Stability-AI/generative-models · error · ValueError

-n/--name and -r/--resume cannot be specified both.If you wa

Error message

-n/--name and -r/--resume cannot be specified both.If you want to resume training in a new log folder, use -n/--name in combination with --resume_from_checkpoint

What it means

main.py's CLI argument validation rejects a launch that passes both -n/--name and -r/--resume, because --resume derives the log folder from the checkpoint path while --name would create a new one — the two are mutually exclusive. The library forces you to choose: resume in-place (only --resume) or resume in a fresh folder (--name plus --resume_from_checkpoint). This check runs immediately after argument parsing, before any model or data loading.

Source

Thrown at main.py:552

    #   callbacks:
    #       callback1:
    #           target: importpath
    #           params:
    #               key: value

    now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")

    # add cwd for convenience and to make classes in this file available when
    # running as `python main.py`
    # (in particular `main.DataModuleFromConfig`)
    sys.path.append(os.getcwd())

    parser = get_parser()

    opt, unknown = parser.parse_known_args()

    if opt.name and opt.resume:
        raise ValueError(
            "-n/--name and -r/--resume cannot be specified both."
            "If you want to resume training in a new log folder, "
            "use -n/--name in combination with --resume_from_checkpoint"
        )
    melk_ckpt_name = None
    name = None
    if opt.resume:
        if not os.path.exists(opt.resume):
            raise ValueError("Cannot find {}".format(opt.resume))
        if os.path.isfile(opt.resume):
            paths = opt.resume.split("/")
            # idx = len(paths)-paths[::-1].index("logs")+1
            # logdir = "/".join(paths[:idx])
            logdir = "/".join(paths[:-2])
            ckpt = opt.resume
            _, melk_ckpt_name = get_checkpoint_name(logdir)
        else:
            assert os.path.isdir(opt.resume), opt.resume

View on GitHub (pinned to e8cd657656)

Solutions

1. Drop -n/--name and keep only -r/--resume to resume inside the original log folder.
2. Keep -n/--name but replace -r/--resume with --resume_from_checkpoint <path> to resume into a new log folder.
3. Fix the launcher script so name and resume flags are set mutually exclusively (e.g. only emit -r when -n is absent).

Example fix

// before
python main.py --base configs/stable-diffusion/v1-finetune.yaml -n myrun -r logs/2024-01-01T00-00-00/checkpoints/last.ckpt
// after
python main.py --base configs/stable-diffusion/v1-finetune.yaml -n myrun --resume_from_checkpoint logs/2024-01-01T00-00-00/checkpoints/last.ckpt

import sys
assert not ('--name' in sys.argv and '--resume' in sys.argv), \
    "Use -n with --resume_from_checkpoint, not -r"
# then run: python main.py ...

When it happens

Trigger: Running the training entrypoint with both -n/--name and -r/--resume set to non-empty values, e.g. `python main.py -n myrun -r logs/2023-.../checkpoints/last.ckpt --base configs/...`. The ValueError fires in the argparse post-processing block at main.py:552 whenever `opt.name and opt.resume` are both truthy.

Common situations: Shell scripts or SLURM launchers that append `-n $RUN_NAME` unconditionally while also enabling resume; developers copying a working resume command and adding a custom run name; shell variable expansion leaving both flags populated (e.g. `-n ""` vs `-n "foo"`).

Related errors

• Cannot find {}

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/be5f085c6d18bce8. Report an issue: GitHub.