MoH: Multi-Head Attention as Mixture-of-Head Attention

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.

MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts

In this work, we aim to simultaneously enhance the effectiveness and efficiency of Mixture-of-Experts (MoE) methods. To achieve this, we propose MoE++, a general and heterogeneous MoE framework that integrates both Feed-Forward Network~(FFN) and zero-computation experts. Specifically, we introduce three types of zero-computation experts: the zero expert, copy expert, and constant expert, which correspond to discard, skip, and replace operations, respectively. This design offers three key advantages: (i) Low Computing Overhead: Unlike the uniform mixing mechanism for all tokens within vanilla MoE, MoE++ allows each token to engage with a dynamic number of FFNs, be adjusted by constant vectors, or even skip the MoE layer entirely. (ii) High Performance: By enabling simple tokens to utilize fewer FFN experts, MoE++ allows more experts to focus on challenging tokens, thereby unlocking greater performance potential than vanilla MoE. (iii) Deployment Friendly: Given that zero-computation experts have negligible parameters, we can deploy all zero-computation experts on each GPU, eliminating the significant communication overhead and expert load imbalance associated with FFN experts distributed across different GPUs. Moreover, we leverage gating residuals, enabling each token to consider the pathway taken in the previous layer when selecting the appropriate experts. Extensive experimental results demonstrate that MoE++ achieves better performance while delivering 1.1-2.1x expert forward throughput compared to a vanilla MoE model of the same size, which lays a solid foundation for developing advanced and efficient MoE-related models.

MUSE: Mamba is Efficient Multi-scale Learner for Text-video Retrieval

Text-Video Retrieval (TVR) aims to align and associate relevant video content with corresponding natural language queries. Most existing TVR methods are based on large-scale pre-trained vision-language models (e.g., CLIP). However, due to the inherent plain structure of CLIP, few TVR methods explore the multi-scale representations which offer richer contextual information for a more thorough understanding. To this end, we propose MUSE, a multi-scale mamba with linear computational complexity for efficient cross-resolution modeling. Specifically, the multi-scale representations are generated by applying a feature pyramid on the last single-scale feature map. Then, we employ the Mamba structure as an efficient multi-scale learner to jointly learn scale-wise representations. Furthermore, we conduct comprehensive studies to investigate different model structures and designs. Extensive results on three popular benchmarks have validated the superiority of MUSE.

Local Action-Guided Motion Diffusion Model for Text-to-Motion Generation

Text-to-motion generation requires not only grounding local actions in language but also seamlessly blending these individual actions to synthesize diverse and realistic global motions. However, existing motion generation methods primarily focus on the direct synthesis of global motions while neglecting the importance of generating and controlling local actions. In this paper, we propose the local action-guided motion diffusion model, which facilitates global motion generation by utilizing local actions as fine-grained control signals. Specifically, we provide an automated method for reference local action sampling and leverage graph attention networks to assess the guiding weight of each local action in the overall motion synthesis. During the diffusion process for synthesizing global motion, we calculate the local-action gradient to provide conditional guidance. This local-to-global paradigm reduces the complexity associated with direct global motion generation and promotes motion diversity via sampling diverse actions as conditions. Extensive experiments on two human motion datasets, i.e., HumanML3D and KIT, demonstrate the effectiveness of our method. Furthermore, our method provides flexibility in seamlessly combining various local actions and continuous guiding weight adjustment, accommodating diverse user preferences, which may hold potential significance for the community. The project page is available at https://jpthu17.github.io/GuidedMotion-project/.

LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference

Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs' KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge. In this work, we introduce LOOK-M, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. LOOK-M demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by 80% in some cases, it not only achieves up to 1.5x faster decoding but also maintains or even enhances performance across a variety of long context multimodal tasks.

RAP: Efficient Text-Video Retrieval with Sparse-and-Correlated Adapter

Text-Video Retrieval (TVR) aims to align relevant video content with natural language queries. To date, most state-of-the-art TVR methods learn image-to-video transfer learning based on large-scale pre-trained visionlanguage models (e.g., CLIP). However, fully fine-tuning these pre-trained models for TVR incurs prohibitively expensive computation costs. To this end, we propose to conduct efficient text-video Retrieval with a sparse-andcorrelated AdaPter (RAP), i.e., fine-tuning the pre-trained model with a few parameterized layers. To accommodate the text-video scenario, we equip our RAP with two indispensable characteristics: temporal sparsity and correlation. Specifically, we propose a low-rank modulation module to refine the per-image features from the frozen CLIP backbone, which accentuates salient frames within the video features while alleviating temporal redundancy. Besides, we introduce an asynchronous self-attention mechanism that first selects the top responsive visual patches and augments the correlation modeling between them with learnable temporal and patch offsets. Extensive experiments on four TVR datasets demonstrate that RAP achieves superior or comparable performance compared to the fully fine-tuned counterpart and other parameter-efficient fine-tuning methods.

LLMBind: A Unified Modality-Task Integration Framework

While recent progress in multimodal large language models tackles various modality tasks, they posses limited integration capabilities for complex multi-modality tasks, consequently constraining the development of the field. In this work, we take the initiative to explore and propose the LLMBind, a unified framework for modality task integration, which binds Large Language Models and corresponding pre-trained task models with task-specific tokens. Consequently, LLMBind can interpret inputs and produce outputs in versatile combinations of image, text, video, and audio. Specifically, we introduce a Mixture-of-Experts technique to enable effective learning for different multimodal tasks through collaboration among diverse experts. Furthermore, we create a multi-task dataset comprising 400k instruction data, which unlocks the ability for interactive visual generation and editing tasks. Extensive experiments show the effectiveness of our framework across various tasks, including image, video, audio generation, image segmentation, and image editing. More encouragingly, our framework can be easily extended to other modality tasks, showcasing the promising potential of creating a unified AI agent for modeling universal modalities.

SPHINX-X: Scaling Data and Parameters for a Family of Multi-modal Large Language Models

We propose SPHINX-X, an extensive Multimodality Large Language Model (MLLM) series developed upon SPHINX. To improve the architecture and training efficiency, we modify the SPHINX framework by removing redundant visual encoders, bypassing fully-padded sub-images with skip tokens, and simplifying multi-stage training into a one-stage all-in-one paradigm. To fully unleash the potential of MLLMs, we assemble a comprehensive multi-domain and multimodal dataset covering publicly available resources in language, vision, and vision-language tasks. We further enrich this collection with our curated OCR intensive and Set-of-Mark datasets, extending the diversity and generality. By training over different base LLMs including TinyLlama1.1B, InternLM2-7B, LLaMA2-13B, and Mixtral8x7B, we obtain a spectrum of MLLMs that vary in parameter size and multilingual capabilities. Comprehensive benchmarking reveals a strong correlation between the multi-modal performance with the data and parameter scales. Code and models are released at https://github.com/Alpha-VLLM/LLaMA2-Accessory

MoE-LLaVA: Mixture of Experts for Large Vision-Language Models

Recent advances demonstrate that scaling Large Vision-Language Models (LVLMs) effectively improves downstream task performances. However, existing scaling methods enable all model parameters to be active for each token in the calculation, which brings massive training and inferring costs. In this work, we propose a simple yet effective training strategy MoE-Tuning for LVLMs. This strategy innovatively addresses the common issue of performance degradation in multi-modal sparsity learning, consequently constructing a sparse model with an outrageous number of parameters but a constant computational cost. Furthermore, we present the MoE-LLaVA, a MoE-based sparse LVLM architecture, which uniquely activates only the top-k experts through routers during deployment, keeping the remaining experts inactive. Extensive experiments show the significant performance of MoE-LLaVA in a variety of visual understanding and object hallucination benchmarks. Remarkably, with only approximately 3B sparsely activated parameters, MoE-LLaVA demonstrates performance comparable to the LLaVA-1.5-7B on various visual understanding datasets and even surpasses the LLaVA-1.5-13B in object hallucination benchmark. Through MoE-LLaVA, we aim to establish a baseline for sparse LVLMs and provide valuable insights for future research in developing more efficient and effective multi-modal learning systems. Code is released at \url{https://github.com/PKU-YuanGroup/MoE-LLaVA}.

Instance Brownian Bridge as Texts for Open-vocabulary Video Instance Segmentation

Temporally locating objects with arbitrary class texts is the primary pursuit of open-vocabulary Video Instance Segmentation (VIS). Because of the insufficient vocabulary of video data, previous methods leverage image-text pretraining model for recognizing object instances by separately aligning each frame and class texts, ignoring the correlation between frames. As a result, the separation breaks the instance movement context of videos, causing inferior alignment between video and text. To tackle this issue, we propose to link frame-level instance representations as a Brownian Bridge to model instance dynamics and align bridge-level instance representation to class texts for more precisely open-vocabulary VIS (BriVIS). Specifically, we build our system upon a frozen video segmentor to generate frame-level instance queries, and design Temporal Instance Resampler (TIR) to generate queries with temporal context from frame queries. To mold instance queries to follow Brownian bridge and accomplish alignment with class texts, we design Bridge-Text Alignment (BTA) to learn discriminative bridge-level representations of instances via contrastive objectives. Setting MinVIS as the basic video segmentor, BriVIS surpasses the Open-vocabulary SOTA (OV2Seg) by a clear margin. For example, on the challenging large-vocabulary VIS dataset (BURST), BriVIS achieves 7.43 mAP and exhibits 49.49% improvement compared to OV2Seg (4.97 mAP).