ML Systems | Ying Liu

Many modeling ideas live or die on systems: orchestration, multi-dimensional parallelism, reward-model offloading, dynamic batching, kernel-level efficiency, and the launch machinery that gets a model from a training run into a product. I have built and led such platforms for LLM post-training, for diffusion and transformer video models, and for real-time production services in ads and search, including budget pacing and RL-based bidding for Meta Ads, whole-page ranking for Facebook Search, and predictive models, smart bidding, and pricing for Google Hotel Ads.

A recent example is TLX (Guan et al., 2026), Triton language extensions for warp-group-level orchestration of modern GPUs, whose kernels are deployed in large-scale training and inference systems.

References

2026

arXiv
TLX: Hardware-Native, Evolvable MIMW GPU Compiler for Large-scale Production Environments

Yue Guan, Hongtao Yu, Peng Chen, Daohang Shi, Karthik Manivannan, Nicholas J. Riasanovsky, Manman Ren, Lei Wang, Shane Nay, Partha Kanuparthy, Zhijing Li, Ying Liu, Zaifeng Pan, Zhengding Hu, and Yufei Ding

arXiv preprint arXiv:2605.10905, May 2026

arXiv Bib Code
@article{guan2026tlx, title = {{TLX}: Hardware-Native, Evolvable {MIMW} {GPU} Compiler for Large-scale Production Environments}, author = {Guan, Yue and Yu, Hongtao and Chen, Peng and Shi, Daohang and Manivannan, Karthik and Riasanovsky, Nicholas J. and Ren, Manman and Wang, Lei and Nay, Shane and Kanuparthy, Partha and Li, Zhijing and Liu, Ying and Pan, Zaifeng and Hu, Zhengding and Ding, Yufei}, journal = {arXiv preprint arXiv:2605.10905}, year = {2026}, month = may, }