Worked on the jax-ml/jax repository to enhance GPU backend capabilities, focusing on concurrent kernel execution and grid dependency control for the Mosaic GPU. Integrated Programmatic Dependency Launch (PDL) primitives using C++ and CUDA, enabling efficient concurrent kernel launches and dependency tracking within a single CUDA stream. Improved error observability for Triton kernel loading by expanding diagnostics and traceability, supporting faster debugging in production. Developed Warp-level grid dependency primitives and implemented lowering rules in Python and MLIR, allowing flexible Warp-based execution for grid-dependent workloads. Added comprehensive tests to validate new features, contributing to improved reliability and extensibility of the Mosaic backend.