Abstract
—Cell-free massive multiple-input multiple-output (CF-mMIMO) systems can provide high spectral efficiency and uniform service. However, downlink beamforming can incur substantial two-way fronthaul overhead for network-wide channel state information (CSI) collection at the central processor (CP) and beamformer delivery to the access points (APs). Fully distributed methods reduce this burden. However, they have limited instantaneous network-wide channel awareness, while centralized or iterative methods retain costly information exchange. We propose a non-iterative distributed beamforming framework based on an approximate factorized representation. Each AP constructs its local beamformer from local CSI and a common Gram-dependent reconstruction factor. The shared Gram matrix provides the information required to infer three reconstruction parameters for each user equipment (UE) and is exchanged through one non-orthogonal aggregation–multicast round whose signaling dimension is independent of the AP and per-AP antenna populations. We realize the resulting Gram-to-parameter mapping using a scalable hierarchical-attention architecture that is permutation-equivariant with respect to UE ordering and supports fixed-model reuse across system sizes. Numerical results show effective sum-rate and min-rate improvements of 18% and 26%, respectively, over an iterative coordination benchmark.