HolographicType Communication (HTC) demands ultra-high bandwidth, substantial computing resources, ultra-low latency, and precise synchronization that exceed the capabilities of existing Radio Access Networks (RANs). These requirements exacerbate the inherent tradeoff between centralized baseband processing and fronthaul bandwidth efficiency, making traditional coarsegrained baseband function deployment and lightpath provision for unicast services insufficient to efficiently support HTC. In particular, 6G AI-RAN, by combining flexible spectrum and baseband processing allocation (i.e., by flexible-unit deployment), offers a promising solution for transmitting holograms and multisensory data. However, an efficient resource assignment in a 6G AI-RAN requires solving multiple sub-problems, involving Flexible Unit (FU) deployment and Holographic Routing and Spectrum Allocation (FHRSA). We formulate the FHRSA problem as an Integer Linear Programming (ILP) model and propose HoMa, a holographic viewpoint rotation prediction (HVRP)-based multi-agent deep-deterministic policy gradient (MADDPG) algorithm, to efficiently address it. HoMa employs a long-short-term memory (LSTM) network for HVRP, enabling accurate estimation of future resource demands. Leveraging these predictions, HoMa explicitly models inter-agent interactions and optimizes high-dimensional actions, achieving faster convergence, enhanced cooperation, and near-global optimal resource allocation. Simulation results demonstrate that the proposed HoMa algorithm significantly outperforms the benchmark algorithms in both small- and large-scale networks, achieving reductions of 20.29%, 38.96%, and 28.24% in multi-objective optimization cost, computational cost, and spectrum resource usage, respectively. Performance is closely aligned with the ILP model, validating its efficiency in resource allocation optimization.