Pragmatic understanding remains a fundamental challenge in natural language processing, particularly for underrepresented languages such as Cantonese. To alleviate the scarcity of resources in this area, we introduce YUE-PUB, the first large-scale pragmatic understanding benchmark for Cantonese. Constructed from the PUB dataset through automatic translation with human verification for quality assurance, YUE-PUB covers three core pragmatic phenomena: Implicature, Presupposition, and Reference. We conduct systematic experiments across multiple baseline models to evaluate their performance under several cross-lingual transfer settings, compare models with and without pragmatic topic awareness, and further analyze their sensitivity to different pragmatic phenomena as well as their Chain-of-Thought (CoT) reasoning behavior. Experimental results show that current large language models still exhibit substantial limitations in Cantonese pragmatic understanding, underscoring the value of YUE-PUB as an important resource for advancing pragmatic reasoning research in low-resource languages.