A. Habibian
Publications
HLA: Hadamard Linear Attention
The attention mechanism is an important reason for the success of transformers. It relies on computing pairwise relations between tokens. To reduce the high computational cost of standard quadratic attention, linear attention has been proposed as an efficient approximation. It employs kernel functions that are applied independently to the inputs before the pairwise similarities are calculated. That allows for an efficient computational procedure which, however, amounts to a low-degree rational function approximating softmax. We propose Hadamard Linear Attention (HLA). Unlike previous works on linear attention, the nonlinearity in HLA is not applied separately to queries and keys, but, analogously to standard softmax attention, after the pairwise similarities have been computed. It will be shown that the proposed nonlinearity amounts to a higher-degree rational function to approximate softmax. An efficient computational scheme for the proposed method is derived that is similar to that of standard linear attention. In contrast to other approaches, no time-consuming tensor reshaping is necessary to apply the proposed algorithm. The effectiveness of the approach is demonstrated by applying it to a large diffusion transformer model for video generation, an application that involves very large amounts of tokens.
Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs
Vision-Language Models (VLMs) achieve strong multimodal performance but are costly to deploy, and post-training quantization often causes significant accuracy loss. Despite its potential, quantization-aware training for VLMs remains underexplored. We propose GRACE, a framework unifying knowledge distillation and QAT under the Information Bottleneck principle: quantization constrains information capacity while distillation guides what to preserve within this budget. Treating the teacher as a proxy for task-relevant information, we introduce confidence-gated decoupled distillation to filter unreliable supervision, relational centered kernel alignment to transfer visual token structures, and an adaptive controller via Lagrangian relaxation to balance fidelity against capacity constraints. Across extensive benchmarks on LLaVA and Qwen families, our INT4 models consistently outperform FP16 baselines (e.g., LLaVA-1.5-7B: 70.1 vs. 66.8 on SQA; Qwen2-VL-2B: 76.9 vs. 72.6 on MMBench), nearly matching teacher performance. Using real INT4 kernel, we achieve 3$\times$ throughput with 54% memory reduction. This principled framework significantly outperforms existing quantization methods, making GRACE a compelling solution for resource-constrained deployment.