In-network computing can dramatically improve the performance of application pipelines through accelerated processing and predictable chaining behavior. But to achieve practical adoption, we need to evolve in-network computing from network packet-level offloads towards application-level service. This requires networked computing nodes that preserve application semantics while executing functions on the network fast path. We design and implement such capability on a hostless FPGA framework that enables chaining of application-level functions executing on FPGAs across the network without host intervention. Our proposed hardware accelerator datapath has predictable latency and throughput, simplifying provisioning to meet application throughput. We use pre-processing for deep learning recommendation model training as a case study. Our evaluation shows that our framework adds minimal abstraction overhead while providing stable latency, reducing cross-device composition latency by 40% compared to strong CPU baselines.