Retrieval Or Holistic Understanding? Dolce: Differentiate Our Long Context Evaluation Tasks

We argue that there are two major distinct capabilities in long context understanding: retrieval and holistic understanding. Understanding and further improving LLMs' long context capabilities would not be possible without knowing the tasks' focus categories. We aim to automatically identify retrieval focused and holistic understanding focused problems from suites of benchmarks and quantitatively measure the difficulty within each focus. In this paper, we present the Dolce framework, which parameterizes each problem by $\lambda$ (complexity) and $k$ (redundancy) and assigns to one of five predefined focus categories. We propose to sample short contexts from the full context and estimate the probability an LLM solves the problem using the sampled spans. To find the $\lambda$ and $k$ for each problem, we further propose a mixture model of a non-parametric background noise component and a parametric/non-parametric hybrid oracle component, where we derive the probability functions parameterized by $\lambda$ and $k$ for both the correct-or-wrong (COW) scenario and the partial-point-in-grading (PIG) scenario. Our proposed methods can identify 0% to 67% of the problems are retrieval focused and 0% to 90% of the problems are holistic understanding focused across 44 existing long context evaluation tasks.

MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization

In this paper, we present a simple optimization-based preprocessing technique called Weight Magnitude Reduction (MagR) to improve the performance of post-training quantization. For each linear layer, we adjust the pre-trained floating-point weights by solving an $\ell_\infty$-regularized optimization problem. This process greatly diminishes the maximum magnitude of the weights and smooths out outliers, while preserving the layer's output. The preprocessed weights are centered more towards zero, which facilitates the subsequent quantization process. To implement MagR, we address the $\ell_\infty$-regularization by employing an efficient proximal gradient descent algorithm. Unlike existing preprocessing methods that involve linear transformations and subsequent post-processing steps, which can introduce significant overhead at inference time, MagR functions as a non-linear transformation, eliminating the need for any additional post-processing. This ensures that MagR introduces no overhead whatsoever during inference. Our experiments demonstrate that MagR achieves state-of-the-art performance on the Llama family of models. For example, we achieve a Wikitext2 perplexity of 5.95 on the LLaMA2-70B model for per-channel INT2 weight quantization without incurring any inference overhead.

CoMERA: Computing- and Memory-Efficient Training via Rank-Adaptive Tensor Optimization

Training large AI models such as deep learning recommendation systems and foundation language (or multi-modal) models costs massive GPUs and computing time. The high training cost has become only affordable to big tech companies, meanwhile also causing increasing concerns about the environmental impact. This paper presents CoMERA, a Computing- and Memory-Efficient training method via Rank-Adaptive tensor optimization. CoMERA achieves end-to-end rank-adaptive tensor-compressed training via a multi-objective optimization formulation, and improves the training to provide both a high compression ratio and excellent accuracy in the training process. Our optimized numerical computation (e.g., optimized tensorized embedding and tensor-vector contractions) and GPU implementation eliminate part of the run-time overhead in the tensorized training on GPU. This leads to, for the first time, $2-3\times$ speedup per training epoch compared with standard training. CoMERA also outperforms the recent GaLore in terms of both memory and computing efficiency. Specifically, CoMERA is $2\times$ faster per training epoch and $9\times$ more memory-efficient than GaLore on a tested six-encoder transformer with single-batch training. With further HPC optimization, CoMERA may significantly reduce the training cost of large language models.

Equipping Transformer with Random-Access Reading for Long-Context Understanding

Long-context modeling presents a significant challenge for transformer-based large language models (LLMs) due to the quadratic complexity of the self-attention mechanism and issues with length extrapolation caused by pretraining exclusively on short inputs. Existing methods address computational complexity through techniques such as text chunking, the kernel approach, and structured attention, and tackle length extrapolation problems through positional encoding, continued pretraining, and data engineering. These approaches typically require $\textbf{sequential access}$ to the document, necessitating reading from the first to the last token. We contend that for goal-oriented reading of long documents, such sequential access is not necessary, and a proficiently trained model can learn to omit hundreds of less pertinent tokens. Inspired by human reading behaviors and existing empirical observations, we propose $\textbf{random access}$, a novel reading strategy that enables transformers to efficiently process long documents without examining every token. Experimental results from pretraining, fine-tuning, and inference phases validate the efficacy of our method.

Parameter-Efficient Fine-Tuning With Adapters

In the arena of language model fine-tuning, the traditional approaches, such as Domain-Adaptive Pretraining (DAPT) and Task-Adaptive Pretraining (TAPT), although effective, but computational intensive. This research introduces a novel adaptation method utilizing the UniPELT framework as a base and added a PromptTuning Layer, which significantly reduces the number of trainable parameters while maintaining competitive performance across various benchmarks. Our method employs adapters, which enable efficient transfer of pretrained models to new tasks with minimal retraining of the base model parameters. We evaluate our approach using three diverse datasets: the GLUE benchmark, a domain-specific dataset comprising four distinct areas, and the Stanford Question Answering Dataset 1.1 (SQuAD). Our results demonstrate that our customized adapter-based method achieves performance comparable to full model fine-tuning, DAPT+TAPT and UniPELT strategies while requiring fewer or equivalent amount of parameters. This parameter efficiency not only alleviates the computational burden but also expedites the adaptation process. The study underlines the potential of adapters in achieving high performance with significantly reduced resource consumption, suggesting a promising direction for future research in parameter-efficient fine-tuning.

COMQ: A Backpropagation-Free Algorithm for Post-Training Quantization

Post-training quantization (PTQ) has emerged as a practical approach to compress large neural networks, making them highly efficient for deployment. However, effectively reducing these models to their low-bit counterparts without compromising the original accuracy remains a key challenge. In this paper, we propose an innovative PTQ algorithm termed COMQ, which sequentially conducts coordinate-wise minimization of the layer-wise reconstruction errors. We consider the widely used integer quantization, where every quantized weight can be decomposed into a shared floating-point scalar and an integer bit-code. Within a fixed layer, COMQ treats all the scaling factor(s) and bit-codes as the variables of the reconstruction error. Every iteration improves this error along a single coordinate while keeping all other variables constant. COMQ is easy to use and requires no hyper-parameter tuning. It instead involves only dot products and rounding operations. We update these variables in a carefully designed greedy order, significantly enhancing the accuracy. COMQ achieves remarkable results in quantizing 4-bit Vision Transformers, with a negligible loss of less than 1% in Top-1 accuracy. In 4-bit INT quantization of convolutional neural networks, COMQ maintains near-lossless accuracy with a minimal drop of merely 0.3% in Top-1 accuracy.

Attendre: Wait To Attend By Retrieval With Evicted Queries in Memory-Based Transformers for Long Context Processing

As LLMs have become capable of processing more complex types of inputs, researchers have recently studied how to efficiently and affordably process possibly arbitrarily long sequences. One effective approach is to use a FIFO memory to store keys and values of an attention sublayer from past chunks to allow subsequent queries to attend. However, this approach requires a large memory and/or takes into the consideration the specific LM architecture. Moreover, due to the causal nature between the key-values in prior context and the queries at present, this approach cannot be extended to bidirectional attention such as in an encoder-decoder or PrefixLM decoder-only architecture. In this paper, we propose to use eviction policies, such as LRA and LFA, to reduce the memory size and adapt to various architectures, and we also propose the Attendre layer, a wait-to-attend mechanism by retrieving the key-value memory (K/V memory) with evicted queries in the query memory (Q memory). As a first step, we evaluate this method in the context length extension setup using the TriviaQA reading comprehension task, and show the effectiveness of the approach.

Quantization-Aware and Tensor-Compressed Training of Transformers for Natural Language Understanding

Fine-tuned transformer models have shown superior performances in many natural language tasks. However, the large model size prohibits deploying high-performance transformer models on resource-constrained devices. This paper proposes a quantization-aware tensor-compressed training approach to reduce the model size, arithmetic operations, and ultimately runtime latency of transformer-based models. We compress the embedding and linear layers of transformers into small low-rank tensor cores, which significantly reduces model parameters. A quantization-aware training with learnable scale factors is used to further obtain low-precision representations of the tensor-compressed models. The developed approach can be used for both end-to-end training and distillation-based training. To improve the convergence, a layer-by-layer distillation is applied to distill a quantized and tensor-compressed student model from a pre-trained transformer. The performance is demonstrated in two natural language understanding tasks, showing up to $63\times$ compression ratio, little accuracy loss and remarkable inference and training speedup.

Leveraging Global Binary Masks for Structure Segmentation in Medical Images

Deep learning (DL) models for medical image segmentation are highly influenced by intensity variations of input images and lack generalization due to primarily utilizing pixels' intensity information for inference. Acquiring sufficient training data is another challenge limiting models' applications. We proposed to leverage the consistency of organs' anatomical shape and position information in medical images. We introduced a framework leveraging recurring anatomical patterns through global binary masks for organ segmentation. Two scenarios were studied.1) Global binary masks were the only model's (i.e. U-Net) input, forcing exclusively encoding organs' position and shape information for segmentation/localization.2) Global binary masks were incorporated as an additional channel functioning as position/shape clues to mitigate training data scarcity. Two datasets of the brain and heart CT images with their ground-truth were split into (26:10:10) and (12:3:5) for training, validation, and test respectively. Training exclusively on global binary masks led to Dice scores of 0.77(0.06) and 0.85(0.04), with the average Euclidian distance of 3.12(1.43)mm and 2.5(0.93)mm relative to the center of mass of the ground truth for the brain and heart structures respectively. The outcomes indicate that a surprising degree of position and shape information is encoded through global binary masks. Incorporating global binary masks led to significantly higher accuracy relative to the model trained on only CT images in small subsets of training data; the performance improved by 4.3-125.3% and 1.3-48.1% for 1-8 training cases of the brain and heart datasets respectively. The findings imply the advantages of utilizing global binary masks for building generalizable models and to compensate for training data scarcity.

Pruning Redundant Mappings in Transformer Models via Spectral-Normalized Identity Prior

Traditional (unstructured) pruning methods for a Transformer model focus on regularizing the individual weights by penalizing them toward zero. In this work, we explore spectral-normalized identity priors (SNIP), a structured pruning approach that penalizes an entire residual module in a Transformer model toward an identity mapping. Our method identifies and discards unimportant non-linear mappings in the residual connections by applying a thresholding operator on the function norm. It is applicable to any structured module, including a single attention head, an entire attention block, or a feed-forward subnetwork. Furthermore, we introduce spectral normalization to stabilize the distribution of the post-activation values of the Transformer layers, further improving the pruning effectiveness of the proposed methodology. We conduct experiments with BERT on 5 GLUE benchmark tasks to demonstrate that SNIP achieves effective pruning results while maintaining comparable performance. Specifically, we improve the performance over the state-of-the-art by 0.5 to 1.0% on average at 50% compression ratio.