I'm a master's student in Computer Technology at Northwestern Polytechnical University (NWPU). I work on AI infrastructure for distributed training, efficient inference, and reinforcement learning (RL) post-training. My work includes GPU/NPU kernels, KV cache management, and rollout and reward systems.
I'm looking for internships in LLM inference optimization and AI systems, remote or on-site.
- vLLM-Omni: Fixed a Qwen3-Omni Code2Wav dtype mismatch by preserving the embedding dtype after the batch-invariant mean, with regression coverage for BF16, FP16, and FP32 (#8327).
- SGLang-Omni: Made PersonaPlex caller-history updates incremental on CPU and GPU, avoiding repeated full-history copies while preserving history for re-prefill (#3103).
- verl-omni: Added opt-in FSDP2 sharding for frozen PickScore reward models across two GPUs, with lifecycle and failure cleanup handling (#855). Initial support targets single-node synchronous diffusion training.
- UniRL: Made WAN2.1 denoising noise reproducible by sample identity across rollout microbatches (#652).
- verl-omni: Added CPU placement for native reward models (#694) and migrated the SD3.5 OCR recipe to named rewards (#695).
- vLLM-Omni: Added remote SeedVR2 video restoration for ComfyUI (#8505).
- verl-omni: Shared-GPU asynchronous diffusion rollout (#742); per-sample named reward streaming (#730), audio inputs for shared reward managers (#920), and safe rollout restoration after synchronous validation scoring (#929).
- Megatron-LM / Transformer Engine: GDN recurrence activation offloading (#7908) and regression coverage (#8243); context-parallel scheduling metrics (#8431); draft work on partial CUDA Graph capture with FSDP v2 and MoE (#9112) and warmup failure cleanup (TE #3862).
- NVIDIA Model Optimizer / Megatron-LM: Proposed streaming vLLM teacher features into Megatron-LM for EAGLE3 fine-tuning (RFC #3474; draft PRs #3473 and #9550).
- vLLM / vLLM-Omni: Multimodal prefix-cache boundary handling for sibling requests on hybrid Mamba models (vLLM #56818); configurable pipeline stage transitions (Omni #8745); proposed same-policy diffusion request recovery (Omni RFC #10529).
- SGLang-Omni: PersonaPlex session LM warmup before server readiness (#3104) and bounded batching of queued session audio (#3591).
- Mooncake / llm-d: Encoder–Prefill–Decode disaggregation and reusable agent state (Mooncake #2836); proposed a shared latency feature schema across llm-d's router and predictor (llm-d RFC #4218).
- UniRL / NVIDIA recsys-examples: Wan2.1 on-policy distillation recipe (UniRL #651); proposed receiver-side DynamicEmb row-index resolution during input distribution (RFC #552, draft prototype #551).
View my upstream pull requests
- MiniCPM-o on Ascend: Optimized multimodal inference across two NPUs, including CPU/NPU overlap, CANN operators, graph execution, and W8A8 quantization.
- Efficient VLM inference: Optimized Qwen3-VL-2B with KV cache reuse, projection fusion, Triton kernels, and CUDA Graphs.
- MetaX C500 kernels: Built and tuned fused quantization kernels with TileLang, profiling, and correctness tests.
- LLM inference, multimodal serving, KV and prefix caching
- Distributed training, activation offloading, and runtime scheduling
- Quantization, GPU/NPU kernels, profiling, and performance analysis
Python · PyTorch · CUDA · Triton · TileLang · CANN / AscendC · vLLM · llama.cpp
- MiniCPM × Ascend Inference Optimization Challenge: 5th on the public leaderboard
- IEEE AICAS 2026 Grand Challenge: 7th place
- CCF Open Source Innovation Competition: Third Prize, Mooncake KVCache
- Huawei MindSpore Model Development Challenge: 3rd place / Silver, S1-MoE
- MetaX Open Source Talent Camp: Best Engineering Practice Award
- Northwestern Polytechnical University: Master's in Computer Technology, 2025–present
- Shaanxi Normal University: Bachelor's in Software Engineering, 2021–2025

