High-performance computing faces a significant productivity gap between high-level programming languages used for rapid prototyping and low-level languages required for optimal performance. This thesis investigates whether large language models can bridge this gap by automating code translation from sequential Python to parallel implementations
in NumPy, JAX, C++ with OpenMP, and CUDA. A novel benchmark covering core HPC computational patterns was constructed and used to evaluate DeepSeek-Coder models ranging from 1.3B to 33B parameters. Base models achieved moderate success on simpler translations but struggled significantly with CUDA and tensor frameworks. Supervised fine-tuning on a manually curated dataset substantially improved both functional correctness and performance optimization, outperforming models trained on synthetic data. The findings indicate that while current-generation coding LLMs cannot yet guarantee sufficient reliability for production HPC workflows, fine-tuning on high-quality datasets shows promise for improving their code translation capabilites