Persist model weights and datasets across Modal containers with a modal.Volume

Documented steps

1. Create or reference a distributed volume: `vol = modal.Volume.from_name("model-cache", create_if_missing=True)`. Without create_if_missing=True a missing volume raises instead of being created.
2. Attach it to the Function by mount path: `@app.function(volumes={"/cache": vol})`. Only files written UNDER the mount path (e.g. /cache/llama.safetensors) land in the Volume; anything written elsewhere lives in the ephemeral container filesystem and is lost.
3. Inside the container, write files normally, then call `vol.commit()` to publish the changes so other containers can see them.
4. In a reader container call `vol.reload()` to pull in changes committed by other containers since the container started; without reload the view is frozen at container start.
5. For bulk uploads from local, use the batch context manager: `with vol.batch_upload() as batch: batch.put_file(local, "/remote/path")` and `batch.put_directory(...)`. Pass force=True to overwrite existing paths.
6. Inspect contents with `vol.listdir(path, recursive=False)`, `vol.iterdir(path)`, `vol.read_file(path)`; manage with `vol.remove_file(path, recursive=...)` and `vol.copy_files(src_paths, dst_path)`.
7. Mount read-only or scope to a subdirectory with `vol.with_mount_options(read_only=True, sub_path="subdir")` — this returns a new Volume handle, it does not mutate in place.
8. For throwaway state in a script use `with modal.Volume.ephemeral() as vol:` — it is destroyed when the context exits.
9. Reference: https://modal.com/docs/guide/volumes