What causes "Expected key `target` to instantiate." in Stability-AI/generative-models?

Ensure every model/submodule config dict has a 'target' key with the full dotted class path

Stability-AI/generative-models · critical · KeyError

Expected key `target` to instantiate.

Error message

Expected key `target` to instantiate.

What it means

instantiate_from_config expects a dict with a 'target' key holding the dotted python path of the class to build. Sentinel configs '__is_first_stage__' and '__is_unconditional__' short-circuit to None; anything else without 'target' raises KeyError('Expected key `target` to instantiate.').

Source

Thrown at sgm/util.py:174

    Take the mean over all non-batch dimensions.
    """
    return tensor.mean(dim=list(range(1, len(tensor.shape))))


def count_params(model, verbose=False):
    total_params = sum(p.numel() for p in model.parameters())
    if verbose:
        print(f"{model.__class__.__name__} has {total_params * 1.e-6:.2f} M params.")
    return total_params


def instantiate_from_config(config):
    if not "target" in config:
        if config == "__is_first_stage__":
            return None
        elif config == "__is_unconditional__":
            return None
        raise KeyError("Expected key `target` to instantiate.")
    return get_obj_from_str(config["target"])(**config.get("params", dict()))


def get_obj_from_str(string, reload=False, invalidate_cache=True):
    module, cls = string.rsplit(".", 1)
    if invalidate_cache:
        importlib.invalidate_caches()
    if reload:
        module_imp = importlib.import_module(module)
        importlib.reload(module_imp)
    return getattr(importlib.import_module(module, package=None), cls)


def append_zero(x):
    return torch.cat([x, x.new_zeros([1])])


def append_dims(x, target_dims):

View on GitHub (pinned to e8cd657656)

Solutions

Ensure every model/submodule config dict has a 'target' key with the full dotted class path
Check YAML indentation so 'target' sits inside the intended node, not a sibling
If the sentinel semantics apply, pass the exact strings '__is_first_stage__' or '__is_unconditional__'
Print the offending config dict at the call site to see what was actually received

Example fix

// before
first_stage_config:
  params:
    ckpt_path: model.ckpt
// after
first_stage_config:
  target: sgm.models.autoencoder.AutoencoderKL
  params:
    ckpt_path: model.ckpt

Defensive patterns

Strategy: validation

Validate before calling

def safe_instantiate(config):
    if isinstance(config, str) and config in ('__is_first_stage__', '__is_unconditional__'):
        return None
    if not isinstance(config, dict) or 'target' not in config:
        raise KeyError(f"config missing 'target': {config!r}")
    return instantiate_from_config(config)

Type guard

def is_instantiable_config(c) -> bool:
    return isinstance(c, dict) and 'target' in c

Try / catch

try:
    stage = instantiate_from_config(first_stage_cfg)
except KeyError as e:
    if 'target' in str(e):
        logging.error('config lacks target: %s', json.dumps(first_stage_cfg, default=str))
    return None  # or treat as __is_first_stage__ sentinel

Prevention

Verify YAML indentation: target/params must be siblings inside the intended node
Keep full first-stage/cond-stage configs in checkpoint metadata instead of trimmed dicts
Never pass an empty dict or params-only dict to instantiate_from_config
Write a CI check that walks all model YAMLs and asserts every node has 'target'

When it happens

Trigger: Passing a config dict to instantiate_from_config (from sgm/util.py:174) that lacks 'target' — e.g. an empty dict, a dict with only 'params', or a mis-nested dict where target ended up in a sub-dict.

Common situations: YAML indentation mistakes in model configs (params merged at the wrong level); passing None-handling sentinels incorrectly; checkpoint-loading code (load_model_from_config, apply_ckpt) hitting a first-stage/cond-stage config that was trimmed; calling configure_optimizers-style paths with partial config objects.

Related errors

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/28e199e54e8e9175. Report an issue: GitHub.