What You'll Be Doing

Write, tune, and debug GPU kernels (CUDA, Triton, or similar) for LLM training and inference operations.
Optimize performance for the latest NVIDIA GPUs, including memory hierarchy, tensor cores, and low-precision formats like FP8 and FP4.
Reduce communication costs in distributed workloads and overlap communication with computation across NVLink and nodes.
Profile end-to-end training and rollout workloads to identify and quantify compute, memory, or communication bottlenecks.
Build benchmarks, profiling tools, and regression tests to guide optimization decisions.
Collaborate with Research Scientists to implement new model architectures efficiently.
Explore emerging techniques such as low-precision training and serving, speculative decoding, and fused or persistent kernels.

Requirements

Experience in at least one area: GPU kernel development and optimization (CUDA or similar), low-latency communication and collective operations (e.g., RDMA, NCCL), or large-scale training infrastructure.
Ability to reason about hardware constraints (memory, bandwidth, interconnect topology) and diagnose bottlenecks.
Proficiency in a systems language (e.g., C++, Rust, CUDA C) and Python.
Strong debugging and profiling skills, including familiarity with tools such as Nsight.
Ability to clearly explain design and performance trade-offs to technical and non-technical audiences.

Experience with Mixture-of-Experts models, expert parallelism, or sparse model inference.
Familiarity with GPU communication libraries such as NCCL or NVSHMEM, and compute/communication overlap techniques.
Experience with Triton, CUTLASS, or similar kernel frameworks.
Knowledge of distributed training frameworks like Megatron-LM, TorchTitan, or DeepSpeed.
Background in building GPU-intensive systems in graphics, rendering, simulation, or scientific computing.
Contributions to open-source machine learning or systems infrastructure projects.