Large-scale tensor completion plays a crucial role in data analysis and anomaly detection, with Alternating Least Squares (ALS)-based CANDECOMP/PARAFAC(CP) decomposition being widely adopted due to its convergence properties and computational stability. However, when accelerating ALS computation on GPUs, load imbalance caused by data partitioning significantly affects efficiency. Due to the sparsity and heterogeneity of data, different thread blocks handle varying amounts of non-zero values, leading to suboptimal utilization of computational resources. Moreover, updating factor matrices in ALS involves frequent global memory accesses, where the latency is 100 times higher than that of shared memory. Efficient utilization of shared memory is therefore critical for improving computational performance. To address these challenges, we propose a GPU-optimized ALS framework that incorporates a heat conduction-based load balancing strategy and a shared memory acceleration mechanism. The load balancing strategy dynamically adjusts subtensor partitioning based on the distribution of non-zero values, ensuring balanced workload allocation across GPU resources. Meanwhile, the shared memory acceleration mechanism caches frequently accessed factor matrices and employs element-wise implicit computation, eliminating explicit intermediate matrix storage and thereby reducing memory overhead and global memory access latency. Based on this, a comparison was made with the other three methods on four data sets. While ensuring accuracy, the time and memory usage were greatly reduced, providing a practical solution for efficient tensor completion.