About the role
Inferact's mission is to grow vLLM as the world's AI inference engine and accelerate AI progress by making inference cheaper and faster. Founded by the creators and core maintainers of vLLM, we sit at the intersection of models and hardware—a position that took years to build.
About
the Role We're looking for a TPU and AMD GPU performance engineer to make vLLM a first-class inference engine across non-NVIDIA accelerators. Frontier inference cannot be locked to one hardware stack. As AMD GPUs, TPUs, and other accelerators become increasingly important, vLLM needs backend paths that are fast, correct, benchmarked, and maintainable across heterogeneous hardware platforms. You'll build and optimize AMD GPU and TPU backends, kernels, compiler integrations, runtime paths, and benchmarking infrastructure. You'll work at the boundary of inference systems, kernels, compilers, and hardware architecture, improving paths such as attention, GEMM, sampling, KV-cache, communication-heavy operations, and model serving on non-NVIDIA hardware. Your work will directly impact how broadly and efficiently the world can run AI inference with vLLM. Skills and
Qualifications Minimum
qualifications: Bachelor's degree or equivalent experience in computer science, engineering, machine learning systems, hardware systems, compilers, or similar. Hands-on experience optimizing workloads on AMD GPUs, TPUs, or another non-NVIDIA accelerator stack. Experience with AMD ecosystem tools such as ROCm, HIP, Triton, CK, AITER, or equivalent GPU performance libraries and tooling. Experience with TPU, XLA, JAX, Pallas, or related compiler and runtime tooling for accelerator workloads. Ability to optimize ML inference paths such as attention, GEMM, sampling, KV-cache, fused kernels, backend runtimes, or communication-heavy operations. Strong performance profiling and benchmarking discipline, including tokens/second, latency, throughput, correctness parity, hardware counters, and reproducible measurement methodology. Ability to navigate immature tooling, incomplete documentation, backend-specific rough edges, and cross-platform performance differences without getting stuck. Preferred