The Machine Learning Frameworks team exists for two reasons: making the research iteration loop as fast and frictionless as possible, and making production run as fast as the hardware allows. Both matter, and figuring out how to do both at once is most of what makes the work fun.
Concretely, we build optimisers, code generators, runtime libraries, and the orchestration layer that runs computation across heterogeneous hardware. The stack spans C++, Rust, Python, MLIR, and CUDA. You don’t need to know all of these things but should want to work at this level of the stack.
The work is broad. Some of it is high-level, looking at the whole compute pipeline, orchestrating the system and creating a cohesive product. Some of it is in the details, getting into hardware behaviour, dissecting compute graphs, squeezing the last bit of performance out of the hot path. Our job is managing the whole stack, top to bottom.