and Flow-matching). For (a), SADA adaptively allocates sparsity based on the sampling trajectory. For (b), SADA introduces principled approximation schemes that leverage the precise gradient information from the numerical ODE solver. Comprehensive evaluations on SD‐2, SDXL, and Flux using both EDM and DPM++ solvers reveal consistent $\ge 1.8\times$ speedups with minimal fidelity degradation (LPIPS $\leq 0.10$ and FID $\leq 4.5$) compared to unmodified baselines, significantly outperforming prior methods. Moreover, SADA adapts seamlessly to other pipelines and modalities: It accelerates ControlNet without any modifications and speeds up MusicLDM by \(1.8\times\) with \(\sim 0.01\) spectrogram LPIPS. Our code is available at: [https://github.com/Ting-Justin-Jiang/sada-icml](https://github.com/Ting-Justin-Jiang/sada-icml).

Large Language Model (LLM) collaborative decoding techniques improve output quality by combining the outputs of multiple models at each generation step, but they incur high computational costs. In this paper, we introduce **Collaborative decoding via Speculation (CoS)**, a novel framework that accelerates collaborative decoding without compromising performance. Inspired by Speculative Decoding—where a small proposal model generates tokens sequentially, and a larger target model verifies them in parallel, our approach builds on two key insights: (1) the verification distribution can be the combined distribution of both the proposal and target models, and (2) alternating each model as the proposer and verifier can further enhance efficiency. We generalize this method to collaboration among *n* models and theoretically prove that CoS is never slower than standard collaborative decoding, typically achieving faster speed. Extensive experiments demonstrate CoS is **1.11x–2.23x** faster than standard collaborative decoding without compromising generation quality. Our code is available at https://github.com/Kamichanw/CoS/.

Activation sparsity denotes the existence of substantial weakly-contributed neurons within feed-forward networks of large language models (LLMs), providing wide potential benefits such as computation acceleration. However, existing works lack thorough quantitative studies on this useful property, in terms of both its measurement and influential factors. In this paper, we address three underexplored research questions: (1) How can activation sparsity be measured more accurately? (2) How is activation sparsity affected by the model architecture and training process? (3) How can we build a more sparsely activated and efficient LLM? Specifically, we develop a generalizable and performance-friendly metric, named CETT-PPL-1\%, to measure activation sparsity. Based on CETT-PPL-1\%, we quantitatively study the influence of various factors and observe several important phenomena, such as the convergent power-law relationship between sparsity and training data amount, the higher competence of ReLU activation than mainstream SiLU activation, the potential sparsity merit of a small width-depth ratio, and the scale insensitivity of activation sparsity. Finally, we provide implications for building sparse and effective LLMs, and demonstrate the reliability of our findings by training a 2.4B model with a sparsity ratio of 93.52\%, showing 4.1$\times$ speedup compared with its dense version. The codes and checkpoints are available at https://github.com/thunlp/SparsingLaw/.

Large Language Models (LLMs) exhibit impressive performance across various tasks, but deploying them for inference poses challenges. Their high resource demands often necessitate complex, costly multi-GPU pipelines, or the use of smaller, less capable models. While quantization offers a promising solution utilizing lower precision for model storage, existing methods frequently experience significant performance drops at lower precision levels. Additionally, they typically provide only a limited set of solutions at specific bit levels, many of which are extensively manually tuned. To address these challenges, we propose a new method called \textbf{SKIM}: Scaled K-means clustering wIth Mixed precision. Our approach introduces two novel techniques: 1. A \textit{greedy algorithm} to solve approximately optimal bit allocation across weight channels, and 2. A \textit{trainable scaling vector} for non-differentiable K-means clustering. These techniques substantially improve the model performance and can be adapted to any given bit. Notably, in terms of perplexity, our method narrows the gap between quantized LLaMA models and their full precision counterparts by around \textbf{14\%} on average.

Sparse tensor programs are essential in deep learning and graph analytics, driving the need for optimized processing. To meet this demand, specialized hardware accelerators are being developed. Optimizing these programs for accelerators is challenging for two reasons: program performance is highly sensitive to variations in sparse inputs, and early-stage accelerators rely on expensive simulators. Therefore, ML-based cost models used for optimizing such programs on general-purpose hardware are often ineffective for early-stage accelerators, as they require large datasets for proper training. To this end, we introduce COGNATE, a novel framework that leverages inexpensive data samples from general-purpose hardware (e.g., CPUs) to train cost models, followed by few-shot fine-tuning on emerging hardware. COGNATE exploits the homogeneity of input features across hardware platforms while effectively mitigating heterogeneity, enabling cost model training with just 5% of the data samples needed by accelerator-specific models to achieve comparable performance. We conduct extensive experiments to demonstrate that COGNATE outperforms existing techniques, achieving average speedups of 1.47× (up to 5.46×) for SpMM and 1.39× (up to 4.22×) for SDDMM.

The vast diversity of styles, domains, and quality levels present in language model pre-training corpora is essential in developing general model capabilities, but efficiently learning and deploying the correct behaviors exemplified in each of these heterogeneous data sources is challenging. To address this, we propose a new method, termed Metadata Conditioning then Cooldown (MeCo), to incorporate additional learning cues during pre-training. MeCo first provides metadata (e.g., URLs like en.wikipedia.org) alongside the text during training and later uses a cooldown phase with only the standard text, thereby enabling the model to function normally even without metadata. MeCo significantly accelerates pre-training across different model scales (600M to 8B parameters) and training sources (C4, RefinedWeb, and DCLM). For instance, a 1.6B language model trained with MeCo matches the downstream task performance of standard pre-training while using 33% less data. Additionally, MeCo enables us to steer language models by conditioning the inference prompt on either real or fabricated metadata that encodes the desired properties of the output: for example, prepending wikipedia.org to reduce harmful generations or factquizmaster.com (fabricated) to improve common knowledge task performance. We also demonstrate that MeCo is compatible with different types of metadata, such as model-generated topics. MeCo is remarkably simple, adds no computational overhead, and demonstrates promise in producing more capable and steerable language models.