Oh! We Freeze: Improving Quantized Knowledge Distillation via Signal Propagation Analysis for Large Language Models

Large generative models such as large language models (LLMs) and diffusion models have revolutionized the fields of NLP and computer vision respectively. However, their slow inference, high computation and memory requirement makes it challenging to deploy them on edge devices. In this study, we propose a light-weight quantization aware fine tuning technique using knowledge distillation (KD-QAT) to improve the performance of 4-bit weight quantized LLMs using commonly available datasets to realize a popular language use case, on device chat applications. To improve this paradigm of finetuning, as main contributions, we provide insights into stability of KD-QAT by empirically studying the gradient propagation during training to better understand the vulnerabilities of KD-QAT based approaches to low-bit quantization errors. Based on our insights, we propose ov-freeze, a simple technique to stabilize the KD-QAT process. Finally, we experiment with the popular 7B LLaMAv2-Chat model at 4-bit quantization level and demonstrate that ov-freeze results in near floating point precision performance, i.e., less than 0.7% loss of accuracy on Commonsense Reasoning benchmarks.

DONNAv2 -- Lightweight Neural Architecture Search for Vision tasks

With the growing demand for vision applications and deployment across edge devices, the development of hardware-friendly architectures that maintain performance during device deployment becomes crucial. Neural architecture search (NAS) techniques explore various approaches to discover efficient architectures for diverse learning tasks in a computationally efficient manner. In this paper, we present the next-generation neural architecture design for computationally efficient neural architecture distillation - DONNAv2 . Conventional NAS algorithms rely on a computationally extensive stage where an accuracy predictor is learned to estimate model performance within search space. This building of accuracy predictors helps them predict the performance of models that are not being finetuned. Here, we have developed an elegant approach to eliminate building the accuracy predictor and extend DONNA to a computationally efficient setting. The loss metric of individual blocks forming the network serves as the surrogate performance measure for the sampled models in the NAS search stage. To validate the performance of DONNAv2 we have performed extensive experiments involving a range of diverse vision tasks including classification, object detection, image denoising, super-resolution, and panoptic perception network (YOLOP). The hardware-in-the-loop experiments were carried out using the Samsung Galaxy S10 mobile platform. Notably, DONNAv2 reduces the computational cost of DONNA by 10x for the larger datasets. Furthermore, to improve the quality of NAS search space, DONNAv2 leverages a block knowledge distillation filter to remove blocks with high inference costs.

ZiCo-BC: A Bias Corrected Zero-Shot NAS for Vision Tasks

Zero-Shot Neural Architecture Search (NAS) approaches propose novel training-free metrics called zero-shot proxies to substantially reduce the search time compared to the traditional training-based NAS. Despite the success on image classification, the effectiveness of zero-shot proxies is rarely evaluated on complex vision tasks such as semantic segmentation and object detection. Moreover, existing zero-shot proxies are shown to be biased towards certain model characteristics which restricts their broad applicability. In this paper, we empirically study the bias of state-of-the-art (SOTA) zero-shot proxy ZiCo across multiple vision tasks and observe that ZiCo is biased towards thinner and deeper networks, leading to sub-optimal architectures. To solve the problem, we propose a novel bias correction on ZiCo, called ZiCo-BC. Our extensive experiments across various vision tasks (image classification, object detection and semantic segmentation) show that our approach can successfully search for architectures with higher accuracy and significantly lower latency on Samsung Galaxy S10 devices.