The massive parameter scale of sparsely-activated Mixture-of-Experts (MoE) models necessitates distributed training with hybrid parallelism. Placing such training tasks,i.e.mapping the logical partitions of an MoE model to available physical NPUs, is challenging. Due to the bandwidth and latency discrepancies between intra- and inter- Pods, the cross-Pod communication usually becomes a bottleneck. The high dispersion of NPUs in multi-tenant clusters exacerbates this issue further. However, a paucity of studies has considered the cross-Pod model placement problem. To address this challenge, we propose a novel model placement scheme tailored for MoE model training with hybrid parallelism in multi-tenant clusters. By quantifying the cross-Pod communication overhead incurred during MoE model training, the model placement is formulated as a 0-1 integer quadratic problem, which is NP hard. Motivated by the traffic difference between different parallelism, we decompose this problem into two subproblems. To solve the subproblems, we propose a lightweight two-stage algorithm based on Best-Fit strategy and neighborhood search. Experiments under different models and network topologies show that our model placement scheme can reduce cross-Pod traffic by 35.9% and cut communication time by 18.7% compared to state-of-the-art methods.

Fault detection (FD) is essential for wafer quality control in semiconductor manufacturing (SM), as it can identify abnormal wafers in the early stages. However, frequent chamber discrepancies leads to distribution shifts in the data from different chambers, which causes performance degradation of the existing model. In this paper, we conceive a robust cross-chamber fault detection method in SM, which to the best of our knowledge is the first work that employs domain generalization to address the issue of cross-chamber fault detection in SM. Our basic idea is to map the samples from the source chambers and target chambers to the same hypersphere space, making normal samples cluster around a specific feature center while faulty samples stay away from it. Due to the scarcity of faulty samples, we propose a one-class classification-based fault detection method relying on normal samples to establish classification boundaries. To generalize the model to unseen target chambers, we design a meta-learning-based one-class domain generalization approach. We also devise a strategy to enhance distribution alignment within hypersphere, making the classification boundaries of various chambers be close to each other. The evaluations on real-world data collected from a wet process equipment in SM verify the high robustness of the model to chamber discrepancies with limited faulty samples.