AMD FastFlowLM FastFlowLM
Models Benchmarks
Docs
GitHub Discord YouTube Email

Building FastFlowLM from Source on Windows

This page is for building flm.exe yourself, for example to work on FastFlowLM. If you only want to run models, install the packaged release instead; see Get Started (Windows).

The build is the same one CI uses for the Windows release. The full manual walkthrough, including every dependency and the alternative dependency routes, is in src/WinSetup.md in the repository.

⚙️ Requirements

  • 🖥️ OS: Windows 11 x64
  • 🌐 Network: the first build downloads codec sources and Rust crates
  • ⚡ To run the result: an AMD Ryzen™ AI NPU with the NPU driver installed (see Get Started (Windows)). Building does not need the NPU or its driver.

🚀 Quick Start

src\WinSetup.ps1 sets up the build machine and builds flm.exe in one go. Run it from a regular 64-bit PowerShell window. Do not use “Developer PowerShell for VS”: that shortcut targets x86, and the script sets up its own x64 build environment.

You need Git to clone the repository. On a fresh machine, install it with winget install Git.Git and open a new PowerShell window. The script installs everything else.

git clone https://github.com/ROCm/FastFlowLM.git C:\dev\FastFlowLM
cd C:\dev\FastFlowLM
powershell -ExecutionPolicy Bypass -File src\WinSetup.ps1 -InstallPrereqs

-InstallPrereqs uses winget to install any tools that are missing. Visual Studio Build Tools installs machine-wide and shows a UAC prompt, which an administrator has to approve. If you leave the switch off, the script lists what is missing and stops.

When the script finishes, src\out holds a runnable build:

cd src\out
.\flm.exe validate
.\flm.exe run llama3.2:1b

Rerunning the script is safe. Steps whose output already exists are skipped, so a rerun only rebuilds what changed.

Notes:

  • Keep the clone path short, for example C:\dev\FastFlowLM. The Rust and codec build trees nest deeply, and a long path can exceed the Windows 260-character path limit.
  • The dependencies go under C:\dev, because the CMake files look for them there.

🧰 What the Script Does

  1. Checks the tools:
    • Git
    • Visual Studio 2026 or 2022 with Desktop development with C++ (the free Build Tools edition is enough)
    • CMake and Ninja (the copies bundled with Visual Studio work)
    • Rust through rustup, with the stable-x86_64-pc-windows-msvc toolchain
  2. Checks out the submodules (third_party/tokenizers-cpp, third_party/xgrammar).
  3. Sets up the native dependencies under C:\dev:
    • Boost 1.88.0 with Boost.ProgramOptions built
    • the XRT headers at a pinned commit
    • xrt_coreutil.lib, made from a .def file in the repository

    It runs the same script CI uses, .github\script\ci\setup_windows_deps.ps1. The curl headers are part of the repository, so you don’t need vcpkg.

  4. Configures and builds. It runs cmake --preset windows-ninja in an x64 Visual Studio developer shell.
  5. Stages the runtime files in src\out: flm.exe, the engine DLLs, model_list.json, model_info.json and xclbins.

🔧 Script Options

Option Effect
-Variant XRT The default build, the same as the release package. Includes the hip-ep (GPU) backend; pass -NoOgaHipep to leave it out. Builds into src\build.
-Variant HRX Builds against the pinned HRX runtime release. Builds into src\build-hrx.
-Variant Source Builds the Gemma 4 engine (gemma4e_npu.dll) from source instead of using the prebuilt DLL. Builds into src\build-models-from-source.
-BuildDir <dir> Uses a different build directory under src.
-Clean Deletes the build directory and src\out first.
-Jobs <n> Sets the number of parallel build jobs. The default is all cores, or 2 for Source.
-SkipBuild Sets up the tools and dependencies only.
-NoStage Skips copying the runtime files into src\out.

All variants stage into the same src\out, so it holds whichever variant was built last.

🛠️ Building by Hand

Once the dependencies are in place, you can rebuild without the script. Open an x64 developer environment, for example the “x64 Native Tools Command Prompt for VS”, then run:

cd src
cmake --preset windows-ninja -DFLM_ENABLE_OGA_HIPEP=ON
cmake --build build --parallel

You can also use the Visual Studio generator, which does not need a developer shell:

cd src
cmake --preset windows-vs18 -DFLM_ENABLE_OGA_HIPEP=ON   # VS 2026; use windows-default for VS 2022
cmake --build build --config Release --parallel

For the manual dependency setup, the HRX and Source variants, staging, and the vcpkg and conda dependency routes, see src/WinSetup.md.

🩺 Troubleshooting

  • LNK1104: cannot open file 'kernel32.lib'. The developer shell could not find the Windows SDK. This happens on machines where policy blocks reg.exe. WinSetup.ps1 detects this and adds the SDK paths itself. For a manual Ninja build, use the Visual Studio generator instead.
  • “This is a 32-bit PowerShell” or “This developer shell targets x86”. Run the script from a regular 64-bit PowerShell window instead.
  • A build directory configured with another generator. A build directory stays tied to the generator that first configured it. Rerun with -Clean, or pick another directory with -BuildDir.