Support matrix

What mojolearn 0.8.41 supports: platforms, Python versions, devices and numeric modes. Release notes are in Release notes. mojolearn doctor reports what the installed wheel supports on your machine.

Platforms and wheels

Platform	Packages	GPU backend	GPU architectures	CPU
macOS arm64	mojolearn	Apple Metal	Apple silicon, built at the M1 instruction-set floor	Apple silicon
Linux x86-64	mojolearn plus mojolearn-nvidia and mojolearn-amd, installed automatically	NVIDIA CUDA, AMD HIP	CUDA sm_89; HIP gfx942	x86-64

Both platforms carry the CPU bindings, so both train and predict on a machine with no GPU.

NVIDIA needs only the driver (libcuda); no CUDA toolkit is required. AMD needs the amdgpu driver and the ROCm HIP runtime (ROCm 6 or 7).

In 0.8.41 the NVIDIA package ships sm_89 code only. A device loads the exact architecture the install carries, or the highest carried architecture of the same NVIDIA family that does not exceed it, so other NVIDIA generations (for example A100 and H100) are not served by this release. On AMD the match must be exact. An architecture the wheel does not carry needs a source build. mojolearn never substitutes the CPU for a GPU it cannot serve.

Python versions

CPython 3.10, 3.11, 3.12, 3.13 and 3.14, on both platforms.

Verified devices

Vendor	Device	Architecture	Backend
Apple	M4	Apple silicon	Metal
NVIDIA	RTX 4090	sm_89	CUDA
NVIDIA	L40S	sm_89	CUDA
AMD	Instinct MI325X	gfx942	HIP
AMD	Instinct MI300X	gfx942	HIP
CPU	x86-64 (Linux)	x86-64	CPU bindings
CPU	Apple silicon (macOS)	arm64	CPU bindings

Other GPUs of a carried architecture (for example the L40, L4 and RTX 6000 Ada on sm_89) load the same code. Use verify to check them.

Numeric modes

Mode	Contract
identical	The default. The same output bits for the same release, input bytes, model state, settings and supported configuration on NVIDIA and AMD GPUs. Apple GPU and CPU results are recorded beside them.
deterministic	Repeated runs on the same device and build give the same bits. No cross-vendor promise.
fast	Apple GPUs only. Tuned for throughput. No identity or repeatability promise.

Family	identical	deterministic	fast
Gradient boosting	yes	yes	Apple GPUs
Random forests	yes	yes	Apple GPUs
Extra Trees	yes	yes	Apple GPUs
Other tree models (decision trees, AdaBoost, random-trees embedding, isolation forest)	yes	no	Apple GPUs
Neural networks, GPU surface	yes	no	Apple GPUs
Every other family	yes	no	Apple GPUs
CPU bindings (all families)	yes	no	no

fast is opt-in, tuned on Apple and runs on Apple GPUs only. Asking a family for a mode it does not ship raises a named error. Bits may change between releases. See Numeric modes.

The CPU surface

On a CPU-only install, fit trains on the CPU for every estimator with a CPU binding and refuses by name otherwise, and saved models predict on the CPU. Every CPU binding is checked bit for bit against the recorded GPU results. A few gradient boosting training configurations have no CPU path (sample weights, CTR-encoded categorical columns, and some categorical and eval-set combinations outside symmetric trees). See Devices and the CPU and Saving and loading models.

What counts as verified

A cross-vendor result is accepted only when the same source, input bytes and numeric profile ran on each vendor, each run recorded its device and toolchain, the raw-bit hashes agree, and refused configurations are reported as refusals rather than passes. python -m mojolearn verify checks your installed package against the shipped reference results.