In this paper, we present novel methods for dynamically streaming parameters in order to implement a traditional CNN architecture. We also propose a library-based approach for designing scalable and dynamic distributed CNN inference on the fly using partial-reconfiguration techniques, which is particularly appropriate for resource-constrained edge devices. To demonstrate the concept, the proposed approach is implemented on the Xilinx PYNQ-Z2 board using the LeNet-5 CNN model. The results demonstrate that the proposed methodologies are effective, with classification accuracy rates of 92%, 86%, and 94%, respectively.