NVIDIA CUDA

Requirements

• An NVIDIA driver (nvidia-smi works).
• A CUDA build of PyTorch in the interpreter that runs jobs. Check with python -c "import torch; print(torch.cuda.is_available())".

gpulayer resources detect

Each GPU is listed with its memory and compute capability. Any CUDA workload can use them once the driver works: CUDA binaries, CuPy, JAX, TensorRT, PyTorch. PyTorch jobs additionally need the torch.cuda capability (a CUDA build of torch in the worker's Python); if it is missing, resources detect prints a note and PyTorch jobs are not placed there.

Train

trainer = glt.Trainer(max_epochs=10, accelerator="cuda", devices=4, precision="bf16-mixed")
trainer.fit(model, train_loader, val_loader)

• devices is the number of GPUs (one process per GPU) on each node. The job gets CUDA_VISIBLE_DEVICES set to that many GPUs.
• Multi-GPU and multi-node runs use DDP with the nccl backend.
• precision="16-mixed" uses a gradient scaler; bf16-mixed needs a GPU with bfloat16 support.

Placement constraints

Ask for specific hardware through @gl.remote(...), launch(...) or a JobSpec:

@gl.remote(accelerator="cuda", gpus=2, gpu_memory_gb=40)
def finetune(...): ...

ResourceRequest(min_cuda_compute_capability="8.0") restricts jobs to Ampere or newer.

Validation status

CUDA detection and scheduling are tested with mocked nvidia-smi output, and execution uses the same process path as CPU and MPS jobs. Neither single-GPU nor NCCL multi-GPU runs have been validated on NVIDIA hardware in this repository yet. The hardware test (pytest -m hardware_cuda tests/hardware) runs automatically on a machine with a GPU; please report the result.