Stability-AI/generative-models · error · NotImplementedError

NotImplementedError

Error message

NotImplementedError

What it means

FrozenOpenCLIPVisualStyleEmbedder (around modules.py:439) accepts layer='last' or layer='penultimate' to select which CLIP transformer output to use; any other string raises NotImplementedError in __init__. Unlike some OpenCLIP wrappers there is no support for indexed/hidden-layer names here.

Source

Thrown at sgm/modules/encoders/modules.py:439

            arch,
            device=torch.device("cpu"),
            pretrained=version,
        )
        del model.visual
        self.model = model

        self.device = device
        self.max_length = max_length
        self.return_pooled = always_return_pooled
        if freeze:
            self.freeze()
        self.layer = layer
        if self.layer == "last":
            self.layer_idx = 0
        elif self.layer == "penultimate":
            self.layer_idx = 1
        else:
            raise NotImplementedError()
        self.legacy = legacy

    def freeze(self):
        self.model = self.model.eval()
        for param in self.parameters():
            param.requires_grad = False

    @autocast
    def forward(self, text):
        tokens = open_clip.tokenize(text)
        z = self.encode_with_transformer(tokens.to(self.device))
        if not self.return_pooled and self.legacy:
            return z
        if self.return_pooled:
            assert not self.legacy
            return z[self.layer], z["pooled"]
        return z[self.layer]

View on GitHub (pinned to e8cd657656)

Solutions

1. Set layer='last' (layer_idx 0) or layer='penultimate' (layer_idx 1) in the embedder config
2. Check the exact string in the config YAML for case/whitespace/typos
3. If you need another layer, patch the class to map your layer string to the right layer_idx

Example fix

// before
params:
  layer: hidden
// after
params:
  layer: penultimate

ALLOWED_LAYERS = {'last', 'penultimate'}
layer = cfg.model.params.embedder.get('params', {}).get('layer', 'last')
if layer not in ALLOWED_LAYERS:
    raise ValueError(f"layer must be one of {ALLOWED_LAYERS}, got {layer!r}")

try:
    embedder = instantiate_from_config(emb_config)
except NotImplementedError:
    emb_config['params']['layer'] = 'last'
    embedder = instantiate_from_config(emb_config)

When it happens

Trigger: Constructing the embedder with layer set to something other than 'last' or 'penultimate', e.g. layer='hidden', layer=7, or a config omitting/misspelling the layer param.

Common situations: Copying config snippets from the Stable Diffusion (LDM-style) CLIP embedder which used layer='hidden'; typos like 'penultimate ' with whitespace; passing None when the config lacks params.

Related errors

• Unknown loss type {self.loss_type}
• rearranging not available for {len(in_shape)}-dimensional in
• unknown merge strategy {self.merge_strategy}
• provide num_res_blocks either as an int (globally constant)
• NotImplementedError

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/a6561ac993a31838. Report an issue: GitHub.