Xinpeng Zhao

Xun Sun, Shaoyuan Chen, Pingchuan Ma et al.

Mixture-of-Experts (MoE) serving relies on wide expert parallelism (EP) to aggregate the memory capacity and bandwidth of many GPUs within one inference instance. This efficiency comes with a systems cost: every decoding step depends on token dispatch and combination across all active EP ranks, so even one rank failure can disrupt the entire service. Existing EP stacks handle such failures poorly because they treat membership as a fixed configuration established at initialization. The same rank set determines communicator state, expert placement, and the routing metadata baked into CUDA execution graphs, leaving the system with no way to shrink around a failure while keeping the instance valid. This paper argues that partial-failure tolerance should instead be formulated as a live EP validity problem. We present EEP, a communication and runtime substrate that represents membership as explicit, mutable runtime state. EEP repairs the specific state invalidated by a fault: it restores peer reachability without rebuilding the communication substrate, repairs lost expert coverage through a bandwidth-aware hierarchy, and reintegrates repaired ranks without forcing healthy ranks to recapture their CUDA graphs. We implement EEP in an EP serving stack integrated with SGLang and evaluate it under steady-state serving, failure recovery, and rank reintegration. The results show that explicit mutable membership preserves the steady-state fast path, staying within 4.4% of a fixed-membership DeepEP baseline under static serving, while turning a local rank fault from whole-instance downtime into two bounded interruptions. On a single-rank failure workload, EEP incurs an 11s recovery pause and an 8s reintegration pause, and restores throughput to within 95% of the pre-fault level within 52s, whereas a fixed-membership full-restart baseline remains unavailable until 348s.

Ruiyang Ma, Teng Ma, Zhiyuan Su et al.

Engram conditional memory has emerged as a promising component for LLMs by decoupling static knowledge lookup from dynamic computation. Since Engram exhibits sparse access patterns and supports prefetching, its massive embedding tables are well-suited for offloading to lower-tier memory. In this paper, we propose using Compute Express Link (CXL) memory pool for Engram storage. Compared to RDMA, CXL provides fine-grained and low-latency access required by minimal and discrete retrieval patterns of Engram. We integrate the CXL-based Engram pool into SGLang, achieving near-DRAM end-to-end performance. This provides a scalable and cost-efficient storage solution for future Engram-integrated LLMs without compromising inference performance.

Bowen Huang, Yanwei Zheng, Chuanlin Lan et al.

Vision-and-Language Navigation (VLN) is a challenging task where an agent is required to navigate to a natural language described location via vision observations. The navigation abilities of the agent can be enhanced by the relations between objects, which are usually learned using internal objects or external datasets. The relationships between internal objects are modeled employing graph convolutional network (GCN) in traditional studies. However, GCN tends to be shallow, limiting its modeling ability. To address this issue, we utilize a cross attention mechanism to learn the connections between objects over a trajectory, which takes temporal continuity into account, termed as Temporal Object Relations (TOR). The external datasets have a gap with the navigation environment, leading to inaccurate modeling of relations. To avoid this problem, we construct object connections based on observations from all viewpoints in the navigational environment, which ensures complete spatial coverage and eliminates the gap, called Spatial Object Relations (SOR). Additionally, we observe that agents may repeatedly visit the same location during navigation, significantly hindering their performance. For resolving this matter, we introduce the Turning Back Penalty (TBP) loss function, which penalizes the agent's repetitive visiting behavior, substantially reducing the navigational distance. Experimental results on the REVERIE, SOON, and R2R datasets demonstrate the effectiveness of the proposed method.

Xinpeng Zhao, Yanwei Zheng, Chuanlin Lan et al.

Weakly supervised text-based person retrieval seeks to retrieve images of a target person using textual descriptions, without relying on identity annotations and is more challenging and practical. The primary challenge is the intra-class differences, encompassing intra-modal feature variations and cross-modal semantic gaps. Prior works have focused on instance-level samples and ignored prototypical features of each person which are intrinsic and invariant. Toward this, we propose a Cross-Modal Prototypical Contrastive Learning (CPCL) method. In practice, the CPCL introduces the CLIP model to weakly supervised text-based person retrieval to map visual and textual instances into a shared latent space. Subsequently, the proposed Prototypical Multi-modal Memory (PMM) module captures associations between heterogeneous modalities of image-text pairs belonging to the same person through the Hybrid Cross-modal Matching (HCM) module in a many-to-many mapping fashion. Moreover, the Outlier Pseudo Label Mining (OPLM) module further distinguishes valuable outlier samples from each modality, enhancing the creation of more reliable clusters by mining implicit relationships between image-text pairs. We conduct extensive experiments on popular benchmarks of weakly supervised text-based person retrieval, which validate the effectiveness, generalizability of CPCL.