Max Wei CJ

The definitive vLLM runtime for dual RTX 2080 Ti 22GB + NVLink, delivering Qwen 27B local inference with 100+ tok/s single-request decode with support of FP8 weight

Single-request LLM serving recipes, patches, and benchmarks for modified RTX 2080 Ti 22GB / SM75 systems

high-performance linear attention kernel library built on TileLang

make RDNA1 / Navi1x / gfx101x GPUs (5700XT, W5500 etc.) run modern large language models on ROCm 6 and ROCm 7 through llama.cpp, compatible with modern LLM like Gemma 4 and Qwen3.5, even on a host platform that is not a clean vendor-supported reference system.

AI Proxy Hub is a cross-platform local gateway for AI clients and upstream APIs. It unifies multiple upstream endpoints behind one local control plane, adds protocol-aware routing and failover, and provides both a Web dashboard and an interactive CLI.

A benchmark for LLM variant quant models at agentic coding situation

This is a DIY project to push the Dirtywave M8 Headless into a PCB and make it handheld.

No description.

LLM inference in C/C++

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

2080ti / SM75 / Turin aarch compatible Lucebox optimization hub: hand-tuned LLM inference, built for specific consumer hardware.

Homebrew tap for AI Proxy Hub

No description.

M8 Headless Precompiled Firmware

Automatically exported from code.google.com/p/vim

Call all Node.js modules directly from DOM and enable a new way of writing applications with all Web technologies.

What doesn't kill you --- makes you stronger.

No description.

a blog built on hexo