I'm based in Shanghai and work on high-performance AI inference. I'm also interested in fintech and autonomous driving.
- hpc-ops
- Proposed SM120 fused MoE operators for mixed-precision LLM inference.
- OpenStereo
- Built the TensorRT deployment pipeline for stereo matching, added fused CUDA preprocessing, and integrated CUDA Graphs to speed up LightStereo-S inference by ~25–30% on tested workloads.
- vLLM-Omni
- Made CUDA Graphs the default for Qwen2.5-Omni's talker, delivering a measured 2.56× end-to-end speedup on one RTX 4080 SUPER with byte-identical output.
- CVXPY
- Accelerated dense PSD checks with pivoted Cholesky, reducing initial compilation of a 400-variable Markowitz problem from 1.13 s to 181 ms in a benchmark while preserving support for singular PSD matrices.
- SymPy
- Fixed elementwise matrix and array differentiation to preserve function arguments, and enabled derivatives of Hadamard products through the generic array fallback.
- PySCF
- Fixed selected-CI initial guesses for asymmetric determinant spaces by searching the full Hamiltonian diagonal, and repaired ghost-atom and symmetry handling in Berny geometry optimization.
- CleanYourTV
- Documented a Home launcher handover and contributed a reversible five-package debloat list for a Hisense Android 9 TV, tested on-device for a week with core TV functions intact.