Developed and integrated activation checkpointing for the DeepseekV4 model within the AI-Hypercomputer/maxtext repository, focusing on improving memory efficiency and scalability for large-scale deep learning workloads. The work involved implementing a unified kv_proj layer and systematically applying checkpointing tags to key attention components, ensuring consistent memory-optimized behavior across the model. Leveraging Flax, JAX, and Python, the developer emphasized test-driven development by adding unit tests to verify correct application of checkpointing tags in the forward pass, thereby maintaining regression protection. This feature enables deployment of larger contexts in constrained environments and demonstrates a strong focus on model-layer integration and code quality.