This poster studies how to make distributed dense QR run efficiently on modern GPU clusters. Dense QR factorization is a core primitive for least-squares solvers, orthogonalization, eigensolvers, and many machine learning workloads. On distributed GPU systems, however, existing libraries still leave a large fraction of hardware peak unused because panel factorization and its associated communication remain on the critical path. This poster presents a macro–micro co-designed distributed QR framework. At the macro level, our approach replaces the conventional 2D block-cyclic layout with a 1D column block-cyclic layout and removes idle bubbles through a critical-path-guided pipeline. At the micro level, it optimizes the local TSQR/WY path, reshapes trailing updates with two-level blocking, and reduces workspace pressure with chained GEMM updates. It reaches 86.3% of FP32 peak on a single A100, 86.0% on a single H200, and 75.5% of peak on 8 × H200 with explicit Q. More importantly, it raises 8-GPU H200 strong-scaling efficiency to 73% versus 46% for cuSOLVERMp, directly targeting the intra-node regime most relevant in practice. It also maintains a stable 1.6 × speedup on 16 × H200.

Gait recognition has attracted increasing attention in both academia and industry as a non-intrusive human recognition technology from a distance without requiring cooperation. Triplet loss, which enforces relative distance constraints, is a fundamental component in gait recognition. Recently, several gait-specific triplet losses have been introduced to gait recognition. However, they only focus on sample selection and weighting to enhance constraints without exploring the gradient properties of Cosine/Euclidean metrics, which fundamentally influence the model training efficiency and feature discriminability. In this paper, we theoretically analyze triplet loss gradients combined with weight decay and identify inherent limitations due to inadequate norm-control: Cosine metric triplet loss (Lcos) exhibits excessive gradients resulting from small feature norms, while Euclidean metric triplet loss (Leuc) suffers from a small margin-to-norm ratio due to large feature norms. To address these issues, we propose two norm-control approaches to constrain the feature norm in a stable range: 1) Norm-Variance-Regularized Collaboration. 2) Norm-Based Regularization. Extensive experiments show that our methods outperform state-of-the-art results under both Cosine and Euclidean evaluation metrics on three in-the-wild datasets: Gait3D, GREW, and BUAA-Duke-Gait. The code will be available at https://github.com/bgdpgz/TL-Gait.

Diffusion models have achieved remarkable success in the image and video generation tasks. Nevertheless, they often require a large amount of memory and time overhead during inference, due to the complex network architecture and considerable number of timesteps for iterative diffusion. Recently, the post-training quantization (PTQ) technique has proved a promising way to reduce the inference cost by quantizing the float-point operations to low-bit ones. However, most of them fail to tackle with the large variations in the distribution of activations across distinct channels and timesteps, as well as the inconsistent of input between quantization and inference on diffusion models, thus leaving much room for improvement. To address the above issues, we propose a novel method dubbed Timestep-Channel Adaptive Quantization for Diffusion Models (TCAQ-DM). Specifically, we develop a timestep-channel joint reparameterization (TCR) module to balance the activation range along both the timesteps and channels, facilitating the successive reconstruction procedure. Subsequently, we employ a dynamically adaptive quantization (DAQ) module that mitigate the quantization error by selecting an optimal quantizer for each post-Softmax layers according to their specific types of distributions. Moreover, we present a progressively aligned reconstruction (PAR) strategy to mitigate the bias caused by the input mismatch. Extensive experiments on various benchmarks and distinct diffusion models demonstrate that the proposed method substantially outperforms the state-of-the-art approaches in most cases, especially yielding comparable FID metrics to the full precision model on CIFAR-10 in the W6A6 setting, while enabling generating available images in the W4A4 settings.