The Compute Express Link (CXL) protocol introduces memory semantics into solid-state drives (SSDs), enabling the construction of a unified memory-storage architecture. However, CXL’s byte-addressable access pattern is inherently mismatched with the page- or block-level granularity of conventional SSD cache mechanisms. Specifically, the large cache granularity of existing policies causes significant I/O amplification and inefficient cache space utilization when handling fine-grained requests, ultimately diminishing CXL-SSDs’ performance advantages. In contrast, simply reducing the cache granularity compromises spatial locality and increases write-back overhead. To address this dilemma, we propose Sumeru, a hybrid-granularity cache management strategy in CXL-SSDs. Sumeru designs a multi-granularity hybrid policy to handle I/O requests of different sizes, proposes a selective eviction mechanism to capture data locality within a single cache unit, and introduces a differential probation method to reduce write-back overhead. Experimental results on five real-world workloads show that Sumeru improves the cache hit rate by an average of 9.4%, reduces access latency by 76.4%, and decreases write-back frequency by 46.4% compared to existing caching strategies. These gains fully unlock the performance potential of CXL-SSDs.