AI systems first · AI for Science alongside
Master's student at Zhejiang University. I build reproducible systems to understand language models, with a growing focus on reliable inference and stateful agent APIs.
vLLM-Omni #7065 fixes Higgs Audio v3 voice-clone token validation while preserving reference-audio placement across chunked prefill. Merged after collaborator review on September 4, 2026 (UTC).1
| Track | Engineering focus | Public work |
|---|---|---|
| Agentic API | Bounded response sessions; continuation and storage contracts | #257 — draft; #258 — open2 |
| vLLM-Omni | Audio diagnostics; request lifecycle and timeout correctness | #7098, #7151 — open3 |
| DeePMD-kit | Ragged graph batching; charge/spin and numerical correctness | #6008, #6010 — open4 |
| MACE | Scientific model-download reliability | #1712 — open5 |
These are contribution areas I am working toward maintaining, not assigned module ownership. Open and draft PRs are not accepted contributions. The dated contribution log separates merged work, review candidates, and unpublished experiments.
| Project | Focus |
|---|---|
| wechat-ai-memory | Local-first, traceable conversation context |
| CS336 Assignment 1 | From-scratch language-model foundations |
| cs336.2026 | Systems-first learning and implementation notes |
- Reproduce failures and preserve passing controls before changing behavior.
- Keep patches scoped, with explicit dependency and compatibility boundaries.
- Separate local tests, upstream CI, and actual model/hardware validation.
- Disclose AI assistance and respond to review with reproducible evidence.
Python · Rust · PyTorch · pytest · Git · uv
Build the smallest useful version. Test the assumptions. Explain what happened.
Footnotes
-
vLLM-Omni. Merged PR #7065. ↩
-
Agentic API. Core session draft #257, typed Responses file validation #258. Status checked September 6, 2026, 20:59 UTC; no WebSocket integration or module ownership is claimed. ↩
-
vLLM-Omni. Reference-audio diagnostics #7098, shared RPC deadline #7151. Open at the same check. ↩
-
DeePMD-kit. Ragged charge/spin batching #6008, padding-safe force/Hessian metrics #6010. Open at the same check. ↩
-
MACE. Reject HTML before caching model downloads #1712. Open at the same check. ↩