A.1 Hyperparameters for QLORA We do a hyperparameter search for LoRA over the following variables: LoRA dropout { 0.0, 0.05, 0.1}, LoRA r { 8, 16, 32, 64, 128, 256}, LoRA layers {key+query, all attention layers, all FFN layers, all layers, attention + FFN output layers}. We keep LoRA α fixed and search the learning rate, since LoRA α is always proportional to the learning rate. We find that LoRA dropout 0.05 is useful for small models (7B, 13B), but not for larger models (33B, 65B). Each dot represents a combination of hyperparameters and for each LoRA r we run 3 random seed with each hyperparameter combination. The performance of specific LoRA r values appears to be independent of other hyperparameters.

We use TSN-ResNet [11] as our backbone (see Table 1) and initialize it with ImageNet-pretrained weights [4]. We replace its 7 spatial convolutional layers with the RSA layers; for every two ResNet blocks from the third block in res2 to the second block in res5, each spatial convolutional layer is replaced with the RSA layer. For the bottlenecks including RSA layers, we randomly initialize weights using MSRA initialization [3] and set the gamma parameter of the last batch normalization layer to zero. We resize the resolution of each frame to 240 320, and apply random cropping as 224 224, scale jittering, and random horizontal flipping for data augmentation. Note that we do not flip videos of which action labels include'left' or'right' words, e.g., 'pulling something from left to right'.

A critical approach for efficiently deploying computationally demanding large language models (LLMs) is Key-Value (KV) caching. The KV cache stores key-value states of previously generated tokens, significantly reducing the need for repetitive computations and thereby lowering latency in autoregressive generation. However, the size of the KV cache grows linearly with sequence length, posing challenges for applications requiring long context input and extensive sequence generation. In this paper, we present a simple yet effective approach, called MiniCache, to compress the KV cache across layers from a novel depth perspective, significantly reducing the memory footprint for LLM inference. Our approach is based on the observation that KV cache states exhibit high similarity between the adjacent layers in the middle-to-deep portion of LLMs.

Internet of Things and Deep Learning are synergetically and exponentially growing industrial fields with a massive call for their unification into a common framework called Edge AI. While on-device inference is a well-explored topic in recent research, backpropagation remains an open challenge due to its prohibitive computational and memory costs compared to the extreme resource constraints of embedded devices. Drawing on tensor decomposition research, we tackle the main bottleneck of backpropagation, namely the memory footprint of activation map storage. We investigate and compare the effects of activation compression using Singular Value Decomposition and its tensor variant, High-Order Singular Value Decomposition. The application of low-order decomposition results in considerable memory savings while preserving the features essential for learning, and also offers theoretical guarantees to convergence. Experimental results obtained on main-stream architectures and tasks demonstrate Pareto-superiority over other state-of-the-art solutions, in terms of the trade-off between generalization and memory footprint.

Finetuning large language models (LLMs) in federated learning (FL) settings has become increasingly important as it allows resource-constrained devices to finetune a model using private data. However, finetuning LLMs using backpropagation requires excessive memory (especially from intermediate activations) for resource-constrained devices. While Forward-mode Auto-Differentiation (AD) can significantly reduce memory footprint from activations, we observe that directly applying it to LLM finetuning results in slow convergence and poor accuracy. In this paper, we introduce Spry, an FL algorithm that splits trainable weights of an LLM among participating clients, such that each client computes gradients using forward-mode AD that are closer estimations of the true gradients. Spry achieves a low memory footprint, high accuracy, and fast convergence. We formally prove that the global gradients in Spry are unbiased estimators of true global gradients for homogeneous data distributions across clients, while heterogeneity increases bias of the estimates. We also derive Spry's convergence rate, showing that the gradients decrease inversely proportional to the number of FL rounds, indicating the convergence up to the limits of heterogeneity. Empirically, Spry reduces the memory footprint during training by 1.4-7.1$\times$ in contrast to backpropagation, while reaching comparable accuracy, across a wide range of language tasks, models, and FL settings. Spry reduces the convergence time by 1.2-20.3$\times$

We thank all the reviewers for their constructive comments. We will provide details in the final draft. MCUNet shows consistent improvement across different devices (F746, H743) and tasks (classification, detection). R1: Whether the overall network topology brings major improvement. R2: Why the auto-tuning in TVM fails to work on MCUs.

Neural Information Processing Systems http://nips.cc/