Allen Kuo (kwyshell)·Nov 1Another Route to vLLM-Style Serving, Part 2: CPU, Vulkan, and GGUF Across PlatformsHandwritten and Triton-CPU kernels, Vulkan tiling and data reuse, GGUF format costs, and the gaps that remain
Allen Kuo (kwyshell)·Nov 1vLLM 式 serving 的另一條路(下):CPU、Vulkan 與 GGUF,走向多平台手寫與 Triton-CPU 的混合 kernel、Vulkan 的分塊與資料重用、GGUF 的格式代價,以及還沒追上的地方
Allen Kuo (kwyshell)·Oct 22Another Route to vLLM-Style Serving, Part 1: Running vllm.cpp Natively on WindowsA first full introduction, using v0.0.4-vk.2’s official FP8 MoE support, hybrid prefix caching, and translation-shaped measurements
Allen Kuo (kwyshell)·Oct 22vLLM 式 serving 的另一條路(上):讓 vllm.cpp 原生跑在 Windows第一次完整介紹這個 fork,以 v0.0.4-vk.2 的官方 FP8 MoE、hybrid prefix cache 與翻譯負載量測為起點
Allen Kuo (kwyshell)·Oct 8Splice 1.2.0 and GTG 1.1.0: bringing two open-source projects togetherThe decisions behind its protection worker, bitmap overlay, and Present hooks
Allen Kuo (kwyshell)·Oct 8Splice 1.2.0 在 GTG 1.1.0 的全新升級一個小型原生 Windows 工具,兩個開源專案,以及讓 overlay 不成為遊戲負擔的設計
Allen Kuo (kwyshell)·Sep 3AI 推理工作負載不斷重設我的 GPU,於是我為 RTX PRO 6000 Blackwell 打造了一套熱保護工具一套小型原生 C++ (1.2MB) 防護工具,守住「GPU 正變得危險地熱」與「Windows 已經重設驅動程式」之間的空檔
Allen Kuo (kwyshell)·Sep 3My RTX PRO 6000 Blackwell kept resetting during AI workloads, so I made a thermal guardA tiny native C++ (1.2MB) safeguard for the gap between “the GPU is getting dangerously hot” and “Windows has already reset the driver”
Allen Kuo (kwyshell)·Aug 5邊緣推論引擎:一套原生 API,打通 CPU、GPU、NPU 與 RISC-V十個 execution provider,加上一條讓影格從解碼到編碼全程不進 CPU 記憶體的路 — — 跨 NVIDIA、Android、Vulkan、DirectX 的 buffer 身分轉換,每一格都過了正確性、延遲、記憶體和能耗的關
Allen Kuo (kwyshell)·Aug 5Edge Inference Engine: One Native API Across CPU, GPU, NPU, and RISC-VTen execution providers, plus one path that keeps a frame out of CPU memory from decode all the way to encode — buffer-identity conversion…