v0.4.0 — Linux CUDA build + GPU banner fix

Linux gets a CUDA build, and a GPU banner bug that sent Linux users to a Windows-only download is fixed.

Linux + NVIDIA — CUDA build

Linux gets the same GPU path Windows got in v0.3.0. Same app, compiled against CUDA instead of CPU-only PyTorch. A convert that takes ~40 minutes on a CPU takes a small fraction of that.

It runs on exactly the same distros as the standard Linux build. The bundled CUDA runtime needs nothing newer than glibc 2.35, so if the regular build starts on your machine, this one will too — Ubuntu 22.04+, Debian 12+, Mint 21+, Pop!_OS 22.04+, Fedora 36+, Arch, openSUSE 15.5+.

All it needs from your system is the NVIDIA proprietary driver — if nvidia-smi prints your card, you're set. The CUDA toolkit is inside the bundle; there is nothing to apt install.

The GPU banner was broken on Linux

The launch check asked "is there an NVIDIA card here?" and "can this build use it?" — but never asked what OS it was running on. So a Linux user with an NVIDIA card got a Get the CUDA build link pointing at something that only existed for Windows, which dropped them on a release page whose only Linux file was the CPU build they already had.

Fixed. The banner is platform-aware, and it now links to the repo's GPU acceleration section — which actually explains the multi-part download — instead of the bare releases page. macOS is excluded permanently: there are no CUDA PyTorch wheels for macOS at any version, so an NVIDIA card in an Intel Mac has no upgrade path and shouldn't be advertised one.

Thanks to the person who reported it on reddit.

Also in this release

• The app shows its version in the header. It didn't before, which made "which build are you on?" unanswerable.
• Corrected the CPU runtime claim. The README said a Standard convert takes "a few minutes" on CPU. That's true on a modern desktop chip and wrong on an older one — a 2012 i7-3770 took ~40 minutes at Draft. The ETA also reads optimistic at the start: it extrapolates from finished epochs, and early epochs run faster than later ones, so the first number drifts up.
• Apple silicon is unchanged — still trains on the GPU via Metal (MPS) automatically, nothing to download.

Downloads

Platform	File
macOS (signed + notarized)	nam-a2a1-converter-macos.dmg
Windows	nam-a2a1-converter-windows.zip
Windows + NVIDIA	nam-a2a1-converter-windows-cuda.zip.001 and .002 — see below
Linux x86_64	nam-a2a1-converter-linux-x86_64.tar.gz
Linux x86_64 + NVIDIA	nam-a2a1-converter-linux-x86_64-cuda.tar.gz.001, .002 and .003 — see below

Getting the CUDA build

GitHub caps a single release file at 2 GiB, so both CUDA bundles ship as numbered parts you rejoin. They're a plain byte split, so the tools to rejoin them are already on your machine.

Linux (~3.5 GiB, three parts):

cat nam-a2a1-converter-linux-x86_64-cuda.tar.gz.0* > nam-a2a1-converter-linux-x86_64-cuda.tar.gz
sha256sum -c --ignore-missing nam-a2a1-converter-linux-x86_64-cuda.tar.gz.sha256
tar -xzf nam-a2a1-converter-linux-x86_64-cuda.tar.gz
./nam-a2a1-converter/nam-a2a1-converter

The glob orders the parts correctly — they're numbered .001, .002, .003, so plain lexical sort is the right order.

Windows (~2.6 GiB, two parts) — easiest, download get-cuda-build.ps1 and run:

powershell -ExecutionPolicy Bypass -File get-cuda-build.ps1

It fetches both parts, verifies them against the published SHA256, rejoins them and unpacks the app. (-ExecutionPolicy Bypass is needed because the script is unsigned; it only ever downloads from this repo's releases.) Or by hand:

copy /b nam-a2a1-converter-windows-cuda.zip.001 + nam-a2a1-converter-windows-cuda.zip.002 nam-a2a1-converter-windows-cuda.zip

Each .sha256 lists the expected hash for the assembled archive and for every part — check against it if a convert later fails oddly, since a truncated part is the usual cause.

Requires an NVIDIA card of compute capability 5.0+ (GTX 900-series / 2014 or newer) with a current driver.

Prefer not to download several GiB? CUDA also works from source on Linux and Windows:

python3 -m venv .venv
./.venv/bin/pip install torch --index-url https://download.pytorch.org/whl/cu126
./.venv/bin/pip install -r requirements.txt
./.venv/bin/python -m nam_a2a1

Notes

• Signing is unchanged: macOS is signed + notarized and opens clean, Windows is unsigned so SmartScreen warns — More info → Run anyway.
• macOS, Windows and Linux have all now had real conversions run on them by users — including Linux Mint 22 on an i7-3770, with the output imported and running on a Valeton GP-150. The GPU builds are newer and less travelled, so if you're on CUDA and something looks wrong, please open an issue.