If a whole workflow runs on an HPC cluster. There is no need to upload/download files between the HPC fs and the artifact repository (e.g. S3). Of course one can save the absolute path of the files on HPC as parameters. While dflow can provide a more native support for this.
For example, one can specify input/output artifacts as usual but the storage backend is the HPC fs. It will be better if one can download/preview files on the frontend UI, where the backend implements an FTP client with HPC.