Jinrong Guo

Tongxuan Liu, Tao Peng, Peijun Yang et al.

We introduce xLLM, an intelligent and efficient Large Language Model (LLM) inference framework designed for high-performance, large-scale enterprise-grade serving, with deep optimizations for diverse AI accelerators. To address these challenges, xLLM builds a novel decoupled service-engine architecture. At the service layer, xLLM-Service features an intelligent scheduling module that efficiently processes multimodal requests and co-locates online and offline tasks through unified elastic scheduling to maximize cluster utilization. This module also relies on a workload-adaptive dynamic Prefill-Decode (PD) disaggregation policy and a novel Encode-Prefill-Decode (EPD) disaggregation policy designed for multimodal inputs. Furthermore, it incorporates a distributed architecture to provide global KV Cache management and robust fault-tolerant capabilities for high availability. At the engine layer, xLLM-Engine co-optimizes system and algorithm designs to fully saturate computing resources. This is achieved through comprehensive multi-layer execution pipeline optimizations, an adaptive graph mode and an xTensor memory management. xLLM-Engine also further integrates algorithmic enhancements such as optimized speculative decoding and dynamic EPLB, collectively serving to substantially boost throughput and inference efficiency. Extensive evaluations demonstrate that xLLM delivers significantly superior performance and resource efficiency. Under identical TPOT constraints, xLLM achieves throughput up to 1.7x that of MindIE and 2.2x that of vLLM-Ascend with Qwen-series models, while maintaining an average throughput of 1.7x that of MindIE with Deepseek-series models. xLLM framework is publicly available at https://github.com/jd-opensource/xllm and https://github.com/jd-opensource/xllm-service.

Chunzhao Xie, Tongxuan Liu, Lei Jiang et al.

Large Vision-Language Models have demonstrated remarkable performance across various tasks; however, the challenge of hallucinations constrains their practical applications. The hallucination problem arises from multiple factors, including the inherent hallucinations in language models, the limitations of visual encoders in perception, and biases introduced by multimodal data. Extensive research has explored ways to mitigate hallucinations. For instance, OPERA prevents the model from overly focusing on "anchor tokens", thereby reducing hallucinations, whereas VCD mitigates hallucinations by employing a contrastive decoding approach. In this paper, we investigate the correlation between the decay of attention to image tokens and the occurrence of hallucinations. Based on this finding, we propose Temporal Attention Real-time Accumulative Connection (TARAC), a novel training-free method that dynamically accumulates and updates LVLMs' attention on image tokens during generation. By enhancing the model's attention to image tokens, TARAC mitigates hallucinations caused by the decay of attention on image tokens. We validate the effectiveness of TARAC across multiple models and datasets, demonstrating that our approach substantially mitigates hallucinations. In particular, TARAC reduces $C_S$ by 25.2 and $C_I$ by 8.7 compared to VCD on the CHAIR benchmark.

Jinrong Guo, Wantao Liu, Wang Wang et al.

Typically, Ultra-deep neural network(UDNN) tends to yield high-quality model, but its training process is usually resource intensive and time-consuming. Modern GPU's scarce DRAM capacity is the primary bottleneck that hinders the trainability and the training efficiency of UDNN. In this paper, we present "AccUDNN", an accelerator that aims to make the utmost use of finite GPU memory resources to speed up the training process of UDNN. AccUDNN mainly includes two modules: memory optimizer and hyperparameter tuner. Memory optimizer develops a performance-model guided dynamic swap out/in strategy, by offloading appropriate data to host memory, GPU memory footprint can be significantly slashed to overcome the restriction of trainability of UDNN. After applying the memory optimization strategy, hyperparameter tuner is designed to explore the efficiency-optimal minibatch size and the matched learning rate. Evaluations demonstrate that AccUDNN cuts down the GPU memory requirement of ResNet-152 from more than 24GB to 8GB. In turn, given 12GB GPU memory budget, the efficiency-optimal minibatch size can reach 4.2x larger than original Caffe. Benefiting from better utilization of single GPU's computing resources and fewer parameter synchronization of large minibatch size, 7.7x speed-up is achieved by 8 GPUs' cluster without any communication optimization and no accuracy losses.