gpulayer 0.8.0

Train PyTorch on every GPU you can reach.
One Trainer, from your laptop to many machines.
Install · Quickstart · Performance · AI agents · Architecture · Docs · Examples · Changelog

The GPUlayer Trainer runs the same PyTorch script in your process, on every GPU of a machine, or across many machines: laptops, workstations, servers and cloud VMs. It picks machines and GPUs that can train the model, starts every process, and brings back logs, checkpoints and metrics. It uses PyTorch Lightning's names (LightningModule hooks, manual optimization, DataModule, callbacks, loggers, profilers, tuner), so most Lightning code moves with an import change.

NVIDIA CUDA, Apple Metal (MPS) and CPU are validated on hardware. AMD ROCm is implemented and not yet validated. Docs: gpulayer.dev/docs.

v0.8.0. Validated on a Tesla T4, 4× RTX 3070, 2× A100 (NVLink), an Apple M2 Pro, and distributed CPU training on up to 16 Google Cloud machines. NCCL across machines is implemented but not validated. See Status.

Install

pip install gpulayer

Check what this machine can train on:

gpulayer resources detect

To train with PyTorch, install it too: pip install "gpulayer[torch]".

Quickstart

import gpulayer.pytorch as glt

trainer = glt.Trainer(max_epochs=5, accelerator="cuda", devices=4, num_nodes=2)
trainer.fit(model, train_loader)          # 2 machines x 4 GPUs, DDP over NCCL

# The same file in this process, as 2-process DDP, or on 2 nodes: only the flags change
python examples/trainer_regression/train.py
python examples/trainer_regression/train.py --accelerator cpu --devices 2
python examples/trainer_regression/train.py --accelerator cpu --nodes 2 --devices 2

More machines: start a server on one, join the others, and point your laptop at it. Each line is printed by gpulayer up; on a single machine, use three terminals.

gpulayer up                                                    # machine 1: the server, runs jobs too
gpulayer join https://192.168.1.56:8470 --token glj_... --fingerprint sha256:...   # each GPU machine
export GPULAYER_URL=https://192.168.1.56:8470 GPULAYER_TOKEN=glc_... GPULAYER_FINGERPRINT=sha256:...
python examples/trainer_regression/train.py --accelerator cpu --nodes 2   # one node per worker

The full quickstart shows the output of each step.

Performance

Against PyTorch Lightning, with the same model, data and optimizer and each trainer's defaults, the GPUlayer Trainer was faster in 10 of 11 GPU cases and in every CPU case.

Keeping a 4-machine run in one zone instead of two regions made it 1.38x faster (42.5 s vs 58.8 s). Method and logs: Performance; scripts in benchmarks/.

How it works

The Trainer submits a run that says what it needs (accelerator, number of GPUs and machines, GPU memory). A control plane, inside your process or run as a server with gpulayer up, picks machines that can train it: the best-connected GPUs of a machine, and machines in one zone when it can. A worker on each machine starts the ranks, streams logs and uploads checkpoints. Workers open every connection themselves, so machines can sit behind NAT. More in the Architecture overview.

The same engine runs plain commands and Python functions as jobs, and gpulayer mcp lets coding agents start runs and read their results (AI agents).

With one GPU on the machine you are using, a plain training loop is simpler. Jobs are not sandboxed, so use GPUlayer among people who trust each other.

Development

uv pip install -e '.[dev,examples]'
.venv/bin/pytest && .venv/bin/ruff check src tests && .venv/bin/mypy src

See Contributing.

License

Apache License 2.0, see LICENSE.