Skip to content
View Li-changwu's full-sized avatar
  • Huazhong University of Science and Technology (CGCL / IntelliStream)
  • Wuhan, China

Highlights

  • Pro

Block or report Li-changwu

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.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. 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
Li-changwu/README.md

Changwu Li

华中科技大学(HUST)博士研究生 · CGCL / IntelliStream Lab · vLLM-HUST

专注于 大模型推理系统(LLM Serving Systems) 与 ACL Graph 图兼容机制(Graph-Compatible Execution) 研究,围绕静态计算图捕获、回放约束下的动态状态管理与设备控制边界开展系统构建:

  • ACL Graph 图兼容专家缓存与卸载:面向动态路由与受限显存场景,通过预分配地址稳定固定槽位、持久设备映射、图外异步 Staging 与有界溢出分波,解决权重动态换入换出破坏静态图捕获与回放的问题。
  • 单步图回放之上的设备驻留控制面:在计算图消除单次执行算子下发开销的基础上,利用昇腾 DataFlow Device UDF 将跨步解码循环、Paged-KV 更新与采样下沉至设备侧按有界 Epoch 连续执行,消除逐 Token Host 协调往返。
  • 推理图执行综述与数据流系统:负责维护社区 ACL Graph 技术专题综述 与 vLLM / vLLM-Ascend 性能演进研报工作台,并开展非平稳数据流在线指纹选择研究(WWW '26)。

代表推理系统

项目 方向与平台 核心机制与关键结果
LatchMoE (NPU)
LatchMoE-GPU
图兼容 MoE 专家缓存卸载
ACL Graph · vLLM · CUDA Graph
面向受限显存的地址稳定、图回放兼容 MoE 专家缓存框架(论文工作,昇腾 NPU 主实现 + CUDA PIECEWISE 图对照实现);通过预分配固定槽位、持久设备映射、图外异步 Staging、有界溢出分波(Bounded Waves)与槽位租约保护,保持捕获计算图权重地址稳定并消除远端按需访存瓶颈。
Cruise 设备驻留解码控制面
Ascend 910B2 · CANN · vLLM V1
探索单次图回放之上的跨步控制边界:基于 CANN DataFlow Device UDF 将解码内层控制循环、Paged-KV 状态更新与采样下沉至昇腾设备侧按有界 Epoch 连续执行;单 Epoch 控制边界载荷从 136.9 MB 降至 628 B,Host 控制耗时降低 3.54×。

学术论文

  • [WWW '26] Changwu Li, Tongjun Shi, Shuhao Zhang, Binbin Chen, Bingsheng He, Xiaofei Liao, Hai Jin.
    StreamFP: Fingerprint-guided Data Selection for Efficient Stream Learning.
    In Proceedings of the ACM Web Conference 2026. [Code] · [DOI]

开源社区建设


技术栈

  • 图执行与推理运行时:vLLM (V1 Engine / vLLM-Ascend) · Ascend CANN (ACL Graph / DataFlow Device UDF) · PyTorch · CUDA Graph · Triton
  • 系统开发与性能剖析:Python · C / C++ · Ascend Profiling / PyTorch Profiler · Linux

Popular repositories Loading

  1. FlowRAG FlowRAG Public

    Forked from CGCL-codes/FlowRAG

    FlowRAG: Continual Learning for Retrieval-Augmented Generation

    Python

  2. vllm-ascend-hust vllm-ascend-hust Public

    Forked from intellistream/vllm-ascend-hust-legacy-20260831

    IntelliStream fork of vllm-ascend for vllm-hust integration and Ascend runtime convergence

    Python

  3. vllm-hust vllm-hust Public

    Forked from intellistream/vllm-hust-legacy-20260831

    Upstream-compatible vLLM fork for domestic hardware enablement and AGI4S serving

    Python

  4. LatchMoE-GPU LatchMoE-GPU Public

    LatchMoE (GPU):面向受限显存的地址稳定、图重放兼容 MoE 专家缓存卸载框架(CUDA + vLLM PIECEWISE Graph 实现,与昇腾 NPU 主仓对应)

    Python

  5. Cruise Cruise Public

    Cruise:面向昇腾 NPU 的设备驻留解码(Device-Resident Decode Epochs)系统原型,基于 CANN DataFlow 消除逐 Token Host 往返

    Python

  6. codex-provider-switcher codex-provider-switcher Public

    VS Code 扩展:安全管理与切换多套本地及 Remote-SSH Codex Provider 配置,支持事务级回滚保护与原生会话续接

    TypeScript