Transformer-based large language models (LLMs) have become a dominant force in natural language processing, advancing both research and industry. As model sizes have grown from billions to hundreds of billions of parameters, training them on a single GPU is no longer feasible, making distributed training essential. Existing distributed training methods partition models into non-overlapping segments of equal-size to cut memory usage on each device, but this creates cross-device data dependencies. These dependencies lead to delays and reduced throughput, particularly during the forward phase. This paper introduces AsymFB, a novel model parallelism algorithm designed to address these limitations. By adopting an asymmetric partitioning approach, AsymFB assigns varying numbers of devices to different training phases—forward, recompute, and backward—optimizing resource allocation and minimizing performance penalties. Our approach demonstrates up to a 1.25 \(\times \) speedup compared to traditional methods, significantly improving the efficiency of training large-scale LLMs and mitigating common performance bottlenecks.