Large context window is a desirable feature in large language models (LLMs). However, due to high fine-tuning costs, scarcity of long texts, and catastrophic values introduced by new token positions, current extended context windows are limited to around 128k tokens. This paper introduces LongRoPE that, for the first time, extends the context window of pre-trained LLMs to an impressive 2048k tokens, with up to only 1k fine-tuning steps at within 256k training lengths, while maintaining performance at the original short context window. This is achieved by three key innovations: (i) we identify and exploit two forms of non-uniformities in positional interpolation through an efficient search, providing a better initialization for fine-tuning and enabling an 8x extension in non-fine-tuning scenarios; (ii) we introduce a progressive extension strategy that first fine-tunes a 256k length LLM and then conducts a second positional interpolation on the fine-tuned extended LLM to achieve a 2048k context window; (iii) we readjust LongRoPE on 8k length to recover the short context window performance. Extensive experiments on LLaMA2 and Mistral across various tasks demonstrate the effectiveness of our method. Models extended via LongRoPE retain the original architecture with minor modifications to the positional embedding, and can reuse most pre-existing optimizations.
Large language models (LLMs) have shown impressive capabilities in various tasks, yet they still struggle with math reasoning. Despite efforts to optimize Chain-of-Thoughts (CoT) prompts and fine-tune LLMs, the potential of few-shot learning remains unexplored. In this work, we propose CoT-Influx, a novel approach pushing the boundaries of few-shot CoT learning to improve LLM math reasoning capabilities. CoT-Influx addresses the challenges of the selection of useful examples and limited number of examples due to restricted context window length. Inspired by our observation that natural language inputs contain many redundancy, we propose a coarse-to-fine pruner as a plug-and-play module for LLMs, which first identifies as many crucial CoT examples as possible and then further prunes unimportant tokens within the context window. To train the pruner, we collect a math reasoning dataset with diverse difficulty and steps, introduce a reward to measure both the input's effectiveness for math reasoning and token length constraints, and propose a novel training approach with reinforcement learning. As a result, CoT-Influx significantly outperforms CoT and few-shot prompting baselines across various LLMs (LLaMA2-7B, 13B, 70B) and 5 mathematical datasets, achieving up to 4.55% absolute improvements. Remarkably, without any fine-tuning, LLaMA2-70B with CoT-Influx surpasses GPT-3.5 and a wide range of larger LLMs (PaLM, Minerva, etc.) on the GSM8K.
Despite the remarkable success of Large Language Models (LLMs), the massive size poses significant deployment challenges, particularly on resource-constrained hardware. While existing LLM compression methods focus on quantization, pruning remains relatively unexplored due to the high cost of training-based approaches and data collection challenges. One-shot pruning methods, although cost-effective and data-free, have become dominant in LLM pruning, but lead to performance decline under the structured pruning setting. In this work, we introduce a new paradigm for structurally pruning LLMs, called Compresso. Our approach, through the collaboration of the proposed resource-efficient pruning algorithm and the LLM itself, learns optimal pruning decisions during the training process. Compresso addresses the challenges of expensive training costs and data collection by incorporating Low-Rank Adaptation (LoRA) into the $L_0$ regularization during the instruction tuning process. Then, we further augment the pruning algorithm by introducing a collaborative prompt that fosters collaboration between the LLM and the pruning algorithm, significantly boosting the overall performance. To this end, Compresso prunes LLaMA-7B to 5.4B, maintaining original performance and even surpassing LLaMA-7B in reading comprehension by 2.62%. Extensive experiments demonstrate that Compresso significantly outperforms one-shot pruning baselines across various sparsity ratios, achieving up to 2.21%, 11.43%, 7.04%, and 4.81% higher scores on the commonsense reasoning, reading comprehension, MMLU, and BBH benchmarks, respectively.
Embedding-based retrieval methods construct vector indices to search for document representations that are most similar to the query representations. They are widely used in document retrieval due to low latency and decent recall performance. Recent research indicates that deep retrieval solutions offer better model quality, but are hindered by unacceptable serving latency and the inability to support document updates. In this paper, we aim to enhance the vector index with end-to-end deep generative models, leveraging the differentiable advantages of deep retrieval models while maintaining desirable serving efficiency. We propose Model-enhanced Vector Index (MEVI), a differentiable model-enhanced index empowered by a twin-tower representation model. MEVI leverages a Residual Quantization (RQ) codebook to bridge the sequence-to-sequence deep retrieval and embedding-based models. To substantially reduce the inference time, instead of decoding the unique document ids in long sequential steps, we first generate some semantic virtual cluster ids of candidate documents in a small number of steps, and then leverage the well-adapted embedding vectors to further perform a fine-grained search for the relevant documents in the candidate virtual clusters. We empirically show that our model achieves better performance on the commonly used academic benchmarks MSMARCO Passage and Natural Questions, with comparable serving latency to dense retrieval solutions.
Web applications are increasingly becoming the primary platform for AI service delivery, making in-browser deep learning (DL) inference more prominent. However, current in-browser inference systems fail to effectively utilize advanced web programming techniques and customize kernels for various client devices, leading to suboptimal performance. To address the issues, this paper presents the first in-browser inference system, nn-JIT.web, which enables just-in-time (JIT) auto-generation of optimized kernels for both CPUs and GPUs during inference. The system achieves this by using two novel web programming techniques that can significantly reduce kernel generation time, compared to other tensor compilers such as TVM, while maintaining or even improving performance. The first technique, Tensor-Web Compiling Co-Design, lowers compiling costs by unifying tensor and web compiling and eliminating redundant and ineffective compiling passes. The second technique, Web-Specific Lite Kernel Optimization Space Design, reduces kernel tuning costs by focusing on web programming requirements and efficient hardware resource utilization, limiting the optimization space to only dozens. nn-JIT.web is evaluated for modern transformer models on a range of client devices, including the mainstream CPUs and GPUs from ARM, Intel, AMD and Nvidia. Results show that nn-JIT.web can achieve up to 8.2x faster within 30 seconds compared to the baselines across various models.
Large language models (LLMs) based on transformers have made significant strides in recent years, the success of which is driven by scaling up their model size. Despite their high algorithmic performance, the computational and memory requirements of LLMs present unprecedented challenges. To tackle the high compute requirements of LLMs, the Mixture-of-Experts (MoE) architecture was introduced which is able to scale its model size without proportionally scaling up its computational requirements. Unfortunately, MoE's high memory demands and dynamic activation of sparse experts restrict its applicability to real-world problems. Previous solutions that offload MoE's memory-hungry expert parameters to CPU memory fall short because the latency to migrate activated experts from CPU to GPU incurs high performance overhead. Our proposed Pre-gated MoE system effectively tackles the compute and memory challenges of conventional MoE architectures using our algorithm-system co-design. Pre-gated MoE employs our novel pre-gating function which alleviates the dynamic nature of sparse expert activation, allowing our proposed system to address the large memory footprint of MoEs while also achieving high performance. We demonstrate that Pre-gated MoE is able to improve performance, reduce GPU memory consumption, while also maintaining the same level of model quality. These features allow our Pre-gated MoE system to cost-effectively deploy large-scale LLMs using just a single GPU with high performance.
Deploying pre-trained transformer models like BERT on downstream tasks in resource-constrained scenarios is challenging due to their high inference cost, which grows rapidly with input sequence length. In this work, we propose a constraint-aware and ranking-distilled token pruning method ToP, which selectively removes unnecessary tokens as input sequence passes through layers, allowing the model to improve online inference speed while preserving accuracy. ToP overcomes the limitation of inaccurate token importance ranking in the conventional self-attention mechanism through a ranking-distilled token distillation technique, which distills effective token rankings from the final layer of unpruned models to early layers of pruned models. Then, ToP introduces a coarse-to-fine pruning approach that automatically selects the optimal subset of transformer layers and optimizes token pruning decisions within these layers through improved $L_0$ regularization. Extensive experiments on GLUE benchmark and SQuAD tasks demonstrate that ToP outperforms state-of-the-art token pruning and model compression methods with improved accuracy and speedups. ToP reduces the average FLOPs of BERT by 8.1x while achieving competitive accuracy on GLUE, and provides a real latency speedup of up to 7.4x on an Intel CPU.
Automatic Speech Recognition (ASR) has seen remarkable advancements with deep neural networks, such as Transformer and Conformer. However, these models typically have large model sizes and high inference costs, posing a challenge to deploy on resource-limited devices. In this paper, we propose a novel compression strategy that leverages structured pruning and knowledge distillation to reduce the model size and inference cost of the Conformer model while preserving high recognition performance. Our approach utilizes a set of binary masks to indicate whether to retain or prune each Conformer module, and employs L0 regularization to learn the optimal mask values. To further enhance pruning performance, we use a layerwise distillation strategy to transfer knowledge from unpruned to pruned models. Our method outperforms all pruning baselines on the widely used LibriSpeech benchmark, achieving a 50% reduction in model size and a 28% reduction in inference cost with minimal performance loss.
Efficient deployment of large language models (LLMs) necessitates low-bit quantization to minimize model size and inference cost. While low-bit integer formats (e.g., INT8/INT4) have been the conventional choice, emerging low-bit floating-point formats (e.g., FP8/FP4) offer a compelling alternative and are gaining support from cutting-edge hardware, such as NVIDIA's H100 GPU. However, the superiority of low-bit INT versus FP formats for quantization on LLMs remains unclear. In this study, we conduct a comparative analysis of INT and FP quantization with the same bit-width, revealing that the optimal quantization format varies across different layers due to the complexity and diversity of tensor distribution. Consequently, we advocate the Mixture of Formats Quantization (MoFQ), which selects the optimal format on a layer-wise basis. This simple yet effective approach achieves state-of-the-art results in both weight-only (W-only) and weight-activation (WA) post-training quantization scenarios when tested on LLaMA across various tasks. In 4-bit W-only quantization, MoFQ surpasses GPTQ without complex hyperparameter tuning and with an order of magnitude faster quantization speed. While in 8-bit WA quantization, MoFQ significantly outperforms INT/FP-only methods, achieving performance close to the full precision model. Notably, MoFQ incurs no hardware overhead compared to INT/FP-only quantization, as the bit-width remains unchanged.
While generative modeling has been ubiquitous in natural language processing and computer vision, its application to image retrieval remains unexplored. In this paper, we recast image retrieval as a form of generative modeling by employing a sequence-to-sequence model, contributing to the current unified theme. Our framework, IRGen, is a unified model that enables end-to-end differentiable search, thus achieving superior performance thanks to direct optimization. While developing IRGen we tackle the key technical challenge of converting an image into quite a short sequence of semantic units in order to enable efficient and effective retrieval. Empirical experiments demonstrate that our model yields significant improvement over three commonly used benchmarks, for example, 22.9\% higher than the best baseline method in precision@10 on In-shop dataset with comparable recall@10 score.