Storage

• runpod.NetworkVolume: Storage in a single data center. Resources that mount it run in that data center. Works with GPU and CPU compute.
• runpod.GlobalVolume (beta): Storage that isn’t tied to a data center. Requires GPU compute.

​

Attach a volume

import runpod

app = runpod.App("trainer")

checkpoints = runpod.NetworkVolume("checkpoints", size=10)

@app.task(gpu="4090", mounts={"/checkpoints": checkpoints})
def train(steps: int = 200):
    run_dir = checkpoints.path / "linear"
    run_dir.mkdir(parents=True, exist_ok=True)
    # ... train and save to run_dir ...
    return {"checkpoint": "linear"}

@app.task(gpu="4090", mounts={"/checkpoints": checkpoints})
def evaluate(checkpoint: str):
    model_file = checkpoints.path / checkpoint / "model.pt"
    ...

​

Volume references

runpod.NetworkVolume(name_or_id, size=50, datacenter=None, create=True)
runpod.GlobalVolume(name_or_id, create=True)

• If no volume with that name or ID exists, the SDK creates it the first time it provisions a resource that mounts it. Set create=False to require an existing volume instead.
• size is the size in GB for a new network volume. Global volumes grow with your data, so they don’t take a size.
• datacenter pins a new network volume to a data center. If you leave it unset, the SDK picks a data center that supports network storage and has hardware for every resource that shares the volume, respecting any datacenter setting on those resources. If you pin a data center that doesn’t support network storage, provisioning fails before the volume is created.
• An existing volume keeps its ID and data center. Every resource that mounts it must be able to run in that data center.

​

Mount rules

​

Read the mount path in your code

@app.queue(gpu="4090", mounts={"/runpod-volume": models})
def infer(prompt: str):
    weights = models.path / "llama" / "model.safetensors"
    ...

• Read .path inside remote functions, not at the top level of a module. The SDK imports your modules locally to discover resources, and the volume isn’t mounted there.
• Accessing .path raises an error if the volume isn’t mounted on the current worker, or if it’s mounted more than once.
• fn.local() runs your function on your machine without mounting remote storage.

​

Cache model weights

​

Billing

​

Next steps

• Compare network and global volumes.
• Build an app that shares checkpoints between tasks.