About the job
P-1285
About This Role
Join our dynamic team at Databricks as a Staff Software Engineer specializing in GenAI Performance and Kernel. In this pivotal role, you will take charge of designing, implementing, and optimizing high-performance GPU kernels that drive our GenAI inference stack. Your expertise will lead the development of finely-tuned, low-level compute paths, balancing hardware efficiency with versatility, while mentoring fellow engineers in the intricacies of kernel-level performance engineering. Collaborating closely with machine learning researchers, systems engineers, and product teams, you will elevate the forefront of inference performance at scale.
What You Will Do
- Lead the design, implementation, benchmarking, and maintenance of essential compute kernels (such as attention, MLP, softmax, layernorm, memory management) tailored for diverse hardware backends (GPU, accelerators).
- Steer the performance roadmap for kernel-level enhancements, focusing on areas like vectorization, tensorization, tiling, fusion, mixed precision, sparsity, quantization, memory reuse, scheduling, and auto-tuning.
- Integrate kernel optimizations seamlessly with higher-level machine learning systems.
- Develop and uphold profiling, instrumentation, and verification tools to identify correctness, performance regressions, numerical discrepancies, and hardware utilization inefficiencies.
- Conduct performance investigations and root-cause analyses to address inference bottlenecks, such as memory bandwidth, cache contention, kernel launch overhead, and tensor fragmentation.
- Create coding patterns, abstractions, and frameworks to modularize kernels for reuse, cross-backend compatibility, and maintainability.
- Influence architectural decisions to enhance kernel efficiency (including memory layout, dataflow scheduling, and kernel fusion boundaries).
- Guide and mentor fellow engineers focused on lower-level performance, conducting code reviews and establishing best practices.
- Collaborate with infrastructure, tooling, and machine learning teams to implement kernel-level optimizations in production and assess their impacts.
