Photon: Federated LLM Pre-Training

Scaling large language models (LLMs) demands extensive data and computing resources, which are traditionally constrained to data centers by the high-bandwidth requirements of distributed training. Low-bandwidth methods like federated learning (FL) could enable collaborative training of larger models across weakly-connected GPUs if they can effectively be used for pre-training. To achieve this, we introduce Photon, the first complete system for federated end-to-end LLM training, leveraging cross-silo FL for global-scale training with minimal communication overheads. Using Photon, we train the first federated family of decoder-only LLMs from scratch. We show that: (1) Photon can train model sizes up to 7B in a federated fashion while reaching an even better perplexity than centralized pre-training; (2) Photon model training time decreases with available compute, achieving a similar compute-time trade-off to centralized; and (3) Photon outperforms the wall-time of baseline distributed training methods by 35% via communicating 64x-512xless. Our proposal is robust to data heterogeneity and converges twice as fast as previous methods like DiLoCo. This surprising data efficiency stems from a unique approach combining small client batch sizes with extremely high learning rates, enabled by federated averaging's robustness to hyperparameters. Photon thus represents the first economical system for global internet-wide LLM pre-training.

MoDification: Mixture of Depths Made Easy

Long-context efficiency has recently become a trending topic in serving large language models (LLMs). And mixture of depths (MoD) is proposed as a perfect fit to bring down both latency and memory. In this paper, however, we discover that MoD can barely transform existing LLMs without costly training over an extensive number of tokens. To enable the transformations from any LLMs to MoD ones, we showcase top-k operator in MoD should be promoted to threshold-p operator, and refinement to architecture and data should also be crafted along. All these designs form our method termed MoDification. Through a comprehensive set of experiments covering model scales from 3B to 70B, we exhibit MoDification strikes an excellent balance between efficiency and effectiveness. MoDification can achieve up to ~1.2x speedup in latency and ~1.8x reduction in memory compared to original LLMs especially in long-context applications.

AMPO: Automatic Multi-Branched Prompt Optimization

Prompt engineering is very important to enhance the performance of large language models (LLMs). When dealing with complex issues, prompt engineers tend to distill multiple patterns from examples and inject relevant solutions to optimize the prompts, achieving satisfying results. However, existing automatic prompt optimization techniques are only limited to producing single flow instructions, struggling with handling diverse patterns. In this paper, we present AMPO, an automatic prompt optimization method that can iteratively develop a multi-branched prompt using failure cases as feedback. Our goal is to explore a novel way of structuring prompts with multi-branches to better handle multiple patterns in complex tasks, for which we introduce three modules: Pattern Recognition, Branch Adjustment, and Branch Pruning. In experiments across five tasks, AMPO consistently achieves the best results. Additionally, our approach demonstrates significant optimization efficiency due to our adoption of a minimal search strategy.

StraGo: Harnessing Strategic Guidance for Prompt Optimization

Prompt engineering is pivotal for harnessing the capabilities of large language models (LLMs) across diverse applications. While existing prompt optimization methods improve prompt effectiveness, they often lead to prompt drifting, where newly generated prompts can adversely impact previously successful cases while addressing failures. Furthermore, these methods tend to rely heavily on LLMs' intrinsic capabilities for prompt optimization tasks. In this paper, we introduce StraGo (Strategic-Guided Optimization), a novel approach designed to mitigate prompt drifting by leveraging insights from both successful and failed cases to identify critical factors for achieving optimization objectives. StraGo employs a how-to-do methodology, integrating in-context learning to formulate specific, actionable strategies that provide detailed, step-by-step guidance for prompt optimization. Extensive experiments conducted across a range of tasks, including reasoning, natural language understanding, domain-specific knowledge, and industrial applications, demonstrate StraGo's superior performance. It establishes a new state-of-the-art in prompt optimization, showcasing its ability to deliver stable and effective prompt improvements.

DEPT: Decoupled Embeddings for Pre-training Language Models

Language Model pre-training benefits from a broader data mixture to enhance performance across domains and languages. However, training on such heterogeneous text corpora is complex, requiring extensive and cost-intensive efforts. Since these data sources vary in lexical, syntactic, and semantic aspects, they cause negative interference or the "curse of multilinguality". We propose a novel pre-training framework to alleviate this curse. Our method, DEPT, decouples the embedding layers from the transformer body while simultaneously training the latter in multiple contexts. DEPT enables the model to train without being bound to a shared global vocabulary. DEPT: (1) can train robustly and effectively under significant data heterogeneity, (2) reduces the parameter count of the token embeddings by up to 80% and the communication costs by 675x for billion-scale models (3) enhances model generalization and plasticity in adapting to new languages and domains, and (4) allows training with custom optimized vocabulary per data source. We prove DEPT's potential by performing the first vocabulary-agnostic federated multilingual pre-training of a 1.3 billion-parameter model across high and low-resource languages, reducing its parameter count by 409 million.

VideoLLM-MoD: Efficient Video-Language Streaming with Mixture-of-Depths Vision Computation

A well-known dilemma in large vision-language models (e.g., GPT-4, LLaVA) is that while increasing the number of vision tokens generally enhances visual understanding, it also significantly raises memory and computational costs, especially in long-term, dense video frame streaming scenarios. Although learnable approaches like Q-Former and Perceiver Resampler have been developed to reduce the vision token burden, they overlook the context causally modeled by LLMs (i.e., key-value cache), potentially leading to missed visual cues when addressing user queries. In this paper, we introduce a novel approach to reduce vision compute by leveraging redundant vision tokens "skipping layers" rather than decreasing the number of vision tokens. Our method, VideoLLM-MoD, is inspired by mixture-of-depths LLMs and addresses the challenge of numerous vision tokens in long-term or streaming video. Specifically, for each transformer layer, we learn to skip the computation for a high proportion (e.g., 80\%) of vision tokens, passing them directly to the next layer. This approach significantly enhances model efficiency, achieving approximately \textasciitilde42\% time and \textasciitilde30\% memory savings for the entire training. Moreover, our method reduces the computation in the context and avoid decreasing the vision tokens, thus preserving or even improving performance compared to the vanilla model. We conduct extensive experiments to demonstrate the effectiveness of VideoLLM-MoD, showing its state-of-the-art results on multiple benchmarks, including narration, forecasting, and summarization tasks in COIN, Ego4D, and Ego-Exo4D datasets.

Understanding the RoPE Extensions of Long-Context LLMs: An Attention Perspective

Enabling LLMs to handle lengthy context is currently a research hotspot. Most LLMs are built upon rotary position embedding (RoPE), a popular position encoding method. Therefore, a prominent path is to extrapolate the RoPE trained on comparably short texts to far longer texts. A heavy bunch of efforts have been dedicated to boosting the extrapolation via extending the formulations of the RoPE, however, few of them have attempted to showcase their inner workings comprehensively. In this paper, we are driven to offer a straightforward yet in-depth understanding of RoPE extensions from an attention perspective and on two benchmarking tasks. A broad array of experiments reveals several valuable findings: 1) Maintaining attention patterns to those at the pretrained length improves extrapolation; 2) Large attention uncertainty leads to retrieval errors; 3) Using longer continual pretraining lengths for RoPE extensions could reduce attention uncertainty and significantly enhance extrapolation.

DetectBench: Can Large Language Model Detect and Piece Together Implicit Evidence?

Detecting evidence within the context is a key step in the process of reasoning task. Evaluating and enhancing the capabilities of LLMs in evidence detection will strengthen context-based reasoning performance. This paper proposes a benchmark called DetectBench for verifying the ability to detect and piece together implicit evidence within a long context. DetectBench contains 3,928 multiple-choice questions, with an average of 994 tokens per question. Each question contains an average of 4.55 pieces of implicit evidence, and solving the problem typically requires 7.62 logical jumps to find the correct answer. To enhance the performance of LLMs in evidence detection, this paper proposes Detective Reasoning Prompt and Finetune. Experiments demonstrate that the existing LLMs' abilities to detect evidence in long contexts are far inferior to humans. However, the Detective Reasoning Prompt effectively enhances the capability of powerful LLMs in evidence detection, while the Finetuning method shows significant effects in enhancing the performance of weaker LLMs. Moreover, when the abilities of LLMs in evidence detection are improved, their final reasoning performance is also enhanced accordingly.

Vript: A Video Is Worth Thousands of Words

Advancements in multimodal learning, particularly in video understanding and generation, require high-quality video-text datasets for improved model performance. Vript addresses this issue with a meticulously annotated corpus of 12K high-resolution videos, offering detailed, dense, and script-like captions for over 420K clips. Each clip has a caption of ~145 words, which is over 10x longer than most video-text datasets. Unlike captions only documenting static content in previous datasets, we enhance video captioning to video scripting by documenting not just the content, but also the camera operations, which include the shot types (medium shot, close-up, etc) and camera movements (panning, tilting, etc). By utilizing the Vript, we explore three training paradigms of aligning more text with the video modality rather than clip-caption pairs. This results in Vriptor, a top-performing video captioning model among open-source models, comparable to GPT-4V in performance. Vriptor is also a powerful model capable of end-to-end generation of dense and detailed captions for long videos. Moreover, we introduce Vript-Hard, a benchmark consisting of three video understanding tasks that are more challenging than existing benchmarks: Vript-HAL is the first benchmark evaluating action and object hallucinations in video LLMs, Vript-RR combines reasoning with retrieval resolving question ambiguity in long-video QAs, and Vript-ERO is a new task to evaluate the temporal understanding of events in long videos rather than actions in short videos in previous works. All code, models, and datasets are available in https://github.com/mutonix/Vript.

Efficient k-Nearest-Neighbor Machine Translation with Dynamic Retrieval

To achieve non-parametric NMT domain adaptation, $k$-Nearest-Neighbor Machine Translation ($k$NN-MT) constructs an external datastore to store domain-specific translation knowledge, which derives a $k$NN distribution to interpolate the prediction distribution of the NMT model via a linear interpolation coefficient $\lambda$. Despite its success, $k$NN retrieval at each timestep leads to substantial time overhead. To address this issue, dominant studies resort to $k$NN-MT with adaptive retrieval ($k$NN-MT-AR), which dynamically estimates $\lambda$ and skips $k$NN retrieval if $\lambda$ is less than a fixed threshold. Unfortunately, $k$NN-MT-AR does not yield satisfactory results. In this paper, we first conduct a preliminary study to reveal two key limitations of $k$NN-MT-AR: 1) the optimization gap leads to inaccurate estimation of $\lambda$ for determining $k$NN retrieval skipping, and 2) using a fixed threshold fails to accommodate the dynamic demands for $k$NN retrieval at different timesteps. To mitigate these limitations, we then propose $k$NN-MT with dynamic retrieval ($k$NN-MT-DR) that significantly extends vanilla $k$NN-MT in two aspects. Firstly, we equip $k$NN-MT with a MLP-based classifier for determining whether to skip $k$NN retrieval at each timestep. Particularly, we explore several carefully-designed scalar features to fully exert the potential of the classifier. Secondly, we propose a timestep-aware threshold adjustment method to dynamically generate the threshold, which further improves the efficiency of our model. Experimental results on the widely-used datasets demonstrate the effectiveness and generality of our model.\footnote{Our code is available at \url{https://github.com/DeepLearnXMU/knn-mt-dr}.