Worked on enhancing hardware compatibility and training reliability for deep learning models in the mosaicml/composer and mosaicml/llm-foundry repositories. Addressed cross-hardware support by enabling TE FusedAttention on AMD GPUs, removing the FP8 buffer export requirement to streamline precision handling. Improved large-model training stability by fixing NaN issues during FSDP meta initialization for Hugging Face models, introducing custom parameter initialization for layers such as RMSNorm. Added targeted tests and configuration updates to ensure reproducibility and deployment readiness. Utilized Python and YAML alongside PyTorch and Transformer Engine, demonstrating depth in performance optimization, model initialization, and hardware acceleration within deep learning frameworks.