Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
mgua: added a revised section for cuda enabled builds on laptop, minimal installs, mingw64 gcc and make and nvcc. no edits on *precious* previous text.
marco guardiglifbdef20883cb
diff --git a/docs/windows.md b/docs/windows.mdindex 81dd4ef1..e0939627 100644--- a/docs/windows.md+++ b/docs/windows.md@@ -2,6 +2,63 @@A start-to-finish, reproducible path from a fresh Windows 11 machine to GLM-5.2 generating tokens, with the GPU tier. Every step and every failure mode below was hit and verified on real hardware: Core Ultra 9 285K (AVX-VNNI) / RTX 5080 (sm_120) / 128 GB RAM / Windows 11 24H2 (issue #306). Steps are ordered so the long downloads run while you build.+---+> **2026-08-09: Additional validation with detailed steps for laptop setup** \+Lenovo Thinkpad P16v (Intel Core i7 ultra (155H 1.4GHz), 64GB RAM, 2Tb NVME drive, Nvidia RTX 2000 Ada generation 8GB (AD107, 2023)).+Windows 11 pro english.\+**Software installs:**\+`- nvidia drivers` from https://www.nvidia.com/en-us/drivers/\+`- nvidia cuda toolkit` from https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_local\+`- msys2` from https://github.com/msys2/msys2-installer/releases/download/2026-06-11/msys2-x86_64-20260611.exe\+`- Microsoft Visual Studio 2022 built tools` installer from https://aka.ms/vs/17/release/vs_buildtools.exe then install Desktop Development with C++\+`- winget install git.git python.python.3.14`\+\+`msys2` is installed in `c:\msys64` with its main launchers. (we will use `C:\msys64\mingw64.exe`)\+Its executables tools `sh`, `bash`, `make`, `sed` etc go in `c:\msys64\usr\bin`. \+From within `mingw64` we need to install compiler and make. \+\+`C:\msys64\mingw64.exe`\+`pacman -S --needed mingw-w64-x86_64-gcc make`\+\+**colibri.exe and coli_cuda.*** build from MS Visual Studio and Nvidia CUDA Toolkit:\+Microsoft Visual Studio tools includes a `vcvars64.bat` batch file that appropriately sets all the paths. \+It is in `"C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat"`\+\+CUDA compilation is to be performed from a CMD shell, where we run `vcvars64` and then the path \+extension for mingw64 build tools, as required by the make process (mingw64 `make` is used). \+Open a command prompt shell (`cmd`, not powershell), cd to subfolder `c\` of the cloned repo and run:\+\+`%comspec% /k "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat"`+`set PATH=%PATH%;C:\msys64\usr\bin`\+`cd C:\Users\YOUR_USER\COLIBRI_REPO_FOLDER\c`\+`make cuda-dll CUDA-ARCH=sm_89`\+`make colibri.exe CUDA_DLL=1 ARCH=native`\+`make iobench.exe`\+\+Replace `sm_89` with your Nvidia GPU architecture according to this table++| Architecture | Example GPUs / Products | Compute Capability | `nvcc` Flag (`-arch=sm_XX`) |+| :--- | :--- | :--- | :--- |+| **Blackwell** | B100, B200, RTX 50x0 | 10.0, 12.0 | `sm_100`, `sm_120` |+| **Hopper** | H100, H200, GH200 | 9.0 | `sm_90` / `sm_90a` |+| **Ada Lovelace** | RTX 40x0, RTX 2000, L4, L40 | 8.9 | `sm_89` |+| **Ampere** | A100, RTX 30x0, A10, Orin | 8.0, 8.6, 8.7 | `sm_80`, `sm_86`, `sm_87` |+| **Turing** | RTX 2080, GTX 1660 Ti, T4 | 7.5 | `sm_75` |+| **Volta** | V100, Titan V, Xavier | 7.0, 7.2 | `sm_70`, `sm_72` |+| **Pascal** | P100, GTX 1080 Ti, P40 | 6.0, 6.1, 6.2 | `sm_60`, `sm_61`, `sm_62` |+| **Maxwell** | M60, GTX 980, GTX TITAN X | 5.0, 5.2, 5.3 | `sm_50`, `sm_52`, `sm_53` |++> The last make links `colibri.exe` with the newly built `coli_cuda.dll`\+to just build `colibri.exe` with no nvidia CUDA support omit `CUDA_DLL=1`\+`ARCH=native` ensures that colibri is build with optimizations specific for your CPU.\+\+**caveat**: to be sure the correct `make` is used, run `where.exe make` after path extension.\+after compilation you should have `colibri.exe coli_cuda.lib, coli_cuda.exp, coli_cuda.dll` files.+---++++## 0. What you need| Piece | Why | Get it |