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
DeePMD-kit #6010 excludes virtual
atoms from force, Hessian and spin-force error metrics while preserving padded detail
outputs. Approved and merged by project member njzjz on September 7, 2026 (UTC).2
| Track | Engineering focus | Public work |
|---|---|---|
| Agentic API | Bounded response sessions; continuation and storage contracts | #257 — draft; #258 — open3 |
| vLLM-Omni | Audio diagnostics; request lifecycle and timeout correctness | #7098, #7151 — open4 |
| Vime | Post-training data-source ordering and resume correctness | Issue #414 — reported; local patch unpublished |
| XGrammar | Structured-generation tool input validation | #881 — draft; local regression verified, awaiting review5 |
| DeePMD-kit | Ragged graph batching; charge/spin and numerical correctness | #6008 — open at last check6; #6010 — merged2 |
| MACE | Scientific model downloads and TorchScript export | #1712 — approved, not merged; #1717 — draft7 |
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 The dated contribution log separates merged work, review candidates, and unpublished experiments.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 |
Case study: honoring Router queue deadlines under contention — token, queue and HTTP regressions; CI-pinned Rust validation; PR #248 submitted after personal review and sign-off; awaiting upstream review, not merged.
Case study: preserving heterogeneous conversation data in GuideLLM — identical baseline/candidate regressions and a real CLI-to-localhost HTTP check; issue evidence shared, patch unpublished and personal review pending.
Case study: preserving Vime's sample stream across epochs — matched baseline/candidate tests and real cursor save/load; issue reported, human verification and upstream agreement pending, not an accepted fix.
Case study: validating an XGrammar parser-history regression — attributed fork candidate with deterministic tests and bytewise mask comparisons; upstream coordination pending, not merged.
Case study: checkpoint-save failure and distributed recovery — real CPU/Gloo failure reproduction with a matched successful-save control; reported upstream, no fix or merge claimed.
- 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. ↩
-
DeePMD-kit. Merge commit 28b7d068. Merged September 7, 2026, 22:14 UTC (September 8, 06:14 Asia/Shanghai); verified directly on September 8. This is an accepted contribution, not an assigned module-maintainer role. ↩ ↩2
-
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. ↩
-
XGrammar. Malformed function-tool validation #881, head
a79fd29, checked September 8, 2026. Draft publication is not upstream acceptance; local tests are not upstream CI. ↩ -
DeePMD-kit. Ragged charge/spin batching #6008. Open at the September 6 check; #6010's later merge is recorded separately below. ↩
-
MACE. Model-download validation #1712 is approved but open; gated-block TorchScript export #1717 is an open draft. Status checked September 8, 2026. Neither is counted as merged. ↩