Worked on extending sequence length support in the ServiceNow/Fast-LLM repository by enhancing the Triton rotary kernel to enable training with context lengths beyond 65,000 tokens. This involved updating input handling and frequency calculations, as well as expanding test coverage to validate longer sequences and larger batch sizes. Addressed a kernel-level issue to improve training stability for long-context scenarios, ensuring correctness and reliability. Collaborated closely with other contributors to align on kernel improvements and maintain code quality. The work leveraged CUDA, Python, and machine learning expertise, resulting in improved capabilities for long-context LLM training and enterprise document processing workflows.