SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator

Large Language Models (LLMs) have exhibited exceptional performance across a spectrum of natural language processing tasks. However, their substantial sizes pose considerable challenges, particularly in computational demands and inference speed, due to their quadratic complexity. In this work, we have identified a key pattern: certain seemingly meaningless special tokens (i.e., separators) contribute disproportionately to attention scores compared to semantically meaningful tokens. This observation suggests that information of the segments between these separator tokens can be effectively condensed into the separator tokens themselves without significant information loss. Guided by this insight, we introduce SepLLM, a plug-and-play framework that accelerates inference by compressing these segments and eliminating redundant tokens. Additionally, we implement efficient kernels for training acceleration. Experimental results across training-free, training-from-scratch, and post-training settings demonstrate SepLLM's effectiveness. Notably, using the Llama-3-8B backbone, SepLLM achieves over 50% reduction in KV cache on the GSM8K-CoT benchmark while maintaining comparable performance. Furthermore, in streaming settings, SepLLM effectively processes sequences of up to 4 million tokens or more while maintaining consistent language modeling capabilities.

Efficient Multi-modal Large Language Models via Visual Token Grouping

The development of Multi-modal Large Language Models (MLLMs) enhances Large Language Models (LLMs) with the ability to perceive data formats beyond text, significantly advancing a range of downstream applications, such as visual question answering and image captioning. However, the substantial computational costs associated with processing high-resolution images and videos pose a barrier to their broader adoption. To address this challenge, compressing vision tokens in MLLMs has emerged as a promising approach to reduce inference costs. While existing methods conduct token reduction in the feature alignment phase. In this paper, we introduce VisToG, a novel grouping mechanism that leverages the capabilities of pre-trained vision encoders to group similar image segments without the need for segmentation masks. Specifically, we concatenate semantic tokens to represent image semantic segments after the linear projection layer before feeding into the vision encoder. Besides, with the isolated attention we adopt, VisToG can identify and eliminate redundant visual tokens utilizing the prior knowledge in the pre-trained vision encoder, which effectively reduces computational demands. Extensive experiments demonstrate the effectiveness of VisToG, maintaining 98.1% of the original performance while achieving a reduction of over 27\% inference time.

DAPE V2: Process Attention Score as Feature Map for Length Extrapolation

The attention mechanism is a fundamental component of the Transformer model, contributing to interactions among distinct tokens, in contrast to earlier feed-forward neural networks. In general, the attention scores are determined simply by the key-query products. However, this work's occasional trial (combining DAPE and NoPE) of including additional MLPs on attention scores without position encoding indicates that the classical key-query multiplication may limit the performance of Transformers. In this work, we conceptualize attention as a feature map and apply the convolution operator (for neighboring attention scores across different heads) to mimic the processing methods in computer vision. Specifically, the main contribution of this paper is identifying and interpreting the Transformer length extrapolation problem as a result of the limited expressiveness of the naive query and key dot product, and we successfully translate the length extrapolation issue into a well-understood feature map processing problem. The novel insight, which can be adapted to various attention-related models, reveals that the current Transformer architecture has the potential for further evolution. Extensive experiments demonstrate that treating attention as a feature map and applying convolution as a processing method significantly enhances Transformer performance.

Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding

The current large auto-regressive models can generate high-quality, high-resolution images, but these models require hundreds or even thousands of steps of next-token prediction during inference, resulting in substantial time consumption. In existing studies, Jacobi decoding, an iterative parallel decoding algorithm, has been used to accelerate the auto-regressive generation and can be executed without training. However, the Jacobi decoding relies on a deterministic criterion to determine the convergence of iterations. Thus, it works for greedy decoding but is incompatible with sampling-based decoding which is crucial for visual quality and diversity in the current auto-regressive text-to-image generation. In this paper, we propose a training-free probabilistic parallel decoding algorithm, Speculative Jacobi Decoding (SJD), to accelerate auto-regressive text-to-image generation. By introducing a probabilistic convergence criterion, our SJD accelerates the inference of auto-regressive text-to-image generation while maintaining the randomness in sampling-based token decoding and allowing the model to generate diverse images. Specifically, SJD facilitates the model to predict multiple tokens at each step and accepts tokens based on the probabilistic criterion, enabling the model to generate images with fewer steps than the conventional next-token-prediction paradigm. We also investigate the token initialization strategies that leverage the spatial locality of visual data to further improve the acceleration ratio under specific scenarios. We conduct experiments for our proposed SJD on multiple auto-regressive text-to-image generation models, showing the effectiveness of model acceleration without sacrificing the visual quality.

QuickLLaMA: Query-aware Inference Acceleration for Large Language Models

The capacity of Large Language Models (LLMs) to comprehend and reason over long contexts is pivotal for advancements in diverse fields. Yet, they still stuggle with capturing long-distance dependencies within sequences to deeply understand semantics. To address this issue, we introduce Query-aware Inference for LLMs (Q-LLM), a system designed to process extensive sequences akin to human cognition. By focusing on memory data relevant to a given query, Q-LLM can accurately capture pertinent information within a fixed window size and provide precise answers to queries. It doesn't require extra training and can be seamlessly integrated with any LLMs. Q-LLM using LLaMA3 (QuickLLaMA) can read Harry Potter within 30s and accurately answer the questions. Q-LLM improved by 7.17% compared to the current state-of-the-art on LLaMA3, and by 3.26% on Mistral on the $\infty$-bench. In the Needle-in-a-Haystack task, On widely recognized benchmarks, Q-LLM improved upon the current SOTA by 7.0% on Mistral and achieves 100% on LLaMA3. Our code can be found in https://github.com/dvlab-research/Q-LLM.

DiM: Diffusion Mamba for Efficient High-Resolution Image Synthesis

Diffusion models have achieved great success in image generation, with the backbone evolving from U-Net to Vision Transformers. However, the computational cost of Transformers is quadratic to the number of tokens, leading to significant challenges when dealing with high-resolution images. In this work, we propose Diffusion Mamba (DiM), which combines the efficiency of Mamba, a sequence model based on State Space Models (SSM), with the expressive power of diffusion models for efficient high-resolution image synthesis. To address the challenge that Mamba cannot generalize to 2D signals, we make several architecture designs including multi-directional scans, learnable padding tokens at the end of each row and column, and lightweight local feature enhancement. Our DiM architecture achieves inference-time efficiency for high-resolution images. In addition, to further improve training efficiency for high-resolution image generation with DiM, we investigate ``weak-to-strong'' training strategy that pretrains DiM on low-resolution images ($256\times 256$) and then finetune it on high-resolution images ($512 \times 512$). We further explore training-free upsampling strategies to enable the model to generate higher-resolution images (e.g., $1024\times 1024$ and $1536\times 1536$) without further fine-tuning. Experiments demonstrate the effectiveness and efficiency of our DiM.

CAPE: Context-Adaptive Positional Encoding for Length Extrapolation

Positional encoding plays a crucial role in transformers, significantly impacting model performance and length generalization. Prior research has introduced absolute positional encoding (APE) and relative positional encoding (RPE) to distinguish token positions in given sequences. However, both APE and RPE remain fixed after model training regardless of input data, limiting their adaptability and flexibility. Hence, we expect that the desired positional encoding should be context-adaptive and can be dynamically adjusted with the given attention. In this paper, we propose a Context-Adaptive Positional Encoding (CAPE) method, which dynamically and semantically adjusts based on input context and learned fixed priors. Experimental validation on real-world datasets (Arxiv, Books3, and CHE) demonstrates that CAPE enhances model performances in terms of trained length and length generalization, where the improvements are statistically significant. The model visualization suggests that our model can keep both local and anti-local information. Finally, we successfully train the model on sequence length 128 and achieve better performance at evaluation sequence length 8192, compared with other static positional encoding methods, revealing the benefit of the adaptive positional encoding method.

On the Expressive Power of a Variant of the Looped Transformer

Besides natural language processing, transformers exhibit extraordinary performance in solving broader applications, including scientific computing and computer vision. Previous works try to explain this from the expressive power and capability perspectives that standard transformers are capable of performing some algorithms. To empower transformers with algorithmic capabilities and motivated by the recently proposed looped transformer (Yang et al., 2024; Giannou et al., 2023), we design a novel transformer block, dubbed Algorithm Transformer (abbreviated as AlgoFormer). Compared with the standard transformer and vanilla looped transformer, the proposed AlgoFormer can achieve significantly higher expressiveness in algorithm representation when using the same number of parameters. In particular, inspired by the structure of human-designed learning algorithms, our transformer block consists of a pre-transformer that is responsible for task pre-processing, a looped transformer for iterative optimization algorithms, and a post-transformer for producing the desired results after post-processing. We provide theoretical evidence of the expressive power of the AlgoFormer in solving some challenging problems, mirroring human-designed algorithms. Furthermore, some theoretical and empirical results are presented to show that the designed transformer has the potential to be smarter than human-designed algorithms. Experimental results demonstrate the empirical superiority of the proposed transformer in that it outperforms the standard transformer and vanilla looped transformer in some challenging tasks.

Diffusion of Thoughts: Chain-of-Thought Reasoning in Diffusion Language Models

Diffusion models have gained attention in text processing, offering many potential advantages over traditional autoregressive models. This work explores the integration of diffusion models and Chain-of-Thought (CoT), a well-established technique to improve the reasoning ability in autoregressive language models. We propose Diffusion-of-Thought (DoT), allowing reasoning steps to diffuse over time through the diffusion process. In contrast to traditional autoregressive language models that make decisions in a left-to-right, token-by-token manner, DoT offers more flexibility in the trade-off between computation and reasoning performance. Our experimental results demonstrate the effectiveness of DoT in multi-digit multiplication and grade school math problems. Additionally, DoT showcases promising self-correction abilities and benefits from existing reasoning-enhancing techniques like self-consistency decoding. Our findings contribute to the understanding and development of reasoning capabilities in diffusion language models.

LEGO-Prover: Neural Theorem Proving with Growing Libraries

Despite the success of large language models (LLMs), the task of theorem proving still remains one of the hardest reasoning tasks that is far from being fully solved. Prior methods using language models have demonstrated promising results, but they still struggle to prove even middle school level theorems. One common limitation of these methods is that they assume a fixed theorem library during the whole theorem proving process. However, as we all know, creating new useful theorems or even new theories is not only helpful but crucial and necessary for advancing mathematics and proving harder and deeper results. In this work, we present LEGO-Prover, which employs a growing skill library containing verified lemmas as skills to augment the capability of LLMs used in theorem proving. By constructing the proof modularly, LEGO-Prover enables LLMs to utilize existing skills retrieved from the library and to create new skills during the proving process. These skills are further evolved (by prompting an LLM) to enrich the library on another scale. Modular and reusable skills are constantly added to the library to enable tackling increasingly intricate mathematical problems. Moreover, the learned library further bridges the gap between human proofs and formal proofs by making it easier to impute missing steps. LEGO-Prover advances the state-of-the-art pass rate on miniF2F-valid (48.0% to 57.0%) and miniF2F-test (45.5% to 47.1%). During the proving process, LEGO-Prover also manages to generate over 20,000 skills (theorems/lemmas) and adds them to the growing library. Our ablation study indicates that these newly added skills are indeed helpful for proving theorems, resulting in an improvement from a success rate of 47.1% to 50.4%. We also release our code and all the generated skills.