I'm a master's student in Computer Science at Xidian University, expecting to graduate in June 2027.
Currently working on:
- 🔨 LLM systems and memory-efficient training and inference, with experience in
- KV cache management
- FP8/LoRA optimization
- RISC-V vector optimization.
- 🔍 researching on speculative decoding
- draft tree
- on-policy distillation (OPD)
Reduce redundant resident FP8 weight storage in TransformerEngine during LoRA fine-tuning.
- Reuse rowwise FP8 weights for forward and input-gradient computation, eliminating the persistent columnwise weight copy.
- Implement a custom FP8 dGrad path with the corresponding Transformer Engine integration.
- Reduce resident FP8 weight data from approximately 2 to 1 byte per parameter. Temporary layout buffers are still used during computation.
MCore-Bridge: FP8 Weight-Loading Fix · Merged
Diagnosed abnormally small training gradients caused by stale columnwise FP8 weight data after checkpoint loading. Fixed the loading path to regenerate columnwise data, restoring correct input-gradient computation.
Layer and Time-Adaptive KV Cache Pruning for Reasoning-Intensive LLM Serving
- Adaptive KV cache pruning for memory-efficient LLM serving.
- My contribution: system architecture implementation and integration into the inference framework.
A stage-aware automated optimization framework for LLM inference on RISC-V.
- Stage- and memory-access-aware vectorized kernels using RISC-V Vector Extensions (RVV).
- My contribution: kernel implementation and optimization search.
- Tencent · TEG AI Platform Department — LLM training framework adaptation and deployment across NVIDIA GPUs and Huawei Ascend NPUs; ms-swift integration for LoRA and full-parameter fine-tuning. (Mar–Aug 2026)
- Institute of Software, Chinese Academy of Sciences — RISC-V RVV optimization and testing for musl libc and glibc. (Aug–Nov 2025)
- Institute of Software, Chinese Academy of Sciences — EulixOS package building, RVV optimization, and testing on RISC-V. (Jul–Oct 2024)
| Area | Experience |
|---|---|
| LLM systems | KV cache management, FP8/LoRA memory optimization, training issue diagnosis, inference framework integration |
| Frameworks | vLLM, ms-swift, Transformer Engine |
| Platforms & performance | NVIDIA GPUs, Huawei Ascend NPUs, RISC-V RVV; CUDA operator implementation for FP8 weight-layout handling |
| Programming & tooling | C/C++, Python, Go; Linux, Git, Docker, software builds and container delivery |
- National runner-up, 2nd China Graduate Operating System Open Source Innovation Competition (2025), for RISC-V vector and matrix operator optimization for LLM applications.
- I also enjoy systems programming, networking, and open-source tools.
- Groxy: a proxy using TLS for encryption and authentication written in Golang, which is my undergraduate thesis project.
- Writing: Horizon Blog NEO · Older blog.
- GPG public-key records: Rev03 · Earlier key.