S2D: Sorted Speculative Decoding For More Efficient Deployment of Nested Large Language Models

Deployment of autoregressive large language models (LLMs) is costly, and as these models increase in size, the associated costs will become even more considerable. Consequently, different methods have been proposed to accelerate the token generation process and reduce costs. Speculative decoding (SD) is among the most promising approaches to speed up the LLM decoding process by verifying multiple tokens in parallel and using an auxiliary smaller draft model to generate the possible tokens. In SD, usually, one draft model is used to serve a specific target model; however, in practice, LLMs are diverse, and we might need to deal with many target models or more than one target model simultaneously. In this scenario, it is not clear which draft model should be used for which target model, and searching among different draft models or training customized draft models can further increase deployment costs. In this paper, we first introduce a novel multi-target scenario for the deployment of draft models for faster inference. Then, we present a novel, more efficient sorted speculative decoding mechanism that outperforms regular baselines in multi-target settings. We evaluated our method on Spec-Bench in different settings, including base models such as Vicuna 7B, 13B, and LLama Chat 70B. Our results suggest that our draft models perform better than baselines for multiple target models at the same time.

Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks

Quantitative systems pharmacology (QSP) is widely used to assess drug effects and toxicity before the drug goes to clinical trial. However, significant manual distillation of the literature is needed in order to construct a QSP model. Parameters may need to be fit, and simplifying assumptions of the model need to be made. In this work, we apply Universal Physics-Informed Neural Networks (UPINNs) to learn unknown components of various differential equations that model chemotherapy pharmacodynamics. We learn three commonly employed chemotherapeutic drug actions (log-kill, Norton-Simon, and E_max) from synthetic data. Then, we use the UPINN method to fit the parameters for several synthetic datasets simultaneously. Finally, we learn the net proliferation rate in a model of doxorubicin (a chemotherapeutic) pharmacodynamics. As these are only toy examples, we highlight the usefulness of UPINNs in learning unknown terms in pharmacodynamic and pharmacokinetic models.

Orchid: Flexible and Data-Dependent Convolution for Sequence Modeling

In the rapidly evolving landscape of deep learning, the quest for models that balance expressivity with computational efficiency has never been more critical. This paper introduces Orchid, a novel architecture that reimagines sequence modeling by incorporating a new data-dependent convolution mechanism. Orchid is designed to address the inherent limitations of traditional attention mechanisms, particularly their quadratic complexity, without compromising the ability to capture long-range dependencies and in-context learning. At the core of Orchid lies the data-dependent convolution layer, which dynamically adjusts its kernel conditioned on input data using a dedicated conditioning neural network. We design two simple conditioning networks that maintain shift equivariance in the adaptive convolution operation. The dynamic nature of data-dependent convolution kernel, coupled with gating operations, grants Orchid high expressivity while maintaining efficiency and quasilinear scalability for long sequences. We rigorously evaluate Orchid across multiple domains, including language modeling and image classification, to showcase its performance and generality. Our experiments demonstrate that Orchid architecture not only outperforms traditional attention-based architectures such as BERT and Vision Transformers with smaller model sizes, but also extends the feasible sequence length beyond the limitations of the dense attention layers. This achievement represents a significant step towards more efficient and scalable deep learning models for sequence modeling.

QDyLoRA: Quantized Dynamic Low-Rank Adaptation for Efficient Large Language Model Tuning

Finetuning large language models requires huge GPU memory, restricting the choice to acquire Larger models. While the quantized version of the Low-Rank Adaptation technique, named QLoRA, significantly alleviates this issue, finding the efficient LoRA rank is still challenging. Moreover, QLoRA is trained on a pre-defined rank and, therefore, cannot be reconfigured for its lower ranks without requiring further fine-tuning steps. This paper proposes QDyLoRA -Quantized Dynamic Low-Rank Adaptation-, as an efficient quantization approach for dynamic low-rank adaptation. Motivated by Dynamic LoRA, QDyLoRA is able to efficiently finetune LLMs on a set of pre-defined LoRA ranks. QDyLoRA enables fine-tuning Falcon-40b for ranks 1 to 64 on a single 32 GB V100-GPU through one round of fine-tuning. Experimental results show that QDyLoRA is competitive to QLoRA and outperforms when employing its optimal rank.

Scalable Graph Self-Supervised Learning

In regularization Self-Supervised Learning (SSL) methods for graphs, computational complexity increases with the number of nodes in graphs and embedding dimensions. To mitigate the scalability of non-contrastive graph SSL, we propose a novel approach to reduce the cost of computing the covariance matrix for the pre-training loss function with volume-maximization terms. Our work focuses on reducing the cost associated with the loss computation via graph node or dimension sampling. We provide theoretical insight into why dimension sampling would result in accurate loss computations and support it with mathematical derivation of the novel approach. We develop our experimental setup on the node-level graph prediction tasks, where SSL pre-training has shown to be difficult due to the large size of real world graphs. Our experiments demonstrate that the cost associated with the loss computation can be reduced via node or dimension sampling without lowering the downstream performance. Our results demonstrate that sampling mostly results in improved downstream performance. Ablation studies and experimental analysis are provided to untangle the role of the different factors in the experimental setup.

WERank: Towards Rank Degradation Prevention for Self-Supervised Learning Using Weight Regularization

A common phenomena confining the representation quality in Self-Supervised Learning (SSL) is dimensional collapse (also known as rank degeneration), where the learned representations are mapped to a low dimensional subspace of the representation space. The State-of-the-Art SSL methods have shown to suffer from dimensional collapse and fall behind maintaining full rank. Recent approaches to prevent this problem have proposed using contrastive losses, regularization techniques, or architectural tricks. We propose WERank, a new regularizer on the weight parameters of the network to prevent rank degeneration at different layers of the network. We provide empirical evidence and mathematical justification to demonstrate the effectiveness of the proposed regularization method in preventing dimensional collapse. We verify the impact of WERank on graph SSL where dimensional collapse is more pronounced due to the lack of proper data augmentation. We empirically demonstrate that WERank is effective in helping BYOL to achieve higher rank during SSL pre-training and consequently downstream accuracy during evaluation probing. Ablation studies and experimental analysis shed lights on the underlying factors behind the performance gains of the proposed approach.

Sorted LLaMA: Unlocking the Potential of Intermediate Layers of Large Language Models for Dynamic Inference Using Sorted Fine-Tuning (SoFT)

The rapid advancement of large language models (LLMs) has revolutionized natural language processing (NLP). While these models excel at understanding and generating human-like text, their widespread deployment can be prohibitively expensive. SortedNet is a recent training technique for enabling dynamic inference for deep neural networks. It leverages network modularity to create sub-models with varying computational loads, sorting them based on computation/accuracy characteristics in a nested manner. We extend SortedNet to generative NLP tasks, making large language models dynamic without any pretraining and by only replacing standard Supervised Fine-Tuning (SFT) with Sorted Fine-Tuning (SoFT) at the same costs. Our approach boosts model efficiency, eliminating the need for multiple models for various scenarios during inference. We show that using this approach, we are able to unlock the potential of intermediate layers of transformers in generating the target output. Our sub-models remain integral components of the original model, minimizing storage requirements and transition costs between different computational/latency budgets. By applying this approach on LLaMa 2 13B for tuning on the Stanford Alpaca dataset and comparing it to normal tuning and early exit via PandaLM benchmark, we show that Sorted Fine-Tuning can deliver models twice as fast as the original model while maintaining or exceeding performance.

SortedNet, a Place for Every Network and Every Network in its Place: Towards a Generalized Solution for Training Many-in-One Neural Networks

As the size of deep learning models continues to grow, finding optimal models under memory and computation constraints becomes increasingly more important. Although usually the architecture and constituent building blocks of neural networks allow them to be used in a modular way, their training process is not aware of this modularity. Consequently, conventional neural network training lacks the flexibility to adapt the computational load of the model during inference. This paper proposes SortedNet, a generalized and scalable solution to harness the inherent modularity of deep neural networks across various dimensions for efficient dynamic inference. Our training considers a nested architecture for the sub-models with shared parameters and trains them together with the main model in a sorted and probabilistic manner. This sorted training of sub-networks enables us to scale the number of sub-networks to hundreds using a single round of training. We utilize a novel updating scheme during training that combines random sampling of sub-networks with gradient accumulation to improve training efficiency. Furthermore, the sorted nature of our training leads to a search-free sub-network selection at inference time; and the nested architecture of the resulting sub-networks leads to minimal storage requirement and efficient switching between sub-networks at inference. Our general dynamic training approach is demonstrated across various architectures and tasks, including large language models and pre-trained vision models. Experimental results show the efficacy of the proposed approach in achieving efficient sub-networks while outperforming state-of-the-art dynamic training approaches. Our findings demonstrate the feasibility of training up to 160 different sub-models simultaneously, showcasing the extensive scalability of our proposed method while maintaining 96% of the model performance.

Recurrent Neural Networks and Long Short-Term Memory Networks: Tutorial and Survey

This is a tutorial paper on Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), and their variants. We start with a dynamical system and backpropagation through time for RNN. Then, we discuss the problems of gradient vanishing and explosion in long-term dependencies. We explain close-to-identity weight matrix, long delays, leaky units, and echo state networks for solving this problem. Then, we introduce LSTM gates and cells, history and variants of LSTM, and Gated Recurrent Units (GRU). Finally, we introduce bidirectional RNN, bidirectional LSTM, and the Embeddings from Language Model (ELMo) network, for processing a sequence in both directions.

Improved knowledge distillation by utilizing backward pass knowledge in neural networks

Knowledge distillation (KD) is one of the prominent techniques for model compression. In this method, the knowledge of a large network (teacher) is distilled into a model (student) with usually significantly fewer parameters. KD tries to better-match the output of the student model to that of the teacher model based on the knowledge extracts from the forward pass of the teacher network. Although conventional KD is effective for matching the two networks over the given data points, there is no guarantee that these models would match in other areas for which we do not have enough training samples. In this work, we address that problem by generating new auxiliary training samples based on extracting knowledge from the backward pass of the teacher in the areas where the student diverges greatly from the teacher. We compute the difference between the teacher and the student and generate new data samples that maximize the divergence. This is done by perturbing data samples in the direction of the gradient of the difference between the student and the teacher. Augmenting the training set by adding this auxiliary improves the performance of KD significantly and leads to a closer match between the student and the teacher. Using this approach, when data samples come from a discrete domain, such as applications of natural language processing (NLP) and language understanding, is not trivial. However, we show how this technique can be used successfully in such applications. We evaluated the performance of our method on various tasks in computer vision and NLP domains and got promising results.