Job description

We are looking for an experienced engineer to lead performance autotuning on MTIA — Meta's custom training and inference accelerator. You will lead the MTIA Software Performance Autotuning team (part of Infra Foundations) and own how we extract maximum performance from our hardware, automatically and at scale. Every kernel, every compiled graph, and every runtime configuration has a large space of possible implementations — tile sizes, scheduling, memory layouts, fusion decisions, precision choices — and the right one depends on the chip, the model, and the shape. Hand-tuning does not scale. The team's core mission is to make MTIA fast by default: building the search infrastructure, cost models, and tuning methodology that finds the best configuration without a human in the loop. As a technical leader, you will define our autotuning strategy, architect the search and benchmarking infrastructure, and partner closely with compiler, kernel, runtime, and product (e.g., Ads Ranking, Recommendation Systems, GenAI) teams to turn performance headroom into shipped speedups. Autotuning operates across the full MTIA software stack — FX graphs, compiler, kernels, runtime, PyTorch — which means lots of cross-team collaboration. We partner closely with machine learning engineers across Ads, Instagram/Facebook, and Meta Superintelligence Labs teams whose models run on MTIA. Own the performance autotuning strategy for MTIA inference and training — current and future chip generations Build and maintain our autotuning infrastructure: search spaces, cost models, benchmarking harnesses, and performance regression detection Define how we trade off search cost against performance gains, and where autotuning fits in the compile/deploy pipeline Partner with kernel and compiler teams to expose tunable parameters and encode them into an automated search Debug performance issues that span multiple layers — kernel, runtime, compiler, or serving infra — and drive them to root-cause Collaborate with product teams to accelerate model performance on MTIA Set technical direction and help the team grow. We've scaled fast and need people who can elevate the impact around them, not just ship their own code Bachelor's degree in Computer Science, Computer Engineering, relevant technical field, or equivalent practical experience Strong Python and C++ skills, with hands-on experience across the PyTorch stack 4+ years in ML systems, AI infra, performance engineering, or similar Experience driving problems that span multiple teams, where no one owns the whole picture Experience with accelerator performance concepts — roofline analysis, memory bandwidth, occupancy, and what makes kernels fast or slow on hardware Track record of setting technical direction and mentoring engineers BS in CS, CE, Math, or equivalent experience Experience adhering to and implementing responsible, ethical AI practices (e.g., risk assessment, bias mitigation, quality and accuracy reviews) Demonstrated ability to integrate AI tools to optimize/redesign workflows and drive measurable impact (e.g., efficiency gains, quality improvements) Demonstrated ongoing AI skill development (e.g., prompt/context engineering, agent orchestration) and staying current with emerging AI technologies Experience with production ML models (recommender systems, LLMs, ranking) is useful but not required Hands-on experience with torch.compile, TorchInductor, or other ML compiler stacks (XLA, TVM, MLIR, Triton) Experience with autotuning, cost models, or search-based optimization (e.g., Ansor, AutoTVM, learned schedulers) MS or PhD in CS, CE, compilers/systems, or related Experience with hardware bring-up or accelerator development (GPU, TPU, or custom ASIC) Experience with kernel-level performance optimization — tiling, scheduling, memory layout, fusion, and how they interact on real hardware

Requirements

Requirements not extracted — see the original description below.

Application prep

Use the role’s language and application details while tailoring your materials.

Keywords to mirror

machine learningpythonmlai

Platform: meta
Effort: medium