Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but introduce significant challenges in real-world serving due to inefficient use of the expensive, computation-optimized accelerators. This mismatch arises from the autoregressive nature of LLMs, where the generation phase comprises operators with varying resource demands. Specifically, the attention operator is memory-intensive, exhibiting a memory access pattern that clashes with the strengths of modern accelerators, especially as context length increases. To enhance the efficiency and cost-effectiveness of LLM serving, we introduce the concept of attention offloading. This approach leverages a collection of cheap, memory-optimized devices for the attention operator while still utilizing high-end accelerators for other parts of the model. This heterogeneous setup ensures that each component is tailored to its specific workload, maximizing overall performance and cost efficiency. Our comprehensive analysis and experiments confirm the viability of splitting the attention computation over multiple devices. Also, the communication bandwidth required between heterogeneous devices proves to be manageable with prevalent networking technologies. To further validate our theory, we develop Lamina, an LLM inference system that incorporates attention offloading. Experimental results indicate that Lamina can provide 1.48x-12.1x higher estimated throughput per dollar than homogeneous solutions.
相關內容
Exploiting large language models (LLMs) to tackle reasoning has garnered growing attention. It still remains highly challenging to achieve satisfactory results in complex logical problems, characterized by plenty of premises within the prompt and requiring multi-hop reasoning. In particular, the reasoning capabilities of LLMs are brittle to disorder and distractibility. In this work, we first examine the mechanism from the perspective of information flow and reveal that LLMs exhibit failure patterns akin to human-like cognitive biases when dealing with disordered and irrelevant content in reasoning tasks. However, in contrast to LLMs, disordered and irrelevant content does not significantly decrease human performance, as humans have a propensity to distill the most relevant information and systematically organize their thoughts, aiding them in responding to questions. Stem from that, we further propose a novel reasoning approach named Concise and Organized Perception (COP). COP carefully analyzes the given statements to identify the most pertinent information while eliminating redundancy efficiently. It then prompts the LLMs in a more organized form that adapts to the model's inference process. By perceiving concise and organized context, the reasoning abilities of LLMs can be better elicited. Extensive experimental results on several popular logical benchmarks (ProofWriter, PrOntoQA, PrOntoQA-OOD, and FOLIO) and math benchmark (DI-GSM) show that COP significantly outperforms previous state-of-the-art methods.
Non-autoregressive (NAR) automatic speech recognition (ASR) models predict tokens independently and simultaneously, bringing high inference speed. However, there is still a gap in the accuracy of the NAR models compared to the autoregressive (AR) models. To further narrow the gap between the NAR and AR models, we propose a single-step NAR ASR architecture with high accuracy and inference speed, called EfficientASR. It uses an Index Mapping Vector (IMV) based alignment generator to generate alignments during training, and an alignment predictor to learn the alignments for inference. It can be trained end-to-end (E2E) with cross-entropy loss combined with alignment loss. The proposed EfficientASR achieves competitive results on the AISHELL-1 and AISHELL-2 benchmarks compared to the state-of-the-art (SOTA) models. Specifically, it achieves character error rates (CER) of 4.26%/4.62% on the AISHELL-1 dev/test dataset, which outperforms the SOTA AR Conformer with about 30x inference speedup.
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating additional information from retrieval. However, studies have shown that LLMs still face challenges in effectively using the retrieved information, even ignoring it or being misled by it. The key reason is that the training of LLMs does not clearly make LLMs learn how to utilize input retrieved texts with varied quality. In this paper, we propose a novel perspective that considers the role of LLMs in RAG as ``Information Refiner'', which means that regardless of correctness, completeness, or usefulness of retrieved texts, LLMs can consistently integrate knowledge within the retrieved texts and model parameters to generate the texts that are more concise, accurate, and complete than the retrieved texts. To this end, we propose an information refinement training method named InFO-RAG that optimizes LLMs for RAG in an unsupervised manner. InFO-RAG is low-cost and general across various tasks. Extensive experiments on zero-shot prediction of 11 datasets in diverse tasks including Question Answering, Slot-Filling, Language Modeling, Dialogue, and Code Generation show that InFO-RAG improves the performance of LLaMA2 by an average of 9.39\% relative points. InFO-RAG also shows advantages in in-context learning and robustness of RAG.
We provide an overview of the emergence of large language models for scientific computing applications. We highlight use cases that involve natural language processing of scientific documents and specialized languages designed to describe physical systems. For the former, chatbot style applications appear in medicine, mathematics and physics and can be used iteratively with domain experts for problem solving. We also review specialized languages within molecular biology, the languages of molecules, proteins, and DNA where language models are being used to predict properties and even create novel physical systems at much faster rates than traditional computing methods.
Vision Transformers (ViT), when paired with large-scale pretraining, have shown remarkable performance across various computer vision tasks, primarily due to their weak inductive bias. However, while such weak inductive bias aids in pretraining scalability, this may hinder the effective adaptation of ViTs for visuo-motor control tasks as a result of the absence of control-centric inductive biases. Such absent inductive biases include spatial locality and translation equivariance bias which convolutions naturally offer. To this end, we introduce Convolution Injector (CoIn), an add-on module that injects convolutions which are rich in locality and equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control. We evaluate CoIn with three distinct types of pretrained ViTs (CLIP, MVP, VC-1) across 12 varied control tasks within three separate domains (Adroit, MetaWorld, DMC), and demonstrate that CoIn consistently enhances control task performance across all experimented environments and models, validating the effectiveness of providing pretrained ViTs with control-centric biases.
Multimodal large language models (MLLMs) fine-tuned with multimodal instruction datasets have demonstrated remarkable capabilities in multimodal tasks. However, fine-tuning all parameters of MLLMs has become challenging as they usually contain billions of parameters. To address this issue, we study parameter-efficient fine-tuning (PEFT) methods for MLLMs. We aim to identify effective methods for enhancing the performance of MLLMs in scenarios where only a limited number of parameters are trained. This paper conducts empirical studies using four popular PEFT methods to fine-tune the LLM component of open-source MLLMs. We present a comprehensive analysis that encompasses various aspects, including the impact of PEFT methods on various models, parameters and location of the PEFT module, size of fine-tuning data, model stability based on PEFT methods, MLLM's generalization, and hallucination. We evaluated four PEFT methods on seven datasets from two different categories: unseen and seen datasets. Across all experiments, we show that the adapter is the best-performing PEFT method. At the same time, fine-tuning the connector layers leads to improved performance in most MLLMs. Code and data are available at //github.com/alenai97/PEFT-MLLM.git.
Increasing the number of parameters in large language models (LLMs) usually improves performance in downstream tasks but raises compute and memory costs, making deployment difficult in resource-limited settings. Quantization techniques, which reduce the bits needed for model weights or activations with minimal performance loss, have become popular due to the rise of LLMs. However, most quantization studies use pre-trained LLMs, and the impact of quantization on instruction-tuned LLMs and the relationship between perplexity and benchmark performance of quantized LLMs are not well understood. Evaluation of quantized LLMs is often limited to language modeling and a few classification tasks, leaving their performance on other benchmarks unclear. To address these gaps, we propose a structured evaluation framework consisting of three critical dimensions: (1) knowledge \& capacity, (2) alignment, and (3) efficiency, and conduct extensive experiments across ten diverse benchmarks. Our experimental results indicate that LLMs with 4-bit quantization can retain performance comparable to their non-quantized counterparts, and perplexity can serve as a proxy metric for quantized LLMs on most benchmarks. Furthermore, quantized LLMs with larger parameter scales can outperform smaller LLMs. Despite the memory savings achieved through quantization, it can also slow down the inference speed of LLMs. Consequently, substantial engineering efforts and hardware support are imperative to achieve a balanced optimization of decoding speed and memory consumption in the context of quantized LLMs.
We introduce a low-resource safety enhancement method for aligning large language models (LLMs) without the need for supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF). Our main idea is to exploit knowledge distillation to extract the alignment information from existing well-aligned LLMs and integrate it into unaligned LLMs in a plug-and-play fashion. Methodology, we employ delta debugging to identify the critical components of knowledge necessary for effective distillation. On the harmful question dataset, our method significantly enhances the average defense success rate by approximately 14.41%, reaching as high as 51.39%, in 17 unaligned pre-trained LLMs, without compromising performance.
Recent advances in large language models (LLMs) for code applications have demonstrated remarkable zero-shot fluency and instruction following on challenging code related tasks ranging from test case generation to self-repair. Unsurprisingly, however, models struggle to compose syntactically valid programs in programming languages unrepresented in pre-training, referred to as very low-resource Programming Languages (VLPLs). VLPLs appear in crucial settings including domain-specific languages for internal to tools and tool-chains and legacy languages. Inspired by an HCI technique called natural program elicitation, we propose designing an intermediate language that LLMs ``naturally'' know how to use and which can be automatically compiled to the target VLPL. Specifically, we introduce synthetic programming elicitation and compilation (SPEAK), an approach that enables LLMs to generate syntactically valid code even for VLPLs. We empirically evaluate the performance of SPEAK in a case study and find that, compared to existing retrieval and fine-tuning baselines, SPEAK produces syntactically correct programs more frequently without sacrificing semantic correctness.
High spectral dimensionality and the shortage of annotations make hyperspectral image (HSI) classification a challenging problem. Recent studies suggest that convolutional neural networks can learn discriminative spatial features, which play a paramount role in HSI interpretation. However, most of these methods ignore the distinctive spectral-spatial characteristic of hyperspectral data. In addition, a large amount of unlabeled data remains an unexploited gold mine for efficient data use. Therefore, we proposed an integration of generative adversarial networks (GANs) and probabilistic graphical models for HSI classification. Specifically, we used a spectral-spatial generator and a discriminator to identify land cover categories of hyperspectral cubes. Moreover, to take advantage of a large amount of unlabeled data, we adopted a conditional random field to refine the preliminary classification results generated by GANs. Experimental results obtained using two commonly studied datasets demonstrate that the proposed framework achieved encouraging classification accuracy using a small number of data for training.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191