I'm a PhD graduate from Zhejiang University, working at the intersection of AI infrastructure and computer vision. I contribute to large-scale training frameworks (VeOmni, verl, verl-omni) and kernel DSLs (tilelang), while publishing research at top-tier venues (EMNLP, NeurIPS, ICCV, IJCAI, ECAI, IEEE TMI).

My research spans distributed training systems, vision-language models, and computational imaging.

Selected Publications:

• Dig into Multi-modal Cues for Video Retrieval with Hierarchical Alignment — IJCAI 2021 (first author)
• FlowTrain: Flow-Based Decoupled Training for Industrial-Grade Vision-Language Models — EMNLP 2026 (co-author)
• A Siamese Transformer with Hierarchical Refinement for Lane Detection — NeurIPS 2024
• SinLane: Siamese Visual Transformer via Pyramid Feature Integration for Lane Detection — ECAI 2024
• PR-Net: Preference Reasoning for Personalized Video Highlight Detection — ICCV 2021
• Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation — IEEE TMI 2021
• IFPNet: Integrated Feature Pyramid Network with Fusion Factor for Lane Detection — ICCV Workshop 2023
• Temporal Structured Illumination and Vision-Transformer Enables Large Field-of-View Binary Snapshot Ptychography — Optics Express 2024
• Propagation Dynamics of a Spatiotemporal Vortex Pulse in the Spatial Fractional System — Photonics Research 2024

📖 Full publication list: Google Scholar

🔧 Active Committer, verl-project/verl-omni — listed in official governance and co-owner of the verl_omni/trainer/ module.

I'm an active committer to verl-omni (multimodal RL post-training frameworks) and a contributor to VeOmni (model-centric distributed training recipe zoo), verl, and tilelang (high-performance kernel DSL).

Selected Contributions:

• verl-omni: Qwen3-Omni Thinker DAPO phase one, batched CLAP reward inference, strip responses from worker batches, reject silent fallbacks, and batched PickScore inference
• VeOmni: Ascend Triton clamped SwiGLU, packed ChunkMBS support, Qwen3-VL FSDP2 torch compile, HSDP all-reduce control, and deterministic mapping cycles
• verl: refresh FSDP and SkyPilot guides
• tilelang: validate T.gemm k_pack arguments

Research directions:

• Parallelism & memory optimization — packed sequence packing, ChunkMBS, FSDP2/hsdp integration
• Training-time validation — distributed metric aggregation and evaluation pipeline
• Hardware portability — NPU/Ascend enablement and cross-platform validation
• Kernel & operator development — Triton kernels for NPU, GEMM argument validation
• Multimodal reward engineering — batched reward inference and reward model integration
• Rollout & training pipeline — async semantics, rollout-train consistency, group management
• Efficient post-training — LoRA/FSDP for VLMs, distributed reward handling

Languages      Python · C++ · CUDA · Triton · Shell
Frameworks     PyTorch · VeOmni · verl · FSDP/FSDP2 · DeepSpeed
Kernels        tilelang · Triton · Ascend/NPU
Distributed    DDP · HSDP · FSDP · Tensor Parallel · Expert Parallel · ChunkMBS
Infrastructure Slurm · Ray · wandb · Docker
Vision/RL       Diffusion Models · VLMs · Video Retrieval · Medical Imaging

• 📧 Email: w_wenzhe[AT]163.com
• 🏠 Location: Hangzhou, China

_{💫 "From research papers to production training frameworks — bridging the gap between algorithms and infrastructure."}