Working on reliable agent systems, LLM post-training infrastructure, and Model Context Protocol (MCP) tooling with Python.
我关注 AI Agent、LLM 后训练与 MCP 生态，并通过开源协作把工程问题变成可验证的实验。
- FastMCP · merged PR #4704: CodeMode error propagation — Fixed tool-error propagation so sandboxed CodeMode programs can catch failed tool calls while uncaught failures still surface correctly.
- AReaL · open draft PR #1571: TMS teardown — Proposed race-safe teardown after TMS offload, covering idempotent FSDP cleanup, ordered worker shutdown, and verified process-group termination.
- TRL · open draft PR #6615: activation offloading — Proposed opt-in activation offloading for GRPO and RLOO trainers, with step-scoped lifecycle management, tests, and documentation.
- OPEN · inclusionAI/AReno#473 — test: add native attention GPU equivalence checks
- OPEN · inclusionAI/AReno#472 — docs(agentic): define trajectory observability contract
- OPEN · inclusionAI/AReno#471 — build: adopt file-backed dynamic package version
- OPEN · alibaba/open-code-review#848 — fix(viewer): show running status for active sessions
- OPEN · areal-project/AReaL#1578 — fix(infra): preserve LD_PRELOAD in local launchers
- How should agent runtimes recover from tool, state, and process failures?
- How can offloading make RL post-training more memory-efficient without obscuring failure modes?
- How should MCP tools preserve clear contracts and error semantics across execution boundaries?
Python · PyTorch · FSDP · RL post-training · distributed systems · MCP