Gang Peng

Zhen Zheng, Xin Ji, Taosong Fang et al.

Large language models (LLMs) increasingly play an important role in a wide range of information processing and management tasks. Many of these tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix between requests. The KV context that are about to be reused may prematurely evicted with the implicit cache management. Besides, the streaming oriented systems do not leverage the request-batch information and can not mix the decoding tokens with the prefill chunks to the best for the batched scenarios, and thus fails to saturate the GPU. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks, and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Finally, BatchLLM optimizes the prefix-shared Attention kernel with horizontal fusion to reduce tail effect and kernel launch overhead. Extensive evaluation shows that BatchLLM outperforms vLLM and SGLang by 1.3$\times$ to 10.8$\times$ on a set of microbenchmarks and a typical industry workload under different hardware environments.

Gang Peng, Zhenyu Ren, Hao Wang et al.

To realize a robust robotic grasping system for unknown objects in an unstructured environment, large amounts of grasp data and 3D model data for the object are required, the sizes of which directly affect the rate of successful grasps. To reduce the time cost of data acquisition and labeling and increase the rate of successful grasps, we developed a self-supervised learning mechanism to control grasp tasks performed by manipulators. First, a manipulator automatically collects the point cloud for the objects from multiple perspectives to increase the efficiency of data acquisition. The complete point cloud for the objects is obtained by utilizing the hand-eye vision of the manipulator, and the TSDF algorithm. Then, the point cloud data for the objects is used to generate a series of six-degrees-of-freedom grasp poses, and the force-closure decision algorithm is used to add the grasp quality label to each grasp pose to realize the automatic labeling of grasp data. Finally, the point cloud in the gripper closing area corresponding to each grasp pose is obtained; it is then used to train the grasp-quality classification model for the manipulator. The results of data acquisition experiments demonstrate that the proposed method allows high-quality data to be obtained. The simulated results prove the effectiveness of the proposed grasp-data acquisition method. The results of performing actual grasping experiments demonstrate that the proposed self-supervised learning method can increase the rate of successful grasps for the manipulator.

Shiqi Yang, Gang Peng

In this paper, we introduce a simple but quite effective recognition framework dubbed D-PCN, aiming at enhancing feature extracting ability of CNN. The framework consists of two parallel CNNs, a discriminator and an extra classifier which takes integrated features from parallel networks and gives final prediction. The discriminator is core which drives parallel networks to focus on different regions and learn different representations. The corresponding training strategy is introduced to ensures utilization of discriminator. We validate D-PCN with several CNN models on benchmark datasets: CIFAR-100, and ImageNet, D-PCN enhances all models. In particular it yields state of the art performance on CIFAR-100 compared with related works. We also conduct visualization experiment on fine-grained Stanford Dogs dataset to verify our motivation. Additionally, we apply D-PCN for segmentation on PASCAL VOC 2012 and also find promotion.