This paper proposes a novel generative pre-training framework, GPD, for spatio-temporal few-shot learning with urban knowledge transfer. The framework addresses the challenge of data scarcity in smart city applications by leveraging pre-training on optimized neural network parameters from source cities. Unlike conventional approaches that rely on common feature extraction or intricate few-shot learning designs, GPD recasts spatio-temporal few-shot learning as pre-training a generative diffusion model. This model generates tailored neural networks guided by prompts, enabling adaptability to diverse data distributions and city-specific characteristics. GPD employs a Transformer-based denoising diffusion model, which is model-agnostic and integrates with powerful spatio-temporal neural networks. By addressing challenges arising from data gaps and the complexity of generalizing knowledge across cities, GPD consistently outperforms state-of-the-art baselines on multiple real-world datasets for tasks such as traffic speed prediction and crowd flow prediction. The implementation of GPD is available at https://github.com/tsinghua-fib-lab/GPD. The framework is model-agnostic, ensuring compatibility with state-of-the-art spatio-temporal prediction models. The key contributions include: (1) leveraging pre-training paradigm to achieve effective fine-grained spatio-temporal knowledge transfer across different cities, (2) proposing a novel Generative Pre-training framework based on Diffusion models, called GPD, which leverages a Transformer-based diffusion model and city-specific prompts to generate neural networks, and (3) extensive experiments on multiple real-world scenarios demonstrating that GPD achieves superior performance towards data-scarce scenarios with an average improvement of 7.87% over the best baseline on four datasets.This paper proposes a novel generative pre-training framework, GPD, for spatio-temporal few-shot learning with urban knowledge transfer. The framework addresses the challenge of data scarcity in smart city applications by leveraging pre-training on optimized neural network parameters from source cities. Unlike conventional approaches that rely on common feature extraction or intricate few-shot learning designs, GPD recasts spatio-temporal few-shot learning as pre-training a generative diffusion model. This model generates tailored neural networks guided by prompts, enabling adaptability to diverse data distributions and city-specific characteristics. GPD employs a Transformer-based denoising diffusion model, which is model-agnostic and integrates with powerful spatio-temporal neural networks. By addressing challenges arising from data gaps and the complexity of generalizing knowledge across cities, GPD consistently outperforms state-of-the-art baselines on multiple real-world datasets for tasks such as traffic speed prediction and crowd flow prediction. The implementation of GPD is available at https://github.com/tsinghua-fib-lab/GPD. The framework is model-agnostic, ensuring compatibility with state-of-the-art spatio-temporal prediction models. The key contributions include: (1) leveraging pre-training paradigm to achieve effective fine-grained spatio-temporal knowledge transfer across different cities, (2) proposing a novel Generative Pre-training framework based on Diffusion models, called GPD, which leverages a Transformer-based diffusion model and city-specific prompts to generate neural networks, and (3) extensive experiments on multiple real-world scenarios demonstrating that GPD achieves superior performance towards data-scarce scenarios with an average improvement of 7.87% over the best baseline on four datasets.