LTX-2.5 on Windows or Mac: natten, attention and fallbacks

LTX-2.5 can be installed on Windows or macOS, but its fastest paths are for Linux with CUDA. The LTX-2 README says the natten extra, the fastest backend for the diffusion video VAE, is Linux and CUDA only, and that on Windows and macOS it is skipped automatically and decoding falls back to a Triton or eager implementation. The model card still lists Python 3.12 or higher and CUDA 12.7 or higher, so a Mac without CUDA is outside the card's stated requirements.

Which backend does each part get?

What does a fallback cost?

The README does not give speeds for the fallbacks, so I cannot say how much slower they are. Expect to measure your own. An installed attention backend is selected automatically; forcing one is a Python-API option, not a CLI flag. The memory flags, --quantization fp8-cast and --offload cpu or disk, trade speed for memory.

Is WSL the way out?

The README does not say, so this is a suggestion rather than a vendor claim: Windows users with an NVIDIA GPU often run Linux tooling inside WSL, which gives the Linux and CUDA path the natten extra expects. Check that the CUDA version inside it meets the card's 12.7 floor. On a Mac there is no CUDA, so you are outside the card's stated requirements either way.

Before you spend a day on setup, run the smallest possible job end to end, and time it. If the fallback is too slow for your use, a hosted job is the shorter path.

When is the hosted route the simpler call?

If the machine is a Mac or a Windows laptop and the goal is a clip rather than a pipeline, send the job to a hosted id and download the MP4. Sume does not list LTX (catalog code, read 2026-10-09), so pick from GET /v1/videos/models. A request is plain JSON with model, prompt, duration, resolution, so it runs the same from any OS. See the video docs.