Skip to content
View algebra-MCX's full-sized avatar
17 followers · 76 following

Achievements

Achievement: Pull SharkAchievement: StarstruckAchievement: Quickdraw

Achievements

Achievement: Pull SharkAchievement: StarstruckAchievement: Quickdraw

Block or report algebra-MCX

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. Markdown is supported. 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
algebra-MCX/README.md
Chenxi Ma

Typing animation


vLLM contributor Homepage Zhihu X

What I work on

  • MoE serving: expert parallelism, dispatch / combine, communication–compute overlap
  • vLLM: MoE expert parallelism, dual-batch overlap (DBO) and NIXL dispatch in Model Runner V2
  • GPU kernels: Triton fused kernels, CUDA Graph decode, tensor-core attention
  • Engines: I read production inference engines closely, rebuild their core modules, and measure every change

vLLM

I'm a contributor to vLLM. My work focuses on MoE expert parallelism in Model Runner V2 (MRV2): correctness fixes under dual-batch overlap (DBO), and the performance of NIXL-based dispatch. Two of my DBO correctness fixes have been merged into main.

PR What it does Status
#63476 [Bugfix][MRV2] Report the first failed DBO microbatch Merged
#63372 [Bugfix][MRV2] Keep routed-expert rows of every DBO microbatch Merged
#63485 [Perf][MoE] Size NIXL dispatch from synchronized execution slices In review

Projects

DeepInfer-Lab

A minimal, readable re-implementation of DeepInfer's core inference engine design.

DeepInfer-Lab stars DeepInfer-Lab forks

MoEFlow

MoE expert parallelism: fused Triton kernels, NCCL all-to-all and CUDA Graph decode on B200.

MoEFlow stars MoEFlow forks

  • DeepInfer-Lab: runs the real DeepSeek-V4.1-Flash on 4 × B200. Every optimization is written up in an issue with its baseline, profile, correctness tests and measured result. Roadmap.
  • MoEFlow: fused Triton kernels, 2× NCCL all-to-all and CUDA Graph decode for MoE expert parallelism on B200.

Tech stack

Tech stack

GitHub stats

GitHub stats: commits, PRs, issues, stars Contribution streak

Contribution graph

Contribution snake

Writing

My blog is now live! It brings together all my ideas, with posts about interesting open-source projects I discover and projects I build myself. Stay tuned.

Pinned Loading

  1. vllm-project/vllm vllm-project/vllm Public

    A high-throughput and memory-efficient inference and serving engine for LLMs

    Python 95.9k 24.2k

  2. DeepInfer-Lab DeepInfer-Lab Public

    An open lab for DeepInfer systems research, correctness validation, and end-to-end optimization

    Python 26 5

  3. MoEFlow MoEFlow Public

    MoE expert parallelism for vLLM: fused Triton kernels, 2x NCCL all-to-all, CUDA Graph decode on B200

    Python 1