Zhou Taichang

Founding contributor to vLLM-Omni, the open-source multimodal (any-to-any) LLM serving framework: built the initial engine core — executor, worker, model runner — in the repo's first week; then implemented the AR (autoregressive) engine (AR GPU model runner, scheduling components, Omni model abstractions, and multimodal output processing).

Designed and merged engine-level KV-cache management for the Block Diffusion Engine / DreamZero world model (RFC #4534), validated bit-exact; authored comprehensive documentation in BDE_doc.

Single-handedly performed every recurring upstream vLLM rebase (through v0.24.0): dependency and API alignment across 8 engine modules, wheel/CI pinning, and post-rebase review follow-ups (PR #4830)

Core reviewer for multimodal engine PRs: audio talkers (CosyVoice3, MOSS-TTS, Higgs-Audio), image generation (HunyuanImage3), XPU enablement; shipped production fixes for concurrent PCM audio streaming (issue #4411).

Authored comprehensive ModelRunner correctness refactor (RFC #5450, PR #5452): nine independently-revertable commits fixing six defect classes across generation/scheduling/concurrency boundaries, each with its own regression test.

Led rebase to vLLM 0.29.0 (PR #7230), preserving multimodal and staged-generation behavior while adapting to upstream API.

Authored multi-part worker refactor series (G3/N PR #6988) fixing dead code, FA3 resize, silent failures, and API consistency across GPU model runners; shipped DreamZero warmup bugfix (PR #7548) and Qwen3-Omni talker seed fix (PR #7474)

Led upstream vLLM rebase through v0.30.0 (PR #7820, 761 upstream commits), preserving multimodal and staged-generation behavior while adapting to upstream API

Designed and merged engine-level KV-cache management for the Block Diffusion Engine / DreamZero world model (RFC/PR #4534, https://github.com/vllm-project/vllm-omni/pull/4534), validated bit-exact, with full architecture documentation in BDE_doc

Drove the world-model realtime-inference roadmap (RFC #7074, https://github.com/vllm-project/vllm-omni/issues/7074) from LingBot World 2.0 to MiniMax H3, ABot-World and Echo-WM, opening the AR7 VAE decode optimization line (issue #100)

Cut LingBot World streaming VAE decode cost with width-sharded streaming decode (PR #7651), fused fast paths for the streaming Wan decoder (PR #7749, mean cadence 590ms → 549ms) and one-pass FP8 input quantization with a Triton 96-channel VAE conv (PR #8111), all measured on a two-H200 kernel ledger

Contributed the LingBot-World realtime streaming benchmark (PR #7645) and eleven bit-identical removals of redundant work in the realtime path (PR #7648); shipped TaoMate-H3 (MiniMax-H3 rank-128 LoRA) real-time streaming on the AR-Diffusion runtime with two- and four-GPU H200 configs (PR #8268)

Built a self-hosted personal assistant agent delivering automated daily digests, profile consolidation, PR-driven website refresh (tzhouam.github.io) and OpenClaw bridge multi-channel communication, in continuous daily operation since July 2026

Hardened Mixture-of-Agents routing into stable multi-model inference with crash-safe config, aggregator fallback and per-call retry (PRs #1–2)

Shipped production fixes for WeChat image vision fallback and empty chat replies plus a one-click reboot CLI (PRs #3–5), and authored an RFC for backward-compatible multi-user tenant isolation (PR #6)

Built a LangGraph multi-agent system that automates vLLM rebases end-to-end, reducing manual effort from days to hours: parallel per-module LLM agents, plan-review gating, local pipeline testing, autonomous Buildkite CI debugging, and a self-improving skills/debug-memory knowledge layer

Implemented a read-only MCP server for Claude Code/Codex (PR #1) and dual-path model routing with eco-by-default intent parsing (PR #2), optimizing token spend during long-running orchestration tasks

Stabilized draft/review runs by retiring AI-generated adapters in favor of a pinned vendor git-submodule knowledge base with GPT-gated curation (PRs #3–4), ensemble reducer routing fix (PR #5), and code-mirror verification pinning (PR #6)

Packaged the orchestrator as the vllm-omni-copilot CLI with locked playbooks for reproducible rebase and review runs across engine modules

Authored the strict-review design (RFC #72: deep passes, per-comment verification, self-proving evidence) and shipped it in PR #71, winning the first head-to-head review battle 8–7 against a strong baseline at roughly one-third the cost per a blind gpt-5.6 judge (https://github.com/JiusiServe/InferMatrixCopilot/pull/71)

Built the provider registry that powers multi-backend review runs: cursor-agent backend (PR #81) and claude-code/codex backends (PR #82), specified in RFC #80 subscription-auth harness (https://github.com/JiusiServe/InferMatrixCopilot/issues/80)

Landed unattended-review foundations: an RFC for a GitHub-event trigger layer starting with automatic PR review (issue #3), plus hardening PRs #64–#67 (quick-map status, tool-output bounds, CI pytest fix) and an eval campaign record with wave-2 holdout (PR #62)

Drove cross-repo review-evidence workflows, e.g. adding a failing test node as a CI-ownership signal in zuiho-kai/omni-reviewbot (https://github.com/zuiho-kai/omni-reviewbot/pull/9) for InferMatrixCopilot issue #87

Built PR automation features including verified CI readiness, maintainer notifications, stale reminders, and comment-only review authority.

Built the maintainer-side channels for the InferMatrixCopilot improvement engine: a proposal routine (PR #8) that applies engine outbox actions with its own GitHub token, and artifact publication (PR #9) that verifies source, patch, tests, evaluation and protected paths before pushing a fixed candidate branch and opening a human-reviewed draft PR

Built the M0/M1 foundation of a local Codex GUI: vendored 743 stable TypeScript protocol bindings from codex-cli 0.162.0-alpha.17.2 with SHA-256 manifest and reproducible generate/check (PR #1), then shipped the M1 local Codex App Server with HTTP/WebSocket service, Host/Origin/cookie validation, connection-status display, and 85/85 passing tests (PR #2)