I'm SIRU HE, a Master's student in Electronic Information at SUSTech, jointly trained with SIAT-CAS, focused on LLM inference and AI infrastructure.
My internship experience sits at the inference-system and kernel layers: at ModelBest I worked on the multimodal inference runtime for MiniCPM-o 4.5 and VoxCPM2; at Shanghai Guangyu Xinchen I optimized LLM Decode-stage operators on a self-developed edge NPU and worked on Decode-stage megakernel fusion.
- CUDA / Triton kernels — memory access, Tensor Core, deterministic computation; cycle-level optimization of memory-bound operators
- Inference engines — from building an engine to tuning it: TP sharding, KV Cache management, MoE execution, CUDA Graph, speculative decoding
- Serving & framework contribution — internals of vLLM, Megatron, llama.cpp/ggml; stability and performance PRs to vLLM-Omni
- Distributed & train-serve consistency — Batch Invariance and deterministic reductions; bitwise-identical logprobs on 8×H100
- Profiling & tooling — Nsight Compute, PyTorch Profiler, Perf, GDB; CI and automation workflows
Tech stack: C++ / CUDA / Triton / Python / PyTorch / vLLM / Megatron / llama.cpp (ggml) / NCCL / Docker / CMake / Git
Publications: SteadyFlow-Edge (first author, ACAI), RT-EdgeDetect (first author, APPT)
Email: siruhe666@gmail.com
GitHub: frank-2077