Jinxin Xu

Jiarui Wang, Huichao Chai, Yuanhang Zhang et al.

Real-time recommender systems execute multi-stage cascades (retrieval, pre-processing, fine-grained ranking) under strict tail-latency SLOs, leaving only tens of milliseconds for ranking. Generative recommendation (GR) models can improve quality by consuming long user-behavior sequences, but in production their online sequence length is tightly capped by the ranking-stage P99 budget. We observe that the majority of GR tokens encode user behaviors that are independent of the item candidates, suggesting an opportunity to pre-infer a user-behavior prefix once and reuse it during ranking rather than recomputing it on the critical path. Realizing this idea at industrial scale is non-trivial: the prefix cache must survive across multiple pipeline stages before the final ranking instance is determined, the user population implies cache footprints far beyond a single device, and indiscriminate pre-inference would overload shared resources under high QPS. We present RelayGR, a production system that enables in-HBM relay-race inference for GR. RelayGR selectively pre-infers long-term user prefixes, keeps their KV caches resident in HBM over the request lifecycle, and ensures the subsequent ranking can consume them without remote fetches. RelayGR combines three techniques: 1) a sequence-aware trigger that admits only at-risk requests under a bounded cache footprint and pre-inference load, 2) an affinity-aware router that co-locates cache production and consumption by routing both the auxiliary pre-infer signal and the ranking request to the same instance, and 3) a memory-aware expander that uses server-local DRAM to capture short-term cross-request reuse while avoiding redundant reloads. We implement RelayGR on Huawei Ascend NPUs and evaluate it with real queries. Under a fixed P99 SLO, RelayGR supports up to 1.5$\times$ longer sequences and improves SLO-compliant throughput by up to 3.6$\times$.

Jinxin Xu, Haixin Wu, Yu Cheng et al.

This paper addresses the optimization of scheduling for workers at a logistics depot using a combination of genetic algorithm and simulated annealing algorithm. The efficient scheduling of permanent and temporary workers is crucial for optimizing the efficiency of the logistics depot while minimizing labor usage. The study begins by establishing a 0-1 integer linear programming model, with decision variables determining the scheduling of permanent and temporary workers for each time slot on a given day. The objective function aims to minimize person-days, while constraints ensure fulfillment of hourly labor requirements, limit workers to one time slot per day, cap consecutive working days for permanent workers, and maintain non-negativity and integer constraints. The model is then solved using genetic algorithms and simulated annealing. Results indicate that, for this problem, genetic algorithms outperform simulated annealing in terms of solution quality. The optimal solution reveals a minimum of 29857 person-days.

Cheng Xuan, Zhiqiang Tang, Jinxin Xu

One of the most efficient ways for a learning-based robotic arm to learn to process complex tasks as human, is to directly learn from observing how human complete those tasks, and then imitate. Our idea is based on success of Deep Q-Learning (DQN) algorithm according to reinforcement learning, and then extend to Deep Deterministic Policy Gradient (DDPG) algorithm. We developed a learning-based method, combining modified DDPG and visual imitation network. Our approach acquires frames only from a monocular camera, and no need to either construct a 3D environment or generate actual points. The result we expected during training, was that robot would be able to move as almost the same as how human hands did.