The rapid adoption of Large Language Models (LLMs) in cloud environments has intensified the demand for high-performance AI training and inference, where Remote Direct Memory Access (RDMA) plays a critical role. However, existing RDMA virtualization solutions, such as Single-Root Input/Output Virtualization (SR-IOV), face significant limitations in scalability, performance, and stability. These issues include lengthy container initialization times, hardware resource constraints, and inefficient traffic steering. To address these challenges, we propose Stellar, a new generation RDMA network for cloud AI. Stellar introduces three key innovations: Para-Virtualized Direct Memory Access (PVDMA) for on-demand memory pinning, extended Memory Translation Table (eMTT) for optimized GPU Direct RDMA (GDR) performance, and RDMA Packet Spray for efficient multi-path utilization. Deployed in our large-scale AI clusters, Stellar spins up virtual devices in seconds, reduces container initialization time by 15 times, and improves LLM training speed by up to 14%. Our evaluations demonstrate that Stellar significantly outperforms existing solutions, offering a scalable, stable, and high-performance RDMA network for cloud AI.

Abstract Retinal microaneurysm (MA) is a definite earliest clinical sigh of diabetic retinopathy (DR). Its automatic segmentation is key to realizing intelligent screening for early DR, which can significantly reduce the risk of visual impairment in patients. However, the minute scale and subtle contrast of MAs against the background pose challenges for segmentation. This paper focuses on automatic MA segmentation in fundus images. A novel pyramid feature fusion network (PFFNet) that progressively develops and fuses rich contextual information by integrating two pyramid modules is proposed. Multiple global pyramid scene parsing (GPSP) modules are introduced between the encoder and decoder to provide diverse global contextual information for the decoder through reconstructing skip connections. Additionally, a spatial scale‐aware pyramid (SSAP) module is introduced to dynamically fuse multi‐scale contextual information. This rich contextual information will help to identify MAs from low‐contrast background. Furthermore, to mitigate issue related to category imbalance, a combo loss function is introduced. Finally, to validate the effectiveness of the proposed method, experiments are conducted on two publicly available datasets, IDRiD and DDR, and PFFNet is compared with several state‐of‐the‐art models. The experimental results demonstrate the superiority of our PFFNet in the MA segmentation task.

Recently, automatic diagnostic approaches have been widely used to classify ocular diseases. Most of these approaches are based on a single imaging modality (e.g., fundus photography or optical coherence tomography (OCT)), which usually only reflect the oculopathy to a certain extent, and neglect the modality-specific information among different imaging modalities. This paper proposes a novel modality-specific attention network (MSAN) for multi-modal retinal image classification, which can effectively utilize the modality-specific diagnostic features from fundus and OCT images. The MSAN comprises two attention modules to extract the modality-specific features from fundus and OCT images, respectively. Specifically, for the fundus image, ophthalmologists need to observe local and global pathologies at multiple scales (e.g., from microaneurysms at the micrometer level, optic disc at millimeter level to blood vessels through the whole eye). Therefore, we propose a multi-scale attention module to extract both the local and global features from fundus images. Moreover, large background regions exist in the OCT image, which is meaningless for diagnosis. Thus, a region-guided attention module is proposed to encode the retinal layer-related features and ignore the background in OCT images. Finally, we fuse the modality-specific features to form a multi-modal feature and train the multi-modal retinal image classification network. The fusion of modality-specific features allows the model to combine the advantages of fundus and OCT modality for a more accurate diagnosis. Experimental results on a clinically acquired multi-modal retinal image (fundus and OCT) dataset demonstrate that our MSAN outperforms other well-known single-modal and multi-modal retinal image classification methods.