FADP-GT: A Frequency-Adaptive and Dual-Pooling Graph Transformer Model for Device Placement in Model Parallelism

The increasing scale and complexity of graph-structured data necessitate efficient parallel training strategies for graph neural networks (GNNs). The effectiveness of these strategies hinges on the quality of graph feature representation. To this end, we propose a Frequency-Adaptive Dual-Pooling Graph Transformer (FADP-GT) model to enhance feature learning for computational graphs. We propose a Frequency-Adaptive Dual-Pooling Graph Transformer (FADP-GT) model, which incorporates two modules: a Frequency-Adaptive Graph Attention (FAGA) module and a Dual-Pooling Feature Refinement (DPFR) module. The FAGA module adaptively filters frequency components in the spectral domain to dynamically adjust the contribution of high- and low-frequency information in attention computation, thereby enhancing the model’s ability to capture structural information and mitigating the over-smoothing problem in multi-layer network propagation. On the other hand, the DPFR module refines graph features through dual-pooling operations—Global Average Pooling (GAP) and Global Max Pooling (GMP)—along the node dimension, which captures both global feature distributions and salient local patterns to enrich multi-scale representations. By improving graph feature representation, our FADP-GT model indirectly supports the development of efficient device placement strategies, as enhanced feature extraction enables the more accurate modeling of node dependencies in computational graphs. The experimental results demonstrate that FADP-GT outperforms existing methods, reducing the average computation time for device placement by 96.52% and the execution time by 9.06% to 26.48%.