Add to Chrome
Add to Firefox
Add to EdgePingcheng Dong
Genetic Quantization-Aware Approximation for Non-Linear Operations in Transformers
Mar 29, 2024
Non-linear functions are prevalent in Transformers and their lightweight variants, incurring substantial and frequently underestimated hardware costs. Previous state-of-the-art works optimize these operations by piece-wise linear approximation and store the parameters in look-up tables (LUT), but most of them require unfriendly high-precision arithmetics such as FP/INT 32 and lack consideration of integer-only INT quantization. This paper proposed a genetic LUT-Approximation algorithm namely GQA-LUT that can automatically determine the parameters with quantization awareness. The results demonstrate that GQA-LUT achieves negligible degradation on the challenging semantic segmentation task for both vanilla and linear Transformer models. Besides, proposed GQA-LUT enables the employment of INT8-based LUT-Approximation that achieves an area savings of 81.3~81.7% and a power reduction of 79.3~80.2% compared to the high-precision FP/INT 32 alternatives. Code is available at https:// github.com/PingchengDong/GQA-LUT.
Boundary and Relation Distillation for Semantic Segmentation
Jan 24, 2024Recently, it has been revealed that small semantic segmentation (SS) models exhibit a tendency to make errors in maintaining boundary region completeness and preserving target region connectivity, despite their effective segmentation of the main object regions. To address these errors, we propose a targeted boundary and relation distillation (BRD) strategy using knowledge distillation from large teacher models to small student models. Specifically, the boundary distillation extracts explicit object boundaries from the hierarchical feature maps of the backbone network, subsequently enhancing the student model's mask quality in boundary regions. Concurrently, the relation distillation transfers implicit relations from the teacher model to the student model using pixel-level self-relation as a bridge, ensuring that the student's mask has strong target region connectivity. The proposed BRD is designed concretely for SS and is characterized by simplicity and efficiency. Through experimental evaluations on multiple SS datasets, including Pascal VOC 2012, Cityscapes, ADE20K, and COCO-Stuff 10K, we demonstrated that BRD significantly surpasses the current methods without increasing the inference costs, generating crisp region boundaries and smooth connecting regions that are challenging for small models.
LLM-FP4: 4-Bit Floating-Point Quantized Transformers
Oct 25, 2023We propose LLM-FP4 for quantizing both weights and activations in large language models (LLMs) down to 4-bit floating-point values, in a post-training manner. Existing post-training quantization (PTQ) solutions are primarily integer-based and struggle with bit widths below 8 bits. Compared to integer quantization, floating-point (FP) quantization is more flexible and can better handle long-tail or bell-shaped distributions, and it has emerged as a default choice in many hardware platforms. One characteristic of FP quantization is that its performance largely depends on the choice of exponent bits and clipping range. In this regard, we construct a strong FP-PTQ baseline by searching for the optimal quantization parameters. Furthermore, we observe a high inter-channel variance and low intra-channel variance pattern in activation distributions, which adds activation quantization difficulty. We recognize this pattern to be consistent across a spectrum of transformer models designed for diverse tasks, such as LLMs, BERT, and Vision Transformer models. To tackle this, we propose per-channel activation quantization and show that these additional scaling factors can be reparameterized as exponential biases of weights, incurring a negligible cost. Our method, for the first time, can quantize both weights and activations in the LLaMA-13B to only 4-bit and achieves an average score of 63.1 on the common sense zero-shot reasoning tasks, which is only 5.8 lower than the full-precision model, significantly outperforming the previous state-of-the-art by 12.7 points. Code is available at: https://github.com/nbasyl/LLM-FP4.