Windows (WSL2)

The supported path on Windows hardware is WSL2. A WSL2 distribution is ordinary x86_64-unknown-linux-gnu, so it runs the Linux fdl binary, the Linux libtorch, full CUDA, and full NCCL multi-GPU training. Nothing is emulated and nothing is cut down.

This is not a fallback. floDl’s published benchmarks were produced on this exact configuration (see Benchmarks):

What runs where

fdl.exe exists as a libtorch fetcher for people consuming libtorch directly from Windows (C++ LibTorch, tch-rs, or a PyTorch install that needs the matching C++ archive). It downloads and unpacks the correct libtorch-win-shared-with-deps-* archive for a chosen CUDA version and stops there.

It cannot build or train, and this is not a gap waiting to be filled: flodl-sys has never been compiled against MSVC, and more decisively NCCL has no Windows port (neither does RCCL), so a native Windows build could never do multi-GPU DDP at all. WSL2 has both. Use WSL2.

Prerequisites

• Windows 11, or Windows 10 21H2+ (GPU passthrough needs /dev/dxg).
• The GPU driver installed on Windows, not inside the distribution. This is the one setup mistake worth calling out: installing a Linux GPU driver inside WSL overwrites the passthrough stubs and breaks CUDA. The Windows driver supplies nvidia-smi inside the distro through /usr/lib/wsl/lib/.
• Docker Desktop with the WSL2 backend, or Docker installed inside the distribution. Either works.

Verify the GPU is visible from inside WSL before going further:

nvidia-smi          # should list your GPU(s)
ls /dev/dxg         # the passthrough device

If nvidia-smi is missing, the Windows-side driver is too old or absent. Fix that before installing anything in the distro.

Setup

Inside the WSL distribution, follow the ordinary Linux instructions:

curl -sL https://flodl.dev/fdl -o fdl && chmod +x fdl
./fdl install
fdl init my-project

Then confirm what floDl sees. fdl probe labels the kernel explicitly, so a WSL2 host is unambiguous in bug reports:

OS:          Linux 6.6.87.2-microsoft-standard-WSL2 (WSL2)

Two settings that matter for floDl

Keep the project on the WSL filesystem. Work in ~/src/... inside the distribution, not /mnt/c/.... The Windows drive is reached over a translation layer, and a cargo target directory or a dataset served across it is dramatically slower than the native ext4 one. This dominates every other performance knob on WSL.

Give WSL enough RAM, and know what it reports. WSL2 runs in a lightweight VM whose memory defaults to a fraction of the host’s and grows on demand. floDl’s data plane budgets against MemAvailable, so it sizes its caches to the VM’s view, not the machine’s. If the loader seems to under-cache on a large-RAM box, raise the VM’s ceiling in C:\Users\<you>\.wslconfig:

[wsl2]
memory=48GB
processors=16

Restart with wsl --shutdown from PowerShell for it to take effect.

Multi-GPU

Multi-GPU DDP works under WSL2 with NCCL, which is why WSL is the supported path rather than a compromise. The usual commands apply:

fdl gpu-test          # GPU test suite
fdl gpu-test-nccl     # NCCL / DDP tests, isolated processes

See Multi-GPU Training for the full picture.

AMD GPUs

Use bare-metal Linux. AMD’s ROCm has a WSL path, but floDl’s AMD support detects GPUs through the kernel’s KFD topology (/dev/kfd plus /sys/class/kfd/), which is the bare-metal interface. We have not validated floDl on ROCm-under-WSL and make no claim about it.

Troubleshooting

nvidia-smi works on Windows but not inside WSL

The distribution predates GPU support or the driver is too old. Update the Windows-side driver, then wsl --shutdown and reopen. Do not install a Linux GPU driver inside the distribution to work around this: it replaces the passthrough libraries in /usr/lib/wsl/lib/ and makes the problem permanent.

Builds are far slower than on native Linux

The project is almost certainly under /mnt/c/. Move it into the distribution’s own filesystem (~/) and rebuild. df -h . from the project root tells you which you are on: the WSL filesystem shows as /dev/sd*, the Windows drive as drvfs.

The data loader caches less than expected

The WSL VM’s memory ceiling, not the host’s RAM, is what MemAvailable reports. Raise memory= in .wslconfig (above) and wsl --shutdown.

fdl.exe says a variant is unavailable

ROCm libtorch is published for Linux only, so fdl.exe refuses it rather than fetching a URL that does not exist. CPU and CUDA variants are available for Windows.