Tony CW Liu

I work on GPU performance.

• Organizational & Business Impact. Rebuilt team identity during PyTorch 2.0 transition, scaling from 3 to 15 engineers across 3 time zones. Brought up 26 production models in AITemplate (10.32MW annual savings, $100M+ cost saving and billions in ad revenue). Brought up AOTInductor as the default compiler across Meta’s AI Infra org (12MW annual savings). • AI Leadership. Defined roadmap for GPU kernel and algorithm optimization across Meta's flagship AI models (Llama, Ads, RecSys). Established cross-organizational partnerships with Meta Superintelligence Labs, PyTorch Core, FAIR, and AI Infrastructure teams spanning 100+ engineers. • AI Research. Drove technical strategy resulting in industry-leading innovations: FlashAttention-3 and Flash-Decoding (Triton split-k)—algorithms now adopted across the AI industry. Led integration of advanced techniques (DeepSeek's MLA, Fused MoE) enabling Llama 4 launches. • Talent & Culture. Conducted 400+ interviews as top interviewer among 9,000 AI Infrastructure professionals. Established hiring pipeline across competitive AI talent market for PyTorch. Contributed to open-source repo (CUTLASS, CK, Triton) to enable heterogeneous GPU inference across NVIDIA and AMD, delivering performance gains from kernel to graph level. • Team Publications: https://gist.github.com/tonycwliu/edf942ec521827a3d6b18bb2e7856bab

I led a remote team optimizing Meta’s inference stack for recommendation models through quantizations and pruning. We launched 3 models that saved $3.61M annually and drove millions revenues.

• Organizational Impact. Founded and scaled Instagram's metrics infrastructure from 2 to 20+ engineers. Co-founded Instagram Metrics Council, establishing governance framework for 100+ metric products across Product, Go-to-Market, Legal, Data & Engineering. • Business Impact. Delivered $12M annual savings through global CPU optimization while maintaining 99.99% reliability for systems powering $100M+ daily ad revenue (received Instagram-wide award). Led zero-downtime migration of 40+ petabyte-scale metrics affecting 2B+ users.

One of two engineers supporting Facebook Composer across 30+ product teams; built and scaled core ML data pipelines, providing first exposure to large-scale AI infra.

I built and launched Square's eCommerce platform from concept to $60M annual revenue, integrating complex payment, tax, and shipping infrastructure during the company's IPO period.

I worked on the Display Ads team.