ByteBotpack

Byte, Kilobyte and Megabyte for RLBot v5, with an optional rank selector. Supports 1v1, 2v2 and 3v3.

Play

Add the bot.toml at the root of the built package to RLBot. Keep the complete bot folder together. For a native build deployed into this checkout, use rlbot/bot.toml. The default is full-strength Megabyte with the selector enabled. Choose Byte, Kilobyte or Megabyte at full strength, or Select Skill for estimated Bronze I through SSL difficulty. Auto tries to keep the score even. Changes apply with Apply & respawn.

CPU inference is the default on Windows and Linux. It uses the CPU runtime already installed with the official RLBot botpack. On Windows this is normally %LOCALAPPDATA%/RLBot5/bots/torch-archive/torch; set LIBTORCH_CPU_ROOT to that torch directory if it is installed elsewhere. Byte does not download another runtime when it starts. The selector window is available on Windows.

GPU inference requires a separate Windows build and a compatible CUDA LibTorch installation; see the GPU build instructions below.

Run without the selector

By default, the selector window opens when the bot starts. To disable it, append --manual=Byte, --manual=Kilobyte, or --manual=Megabyte to the existing run_command in the package-root bot.toml. Preserve the executable path that Bob generated. For example, a Windows Bob package can use:

run_command = 'x86_64-pc-windows-msvc\GGLBot.exe --manual=Megabyte'

Each named model runs at full strength without the selector. Replace the argument with --manual=Off, or remove it, to enable the selector again. For Linux CPU packages, append the same argument to the existing run_command_linux.

Rank labels are provisional estimates, not verified human ranks.

Build for the botpack

From a source checkout with Bob and Docker (Linux containers) installed, run:

bob build bob.toml --out-dir "path/to/build-output"

The package is written to path/to/build-output/ByteBotpack/. Without --out-dir, Bob uses bob_build/ByteBotpack/ in this repository.

The checked-in rlbot/device.txt selects cpu. A fresh CPU build automatically downloads the Windows and Linux SDKs from the public RLBot CPU PyTorch archive. Both downloads are pinned to that release and checked against SHA-256 hashes. No local LibTorch SDK or Python installation is needed for a plain Bob CPU build. These are build-time downloads; the finished bot uses the shared runtime already installed by RLBot.

The optional Python 3.10+ helper can create an isolated source snapshot:

python scripts/build_bob.py --bob "path/to/bob.exe" --work-dir "path/to/scratch" --out-dir "path/to/build-output"

Its default CPU build also uses the public SDKs. To supply a different complete Windows CPU SDK, pass --libtorch "path/to/torch" or set LIBTORCH_CPU_ROOT. Use a version matching the runtime on the computer that will run the bot. An explicit invalid SDK produces an error instead of downloading a replacement.

Optional GPU build (Windows)

1. In a source checkout, change rlbot/device.txt from cpu to gpu.
2. Supply a complete release Windows CUDA LibTorch SDK and build with the helper:

python scripts/build_bob.py --bob "path/to/bob.exe" --libtorch "path/to/libtorch" --work-dir "path/to/scratch" --out-dir "path/to/gpu-build-output"

The helper stages only the SDK headers and import libraries, fingerprints them for Bob's cache, and removes its generated SDK copy after the build. The source snapshot remains in the requested scratch directory. Without --work-dir, SDK staging happens temporarily in build-support/libtorch/local/ in the checkout.

To run the GPU package, use a compatible NVIDIA GPU and driver, and the same CUDA LibTorch version used for the build. Set LIBTORCH_CUDA_ROOT to its libtorch directory, or install it at %LOCALAPPDATA%/RLBot5/bots/libtorch. The helper also checks that location when --libtorch is omitted. CUDA runtime DLLs are supplied by the user and are not included in the package.

Device selection happens before compilation. Changing device.txt in an already built package does not switch CPU/GPU; rebuild from source. A GPU build does not fall back to CPU, and plain Bob GPU builds require a staged CUDA SDK. Keep cpu in the source configuration submitted to the botpack.

The six inference files are grouped under rlbot/models/byte, kilobyte and megabyte in the source checkout, or models/ in a built package. Their checksums are in the model manifest.