Quantization has been widely used to compress and accelerate inference of large language models (LLMs). Existing methods focus on exploring the per-token dynamic calibration to ensure both inference acceleration and model accuracy under 4-bit quantization. However, in autoregressive generation inference of long sequences, the overhead of repeated dynamic quantization and dequantization steps becomes considerably expensive. In this work, we propose MergeQuant, an accurate and efficient per-channel static quantization framework. MergeQuant integrates the per-channel quantization steps with the corresponding scalings and linear mappings through a Quantization Step Migration (QSM) method, thereby eliminating the quantization overheads before and after matrix multiplication. Furthermore, in view of the significant differences between the different channel ranges, we propose dimensional reconstruction and adaptive clipping to address the non-uniformity of quantization scale factors and redistribute the channel variations to the subsequent modules to balance the parameter distribution under QSM. Within the static quantization setting of W4A4, MergeQuant reduces the accuracy gap on zero-shot tasks compared to FP16 baseline to 1.3 points on Llama-2-70B model. On Llama-2-7B model, MergeQuant achieves up to 1.77x speedup in decoding, and up to 2.06x speedup in end-to-end compared to FP16 baseline.

Video action recognition has made significant strides, but challenges remain in effectively using both spatial and temporal information. While existing methods often focus on either spatial features (e.g., object appearance) or temporal dynamics (e.g., motion), they rarely address the need for a comprehensive integration of both. Capturing the rich temporal evolution of video frames, while preserving their spatial details, is crucial for improving accuracy. In this paper, we introduce the Temporal Integration and Motion Enhancement (TIME) layer, a novel preprocessing technique designed to incorporate temporal information. The TIME layer generates new video frames by rearranging the original sequence, preserving temporal order while embedding $N^2$ temporally evolving frames into a single spatial grid of size $N \times N$. This transformation creates new frames that balance both spatial and temporal information, making them compatible with existing video models. When $N=1$, the layer captures rich spatial details, similar to existing methods. As $N$ increases ($N\geq2$), temporal information becomes more prominent, while the spatial information decreases to ensure compatibility with model inputs. We demonstrate the effectiveness of the TIME layer by integrating it into popular action recognition models, such as ResNet-50, Vision Transformer, and Video Masked Autoencoders, for both RGB and depth video data. Our experiments show that the TIME layer enhances recognition accuracy, offering valuable insights for video processing tasks.

We investigate parameter-efficient fine-tuning (PEFT) methods that can provide good accuracy under limited computational and memory budgets in the context of large language models (LLMs). We present a new PEFT method called Robust Adaptation (RoSA) inspired by robust principal component analysis that jointly trains $\textit{low-rank}$ and $\textit{highly-sparse}$ components on top of a set of fixed pretrained weights to efficiently approximate the performance of a full-fine-tuning (FFT) solution. Across a series of challenging generative tasks such as grade-school math and SQL query generation, which require fine-tuning for good performance, we show that RoSA outperforms LoRA, pure sparse fine-tuning, and alternative hybrid methods at the same parameter budget, and can even recover the performance of FFT on some tasks. We provide system support for RoSA to complement the training algorithm, specifically in the form of sparse GPU kernels which enable memory- and computationally-efficient training, and show that it is also compatible with low-precision base weights, resulting in the first joint representation combining quantization, low-rank and sparse approximations. Our code is accessible at https://github.com/IST-DASLab/RoSA.