Hardware

Each Call declares the hardware it needs through client.run(). Hardware is allocated per Call, so different parts of the same application can use different combinations of CPUs, memory, and GPUs.

import outerproduct as op

run: op.Run[object] | op.Call[object] = client.run(
    fn=train_model,
    args=(training_data,),
    image=training_image,
    cpu=8,
    memory_bytes=32 * 1024**3,  # 32 GiB
    gpu=1,
    gpu_type="A100-80GB",
)

CPUs

Use cpu to specify the number of CPUs assigned to the Call. Requested CPUs are assigned to one worker. The default is one CPU.

run: op.Run[object] | op.Call[object] = client.run(
    fn=process_data,
    args=(dataset,),
    image=image,
    cpu=4,
)

Memory

Use memory_bytes to set the amount of RAM available to the worker. Memory is specified in bytes. The limit includes temporary shared memory (used by /dev/shm). If memory_bytes is omitted, OuterProduct provides 2 GiB of memory per requested CPU.

run: op.Run[object] | op.Call[object] = client.run(
    fn=process_data,
    args=(dataset,),
    image=image,
    memory_bytes=16 * 1024**3,  # 16 GiB
)

GPUs

Set gpu_type to request a particular GPU and gpu to specify the count. When gpu_type is provided and gpu is omitted, OuterProduct assigns one GPU by default.

run: op.Run[object] | op.Call[object] = client.run(
    fn=train_model,
    args=(training_data,),
    image=training_image,
    gpu_type="L4",
    gpu=2,
)

gpu=2 assigns two GPUs to the same worker and container. It does not start two workers or divide the function into two Calls. Your code must use a multi-GPU framework such as PyTorch Distributed to take advantage of both GPUs. To run independent work across multiple workers, use client.map() to create multiple Calls with one GPU instead. Your Image must also contain the drivers and libraries required by your application, such as a CUDA-enabled build of PyTorch.

Supported GPUs

OuterProduct supports several NVIDIA GPU types, including L4, L40S, A10G, A100, and H100. Available GPU models and supported counts depend on the platform’s execution backend and may change over time. The Python API does not currently expose a complete list of supported (gpu_type, gpu) configurations. You can inspect the GPU types and per-node counts currently available to your account from the Python API:

available_gpus: set[tuple[str, int]] = set()

for node in client.list_nodes():
    resources: op.ComputeResources | None = node.total_resources

    if resources is not None and resources.accelerator:
        available_gpus.add((resources.accelerator, resources.accelerator_count))

print(sorted(available_gpus))

This reports currently available nodes, not every configuration supported by the execution backend.

A supported (gpu_type, gpu) configuration may still take longer to start when capacity is limited. Unsupported configurations return an error.

Getting started

Execution

Governance

CPUs

Memory

GPUs

Supported GPUs

​CPUs

​Memory

​GPUs

​Supported GPUs

CPUs

Memory

GPUs

Supported GPUs