GPU Acceleration - Beam

Running Tasks on GPU

Pass gpu to a Python workload or --gpu to a container deployment. Install your framework and its compatible CUDA dependencies in the runtime image.

from beam import function


@function(gpu="A10G", memory="2Gi")
def inspect_gpu():
    import subprocess
    return subprocess.check_output(["nvidia-smi"], text=True)


if __name__ == "__main__":
    print(inspect_gpu.remote())

Availability

beam machine list
beam pool offers --gpu H100 --nodes 1 --ttl 1h --max-spend 10

beam machine list reports live serverless availability and on-demand inventory. GPU model names supported by the SDK are not a promise that every model is currently available serverless. Review offers for price, provider, region, and capacity, then use compute pools for reserved hardware.

GPU selection

A priority list lets the scheduler consider alternative types that your workload supports:

from beam import function


@function(gpu=["T4", "A10G", "RTX4090"], memory="2Gi")
def handler():
    return {"ok": True}

Ensure the model fits the smallest GPU in the list and that its libraries support each architecture. gpu_count requests multiple GPUs on a suitable machine. It does not distribute an application automatically; configure your framework’s multi-GPU execution separately. Availability and workspace limits still apply.

Regions and pools

Use a reserved offer’s provider and region selectors when placement matters. Set the workload’s pool to route it to that capacity. Use beam machine list --help and beam pool --help for the available filters and subcommands.