Any GPU framework

GPUlayer does not depend on a particular GPU library. Workers report which accelerators they have and which GPU frameworks are installed in the Python that runs jobs; jobs say what they need; the scheduler matches the two.

What workers report

gpulayer resources detect shows it for the current machine:

device    [0] mps Apple arm64 GPU (Metal) (32.0 GB unified) independent=hardware-validated distributed=unsupported

and gpulayer worker status --json includes, for example:

"frameworks": {"mlx": "0.32.4", "numpy": "2.4.6", "torch": "2.15.0", "torch.mps": "2.15.0"}

A GPU counts as usable when its driver or platform is, so a CUDA box without PyTorch can still run CUDA binaries, CuPy or JAX jobs.

What jobs ask for

Name the accelerator and the frameworks with version ranges:

@gl.remote(accelerator="cuda", frameworks={"jax": ">=0.4.30"})
def train_jax(steps: int) -> dict:
    import jax
    ...

@gl.remote(accelerator="mps", frameworks={"mlx": ">=0.10"})
def matmul_benchmark(n: int) -> dict:
    import mlx.core as mx
    ...

@gl.remote(accelerator="cuda", gpus=1, frameworks={"cupy": ">=13"})
def fft(size: int) -> float:
    import cupy as cp
    ...

gpulayer submit --accelerator cuda --framework 'onnxruntime>=1.18' -- python infer.py

For common combinations, a runtime profile such as profile="pytorch-cuda", "jax-cuda" or "mlx" fills in the accelerator and frameworks for you.

A worker missing a framework, or with the wrong version, is rejected with a reason such as framework 'jax' not installed or mlx 0.9.0 does not satisfy >=0.10. The PyTorch integration adds torch plus torch.cuda or torch.mps automatically.

Inside the job, GPULAYER_ACCELERATOR says which accelerator was granted, and for CUDA jobs CUDA_VISIBLE_DEVICES lists the granted GPUs, so any CUDA library uses the right devices.

What has been run

Multi-process jobs in other frameworks

Distributed jobs are launched with torchrun today. Frameworks with their own launchers (JAX multi-host, MPI, Horovod) are planned; until then, run them as independent jobs or coordinate them yourself.