Sentiment analysis is a fundamental task in natural language processing, yet it remains challenging for low-resource languages such as Vietnamese due to limited annotated data and the scarcity of high-capacity pretrained models. This paper investigates an efficient adaptation framework for Vietnamese sentiment classification by fine-tuning Vistral-7B-Chat, a Vietnamese-centric large language model, using parameter-efficient fine-tuning techniques. Specifically, we integrate Low-Rank Adaptation (LoRA), 4-bit weight quantization, and mixed-precision training to substantially reduce computational and memory overhead, enabling practical fine-tuning on consumer-grade hardware. To further mitigate data scarcity, we adopt an instruction-based data augmentation strategy that leverages GPT-4o-mini to generate sentiment-consistent synthetic samples, thereby increasing linguistic diversity while preserving label fidelity. Extensive experiments on two standard Vietnamese sentiment benchmarks demonstrate the effectiveness of the proposed framework. On VLSP 2016, our model achieves 83.1% accuracy and 83.3% macro F1-score, while on AIVIVN 2019 it attains 94.7% accuracy and 94.5% macro F1-score. These results correspond to improvements of up to 23.3% points over the non-fine-tuned Vistral-7B-Chat baseline and up to 11.5% points over a strong zero-shot GPT-4o baseline. Overall, the findings highlight the advantages of language-specific large language model adaptation combined with parameter-efficient optimization and controlled synthetic data augmentation. The proposed framework offers a scalable and resource-efficient solution for Vietnamese sentiment analysis and provides insights applicable to broader natural language processing tasks in under-resourced languages.