Recent advancements in neural networks have significantly improved speaker segmentation and identification across diverse scenarios. Performance of speaker diarization (SD) remains challenging in far-field environments due to background noise, reverberation, and overlapping speech. To address these challenges, this study introduces an adaptive multi-scale SD system designed for the DISPLACE 2024 dataset (Track 1). Building upon an existing speaker embedding techniques, our proposed system use a Multi-Scale Diarization Decoder (MSDD) with overlapping windows at various temporal scales to capture local and global contextual dependencies effectively. A novel scale-weighting mechanism enhances robustness by dynamically adjusting multi-scale feature contributions during inference. Additionally, a sliding window inference approach enables processing of continuous audio streams. Evaluated on the DISPLACE 2024 dataset, our method achieved a Diarization Error Rate (DER) of 9.22%, marking a 69.22% improvement over the baseline DER of 29.96%.