Modern Transformer-based models increasingly employ semantically rich sequence patches to capture long-term dependencies and use channel-independent encoding to process multivariate inputs; moreover, to handle even longer time series, a growing number of models adopt the patching approach for the initial segmentation of temporal sequences. While existing patch-based tokenizers that excel at capturing long-term representations may weaken fine-grained details in each segment. Furthermore, strict channel independence is hard to enforce with shared encoders and averaged loss, as such designs might cause inter-channel crosstalk. To address these challenges, we propose Point-Patch Transformer (PPT), integrating fine-grained details into the long-term feature and enhancing channel independence. It generates a point-patch fusion representation by proposed Multi Granularity Integration composed of Patch Encoder and Point Encoder. Moreover, to achieve distinct channel independence, PPT introduces a Cross-Channel Contrastive Learning Module that enforces a precise correspondence between point and patch representations. Extensive experiments conducted on nine real-world datasets demonstrate that PPT achieves substantial improvements over the state-of-the-art in multivariate long-time series forecasting.