Magnetic resonance imaging (MRI) is a vital non-invasive, radiation-free imaging tool that provides excellent soft-tissue contrast for visualizing anatomical structures and physiological functions, making it invaluable for clinical diagnosis. However, prolonged MRI scan time can reduce examination efficiency, increase motion artifacts, and compromise patient comfort, posing particular challenges for individuals with claustrophobia or physical injuries. Moreover, existing deep learning–based methods often require retraining the same model multiple times for different acceleration settings, leading to inefficient use of computational resources and limited flexibility in clinical practice. In this paper, we propose a Sampling-guided All-in-one Learning framework (SAMR) for accelerated MRI reconstruction, offering a unified solution that enables high-quality image reconstruction across diverse acceleration settings with a single model. SAMR adopts a deep unrolling architecture that integrates a learning-based reconstruction network with k-space consistency operations, alternating reconstruction between the image domain and k-space domain. Specifically, we design a sampling-guided vision state space model as the reconstruction network, capable of simultaneously capturing both coarse-grained and fine-grained contextual information. In addition, we investigate incorporating sampling mask information into the reconstruction process to enhance adaptability to varying acceleration factors, and propose a sampling fusion module to effectively integrate this prior knowledge with learned features. Extensive experiments on two public datasets demonstrate that SAMR achieves superior reconstruction performance and consistently outperforms state-of-the-art methods.