I build agent tools and contribute to open-source inference systems, from multimodal models to serving and runtime correctness.
我关注 Agent 工程 && 模型推理优化，期待成为一名AI infra-er，欢迎交流呀。
Selected merged upstream pull requests.
| Project | Contribution | PR |
|---|---|---|
| FastVideo | Support Comfy NVFP4-AWQ text encoders for MiniMax H3 | #1865 |
| SGLang | Recover HiCache host slots after allocation failures | #46606 |
| FlashInfer | Reject unsupported GPUs before FP4 quantization | #7987 |
Open PRs · proposals and work in progress
Selected open, non-draft proposals. Titles and status sync from GitHub daily.
| Project | Proposed change | PR |
|---|---|---|
| vLLM-Omni | Fuse H3 final affine and FP32 promotion | #8677 |
| vLLM-Omni | Load Alibaba PDD adapters for MiniMax H3 | #7425 |
| FlashInfer | feat(quantization): preflight MXFP8 inputs before JIT | #8793 |
| AIPerf | feat(dataset): add opt-in disjoint warmup conversation pools | #1676 |
| SGLang | [NVIDIA] kimi_k3: Add in-tree TP8 specialized decode with DCP and HiCache | #48510 |
| SGLang | deepseek_v41: track image work after cancellation | #47457 |
| FlashInfer | ci: report failed and cancelled matrix lanes in test summary | #7996 |
| vLLM | [Benchmark] Reject incomplete completion streams | #64703 |
Automatic PR sync · Scheduled daily for 08:17 Beijing time; GitHub may start it later.
| Project | What it does |
|---|---|
| repository-memory | Source-backed memory for coding agents, with Git citations, a local store, CLI, and MCP server. |
| dsh-ops-kit | Evidence-driven memory, orchestration, benchmarks, and release workflows for DeepSeek Harness. |
| dsh-session-search-pro | Search past and live DeepSeek Harness sessions through its session service. |
| review-workflow | Structured review with independent panelists, disagreement resolution, and a separate critic. |
Also exploring coding-agent evaluation and verifiable distributed inference.
Based in Beijing · WeChat: yaole516