Software Engineer, LLVM Compiler - MTIA

About the role

Meta designs and deploys its own AI systems. MTIA, the Meta Training and Inference Accelerator, is Meta's family of in-house AI accelerator ASICs, deployed in production across Meta's data centers and expanding workload coverage from ranking and recommendation to generative AI. The MTIA Software team is part of the AI & Compute Foundation organization. Because the hardware is ours, the software is ours too: we build the entire stack a chip vendor would normally supply: compiler and toolchain, runtime, kernel libraries, developer tooling, and deep PyTorch integration, co-designed with the silicon teams generation over generation. The GenAI Frameworks org owns the layer where a model meets the machine. We take frontier models, including large language, multimodal and diffusion models, and make them run efficiently on MTIA. Our work determines how fast those models respond, how much context they can afford to hold, how many of them a given amount of silicon can serve, and how quickly a newly released model can be running and trusted in production. We are hiring a Software Engineer to own a key area of GenAI inference on MTIA, spanning serving frameworks, distributed inference, graph-mode execution and compilation, and core PyTorch integration. This is a domain-expert role: you will set technical direction for your area, drive ambiguous multi-quarter programs across team boundaries, and own the quality bar for accuracy, stability, performance, and test coverage on custom silicon.

Responsibilities

Serve as technical owner and domain expert for a key GenAI inference framework area: serving runtime, distributed inference, graph-mode execution and compilation, or core PyTorch integration Lead ambiguous, multi-quarter technical programs end to end: technical design, execution, test strategy and CI, rollout, and production hardening across teams and org boundaries Design, implement, and ship features in generative AI inference frameworks and the MTIA integration layers beneath them, from prototype through production deployment Own performance for production inference workloads end to end: profile across frameworks, runtime, compiler, and kernel boundaries, identify where time actually goes, and drive the fix to the correct layer rather than the convenient one Build and extend the distributed inference substrate: hierarchical KV caching, cross-host transfer paths, disaggregated prefetch and decode, and parallelism strategies for long-context and mixture-of-experts serving Optimize the runtime and execution path: graph capture and replay, host-side latency, memory allocation and placement, and throughput under real traffic conditions Enable frontier models on MTIA and validate accuracy against GPU baselines, closing correctness gaps and distinguishing real regressions from stale references Own the quality bar for your area, ensuring accuracy, stability, latency, and throughput, and test coverage with regression detection that keeps a win from quietly eroding Partner with kernel, compiler, and silicon teams on hardware/software co-design, producing framework-level evidence that shapes the future silicon while the design can still change Provide technical leadership: mentor other engineers, raise the design quality through reviews, and communicate decisions clearly through design documents and cross-team reviews