I am a graduate student at Xi'an Jiaotong University (XJTU), focusing on AI infrastructure, LLM serving, and RL post-training systems.
I enjoy turning systems ideas into practical open-source implementations: efficient rollout execution, distributed training workflows, weight synchronization, and cross-platform GPU optimization for GRPO-style workloads.
- 🔭 Currently contributing to: vLLM-Omni, a framework for efficient omni-modality model inference and serving.
- 🚀 Main work: Leading Vime framework research, fork-roadmap planning, and PR delivery across CUDA and ROCm; integrating Vime with RL-Kernel for reproducible RL training and rollout.
- 🔬 Research interests: Efficient inference, distributed attention, GRPO/RLHF systems, linear-logp providers, and cross-platform GPU performance.
| Project | Focus | Status |
|---|---|---|
| vLLM-Omni | Efficient omni-modality model inference and serving in the vLLM ecosystem | 🔥 Contributing |
| Vime | RL framework integration, roadmap planning, and end-to-end training/rollout validation | ⚡ Lead contributor |
| RL-Kernel | GPU kernels and strict runtime validation consumed by Vime on CUDA and ROCm | 🤝 Contributor |
| Area | Selected Work |
|---|---|
| Vime framework and delivery | Framework investigation, fork-version roadmap planning, upstream PR delivery, and reproducible experiment documentation (PR #409, PR #287) |
| CUDA + ROCm integration | Led the Vime provider boundary for linear log-probabilities on both GPU stacks, preserving Vime's loss semantics and native fallback (CUDA PR #424, ROCm PR #423) |
| Distributed Attention | Developed and validated paged/CP attention paths, including FlashInfer RoPE-fused attention, CP drift checks, and bitwise ROCm schedules (PR #279, PR #284, PR #319, PR #394) |
| Deterministic runtime and performance | CUDA Graph capture, tensor-parallel all-reduce optimization, strict runtime modes, and cross-configuration kernel validation (PR #344, PR #360, PR #365, PR #367) |
| End-to-end RL validation | Ran matched native/provider train-rollout consistency experiments, TP/CP ablations, bitwise checks, throughput profiling, and performance tuning on Qwen3 workloads (PR #377, PR #388, PR #400, PR #403) |
| Linear-logp experiments and integration | Designed the structured provider contract, TP vocabulary partition handling, autograd checks, and technical write-ups for portable log-probability execution (PR #336, PR #352, PR #338) |
| Languages & Core | AI Infrastructure | Distributed & Tooling |
|---|---|---|
   |
   |
   |
|
|
|
Building efficient systems, one kernel and one iteration at a time.