Worked on deep learning infrastructure across mosaicml/composer, mosaicml/llm-foundry, and vllm-project/vllm-omni, focusing on hardware compatibility and performance optimization. Delivered AMD support for TE FusedAttention by refining FP8 buffer handling and removed ONNX export dependencies, using Python and PyTorch. Improved training reliability for Hugging Face models by introducing custom parameter initialization for RMSNorm layers, addressing NaN issues during FSDP meta initialization. Enhanced deployment readiness with targeted testing and configuration updates. In vllm-omni, implemented a configurable optimization to skip attention masks for zero-padded sequence parallel tokens, reducing runtime overhead and supporting flexible performance tuning through YAML-driven configuration options.