Structured pruning fundamentally reduces computational and memory overheads of large language models (LLMs) and offers a feasible solution for end-side LLM deployment. Structurally pruned models remain dense and high-precision, highly compatible with further tuning and compression. However, as the coarse-grained structured pruning poses large damage to the highly interconnected model, achieving a high compression ratio for scaled-up LLMs remains a challenge. In this paper, we introduce a task-agnostic structured pruning approach coupled with a compact Transformer architecture design. The proposed approach, named TransAct, reduces transitional activations inside multi-head attention (MHA) and multi-layer perceptron (MLP) modules, while preserving the inter-module activations that are sensitive to perturbations. Hence, the LLM is pruned into an intra-module low-rank architecture, significantly reducing weights, KV Cache and attention computation. TransAct is implemented on the LLaMA model and evaluated on downstream benchmarks. Results verify the optimality of our approach at high compression with respect to both efficiency and performance. Further, ablation studies reveal the strength of activation-guided iterative pruning and provide experimental analysis on the redundancy of MHA and MLP modules.
相關內容
The inherent diversity of computation types within the deep neural network (DNN) models often requires a variety of specialized units in hardware processors, which limits computational efficiency, increasing both inference latency and power consumption, especially when the hardware processor needs to support and execute different neural networks. In this study, we introduce NeuralMatrix, which elastically transforms the computations of entire DNNs into linear matrix operations. This transformation allows seamless execution of various DNN models all with matrix operations and paves the way for running versatile DNN models with a single General Matrix Multiplication (GEMM) accelerator.Extensive experiments with both CNN and transformer-based models demonstrate the potential of NeuralMatrix to accurately and efficiently execute a wide range of DNN models, achieving 2.17-38.72 times computation efficiency (i.e., throughput per power) compared to CPUs, GPUs, and SoC platforms. This level of efficiency is usually only attainable with the accelerator designed for a specific neural network.
Audio-LLM introduces audio modality into a large language model (LLM) to enable a powerful LLM to recognize, understand, and generate audio. However, during speech recognition in noisy environments, we observed the presence of illusions and repetition issues in audio-LLM, leading to substitution and insertion errors. This paper proposes a transcription prompt-based audio-LLM by introducing an ASR expert as a transcription tokenizer and a hybrid Autoregressive (AR) Non-autoregressive (NAR) decoding approach to solve the above problems. Experiments on 10k-hour WenetSpeech Mandarin corpus show that our approach decreases 12.2% and 9.6% CER relatively on Test_Net and Test_Meeting evaluation sets compared with baseline. Notably, we reduce the decoding repetition rate on the evaluation set to zero, showing that the decoding repetition problem has been solved fundamentally.
As large language models (LLMs) expand the power of natural language processing to handle long inputs, rigorous and systematic analyses are necessary to understand their abilities and behavior. A salient application is summarization, due to its ubiquity and controversy (e.g., researchers have declared the death of summarization). In this paper, we use financial report summarization as a case study because financial reports are not only long but also use numbers and tables extensively. We propose a computational framework for characterizing multimodal long-form summarization and investigate the behavior of Claude 2.0/2.1, GPT-4/3.5, and Cohere. We find that GPT-3.5 and Cohere fail to perform this summarization task meaningfully. For Claude 2 and GPT-4, we analyze the extractiveness of the summary and identify a position bias in LLMs. This position bias disappears after shuffling the input for Claude, which suggests that Claude seems to recognize important information. We also conduct a comprehensive investigation on the use of numeric data in LLM-generated summaries and offer a taxonomy of numeric hallucination. We employ prompt engineering to improve GPT-4's use of numbers with limited success. Overall, our analyses highlight the strong capability of Claude 2 in handling long multimodal inputs compared to GPT-4. The generated summaries and evaluation code are available at //github.com/ChicagoHAI/characterizing-multimodal-long-form-summarization.
As AI chips incorporate numerous parallelized cores to scale deep learning (DL) computing, inter-core communication is enabled recently by employing high-bandwidth and low-latency interconnect links on the chip (e.g., Graphcore IPU). It allows each core to directly access the fast scratchpad memory in other cores, which enables new parallel computing paradigms. However, without proper support for the scalable inter-core connections in current DL compilers, it is hard for developers to exploit the benefits of this new architecture. We present T10, the first DL compiler to exploit the inter-core communication bandwidth and distributed on-chip memory on AI chips. To formulate the computation and communication patterns of tensor operators in this new architecture, T10 introduces a distributed tensor abstraction rTensor. T10 maps a DNN model to execution plans with a generalized compute-shift pattern, by partitioning DNN computation into sub-operators and mapping them to cores, so that the cores can exchange data following predictable patterns. T10 makes globally optimized trade-offs between on-chip memory consumption and inter-core communication overhead, selects the best execution plan from a vast optimization space, and alleviates unnecessary inter-core communications. Our evaluation with a real inter-core connected AI chip, the Graphcore IPU, shows up to 3.3$\times$ performance improvement, and scalability support for larger models, compared to state-of-the-art DL compilers and vendor libraries.
Despite the impressive performance on information-seeking tasks, large language models (LLMs) still struggle with hallucinations. Attributed LLMs, which augment generated text with in-line citations, have shown potential in mitigating hallucinations and improving verifiability. However, current approaches suffer from suboptimal citation quality due to their reliance on in-context learning. Furthermore, the practice of citing only coarse document identifiers makes it challenging for users to perform fine-grained verification. In this work, we introduce FRONT, a training framework designed to teach LLMs to generate Fine-Grained Grounded Citations. By grounding model outputs in fine-grained supporting quotes, these quotes guide the generation of grounded and consistent responses, not only improving citation quality but also facilitating fine-grained verification. Experiments on the ALCE benchmark demonstrate the efficacy of FRONT in generating superior grounded responses and highly supportive citations. With LLaMA-2-7B, the framework significantly outperforms all the baselines, achieving an average of 14.21% improvement in citation quality across all datasets, even surpassing ChatGPT.
This paper presents the first study to explore the potential of parameter quantization for multimodal large language models to alleviate the significant resource constraint encountered during vision-language instruction tuning. We introduce a Quantization-aware Scale LeArning method based on multimodal Warmup, termed QSLAW. This method is grounded in two key innovations: (1) The learning of group-wise scale factors for quantized LLM weights to mitigate the quantization error arising from activation outliers and achieve more effective vision-language instruction tuning; (2) The implementation of a multimodal warmup that progressively integrates linguistic and multimodal training samples, thereby preventing overfitting of the quantized model to multimodal data while ensuring stable adaptation of multimodal large language models to downstream vision-language tasks. Extensive experiments demonstrate that models quantized by QSLAW perform on par with, or even surpass, their full-precision counterparts, while facilitating up to 1.4 times reduction in VL tuning time and GPU consumption. Our code is released at //github.com/xjjxmu/QSLAW.
The ever-growing popularity of large language models (LLMs) has resulted in their increasing adoption for hardware design and verification. Prior research has attempted to assess the capability of LLMs to automate digital hardware design by producing superior-quality Register Transfer Logic (RTL) descriptions, particularly in Verilog. However, these tests have revealed that Verilog code production using LLMs at current state-of-the-art lack sufficient functional correctness to be practically viable, compared to automatic generation of programs in general-purpose programming languages such as C, C++, Python, etc. With this as the key insight, in this paper we assess the performance of a two-stage software pipeline for automated Verilog RTL generation: LLM based automatic generation of annotated C++ code suitable for high-level synthesis (HLS), followed by HLS to generate Verilog RTL. We have benchmarked the performance of our proposed scheme using the open-source VerilogEval dataset, for four different industry-scale LLMs, and the Vitis HLS tool. Our experimental results demonstrate that our two-step technique substantially outperforms previous proposed techniques of direct Verilog RTL generation by LLMs in terms of average functional correctness rates, reaching score of 0.86 in pass@1 metric.
A major factor in the recent success of large language models is the use of enormous and ever-growing text datasets for unsupervised pre-training. However, naively training a model on all available data may not be optimal (or feasible), as the quality of available text data can vary. Filtering out data can also decrease the carbon footprint and financial costs of training models by reducing the amount of training required. Data selection methods aim to determine which candidate data points to include in the training dataset and how to appropriately sample from the selected data points. The promise of improved data selection methods has caused the volume of research in the area to rapidly expand. However, because deep learning is mostly driven by empirical evidence and experimentation on large-scale data is expensive, few organizations have the resources for extensive data selection research. Consequently, knowledge of effective data selection practices has become concentrated within a few organizations, many of which do not openly share their findings and methodologies. To narrow this gap in knowledge, we present a comprehensive review of existing literature on data selection methods and related research areas, providing a taxonomy of existing approaches. By describing the current landscape of research, this work aims to accelerate progress in data selection by establishing an entry point for new and established researchers. Additionally, throughout this review we draw attention to noticeable holes in the literature and conclude the paper by proposing promising avenues for future research.
Retrieval-Augmented Generation (RAG) merges retrieval methods with deep learning advancements to address the static limitations of large language models (LLMs) by enabling the dynamic integration of up-to-date external information. This methodology, focusing primarily on the text domain, provides a cost-effective solution to the generation of plausible but incorrect responses by LLMs, thereby enhancing the accuracy and reliability of their outputs through the use of real-world data. As RAG grows in complexity and incorporates multiple concepts that can influence its performance, this paper organizes the RAG paradigm into four categories: pre-retrieval, retrieval, post-retrieval, and generation, offering a detailed perspective from the retrieval viewpoint. It outlines RAG's evolution and discusses the field's progression through the analysis of significant studies. Additionally, the paper introduces evaluation methods for RAG, addressing the challenges faced and proposing future research directions. By offering an organized framework and categorization, the study aims to consolidate existing research on RAG, clarify its technological underpinnings, and highlight its potential to broaden the adaptability and applications of LLMs.
Deploying large language models (LLMs) is challenging because they are memory inefficient and compute-intensive for practical applications. In reaction, researchers train smaller task-specific models by either finetuning with human labels or distilling using LLM-generated labels. However, finetuning and distillation require large amounts of training data to achieve comparable performance to LLMs. We introduce Distilling step-by-step, a new mechanism that (a) trains smaller models that outperform LLMs, and (b) achieves so by leveraging less training data needed by finetuning or distillation. Our method extracts LLM rationales as additional supervision for small models within a multi-task training framework. We present three findings across 4 NLP benchmarks: First, compared to both finetuning and distillation, our mechanism achieves better performance with much fewer labeled/unlabeled training examples. Second, compared to LLMs, we achieve better performance using substantially smaller model sizes. Third, we reduce both the model size and the amount of data required to outperform LLMs; our 770M T5 model outperforms the 540B PaLM model using only 80% of available data on a benchmark task.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191