Skip to content
View ChaoyuWang04's full-sized avatar
17 followers · 24 following
  • 17:23 (UTC -06:00)

Achievements

Achievement: YOLO

Achievements

Achievement: YOLO

Block or report ChaoyuWang04

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ChaoyuWang04/README.md

Hi, I'm Chaoyu Wang 👋

AI researcher working at the intersection of LLM post-training and inference systems — training agents with RL, then making them run fast.

🔭 What I'm working on

  • Agentic RL & post-training — sync vs. fully-async multi-turn tool-calling GRPO on verl; diffusion RL post-training (Flow-GRPO on Qwen-Image) with a CPU OCR reward at 6.9× headroom
  • Inference systems — hand-built MoE expert parallelism (fused grouped-GEMM Triton kernels, 1.28×), INT8 KV-cache quantization (0.50× memory, near-lossless PPL), CUDA pipeline-bubble profiling on sm_120
  • Open source — contributing to SGLang and vLLM
  • Research — conditional injection in diffusion transformers for compositional 3D scene generation

Seeking PhD / RA opportunities in LLM alignment & systems (Fall 2027).

🛠️ Tech Stack

Training: PyTorch · verl · HuggingFace Transformers · PEFT/LoRA Systems: CUDA · Triton · vLLM · SGLang · Nsight Systems/Compute Dev: Python · Docker · Next.js · PostgreSQL

🎓 Background

  • M.S. Engineering Sciences & Applied Mathematics — Northwestern University (2025)
  • B.S. Applied Mathematics — University of California, San Diego (2024)

📫 Reach me

✉️ samuelwang997@gmail.com  ·  🌐 chaoyuwang.vercel.app  ·  🤗 huggingface.co/SamWang0405

Pinned Loading

  1. Syncopate_Async_AgenticRL Syncopate_Async_AgenticRL Public

    Sync vs fully-async agentic RL on verl: multi-turn GRPO, long-tail rollout profiling, staleness ablations — quantifying when async pays off.

  2. Darkroom_VeRL-Omni Darkroom_VeRL-Omni Public

    Diffusion RL post-training on verl: Flow-GRPO on Qwen-Image, denoising-trajectory rollouts, CPU OCR reward replacing the VLM judge (6.9× headroom), budgeted for 4×5090.

    Python

  3. Laminar_GPU-Bubble-Lab Laminar_GPU-Bubble-Lab Public

    Profile-first CUDA lab for killing GPU pipeline bubbles: stream overlap, CUDA Graphs, torch.compile, spec-decode overlap — Nsight-verified

    Python

  4. Switchboard_MoE-DeepEP Switchboard_MoE-DeepEP Public

    Hand-built MoE expert parallelism: routing-skew capture, fused grouped-GEMM Triton kernels, 2-all2all dispatch/combine, overlap ablations — proving by counter-example why DeepEP exists.

    Python

  5. Halftone_QuantizedKVCache Halftone_QuantizedKVCache Public

    Data-driven INT8 KV-cache quantization: SQNR-guided granularity, near-lossless PPL (+0.2%), 0.50× memory, fused Triton int8 decode attention at 1.56× — gap to 2× fully explained.

    Python

  6. AdCampaignAgent-SFT AdCampaignAgent-SFT Public

    End-to-end training pipeline for mobile game UA tool-calling agents, covering rule-based synthetic data generation across 7 workflows, OpenAI Messages conversion, Qwen3 LoRA SFT, GRPO/RLVR alignmen…

    Python 2