About

I am an AI Framework Researcher at the Beijing Academy of Artificial Intelligence (BAAI). I care about language models people can run locally—privately, on hardware they own—and whether small or compressed models can still support coding agents and multi-step reasoning, instead of sending every prompt through an untrusted cloud API.

My current systems work is Torch-FL, a PyTorch device plugin in the FlagOS stack that exposes a unified flagos device across accelerators. I treat that stack as a way to measure what survives off datacenter GPUs, not as the research question itself.

From July 2021 to July 2026 I was a Senior AI Infrastructure Engineer at Huawei on the MindSpore team (MindNLP, MSAdapter / MindTorch). I received my M.S. and B.S. in Computer Science from Chongqing University; my thesis was on multi-domain abstractive summarization.

Research interests

Small and compressed language models under tight compute; agentic skills (tool use, coding agents, long-horizon tasks); efficient and heterogeneous LLM inference.

News

• 2026.09   Gave Torch-FL: A Virtual Device Layer for PyTorch at PyTorch Conference China 2026
• 2026.07   Joined BAAI as an AI Framework Researcher
• 2026.05   Open-sourced PocketLLM, starting from DeepSeek-V4-Flash on 4× RTX 2080 Ti
• 2026.04   Fate accepted to The ACM Web Conference (WWW 2026)

Selected publications

• Klotski (ASPLOS 2025). Expert-aware multi-batch pipelining for MoE inference. I co-designed the prefetch and offload path. arXiv · PDF
• Fate (WWW 2026). Cross-layer gating for faster MoE inference on edge devices. arXiv

All publications →

Selected systems

• Torch-FL (BAAI, Jul. 2026 – present): PyTorch device plugin for FlagOS. One flagos device (PrivateUse1) with per-operator routing to vendor kernels, FlagGems/Triton, compatibility boxing, or CPU fallback.
• PocketLLM (2026 – present): C++/CUDA and PyTorch inference on consumer GPUs. Started as DeepSeek-V4-Flash on 4× RTX 2080 Ti; now also covers MiniMax-M2.7, GLM-5.2, and Qwen3.8-27B-FP8. On that 2080 Ti box, C++ FP4 decode is about 3.7 tok/s—enough to run a large MoE at home, not enough for productive coding-agent work. Docs · Technical report (PDF)