xai-org/x-algorithm · error · ValueError

The value of top_feedforward specified

Error message

The value of top_feedforward specified ({top_feedforward}) does not match that in the checkpoint for {clip_model_type} ({config['top_feedforward']}).

What it means

When loading a fine-tuned checkpoint, Model.__init__ compares the top_feedforward hyperparameter you pass to the value stored in the checkpoint's config; a mismatch means the caller's model architecture would differ from the trained one, so weight loading would be invalid and it raises.

Solutions

1. Open the checkpoint's config dict and read config['top_feedforward'], then pass exactly that value
2. Or do not override top_feedforward; let the code take the checkpoint's value
3. If you intentionally changed the architecture, retrain and export a new checkpoint
4. Keep a single source of truth for hyperparameters alongside checkpoints

Example fix

# before
Model(clip_model_type='ViT-B/32-finetuned', top_feedforward=4096)
# after
Model(clip_model_type='ViT-B/32-finetuned', top_feedforward=1024)  # from checkpoint config

ckpt = torch.load(local_model_path, map_location='cpu')
expected_ff = ckpt['config']['top_feedforward']
assert top_feedforward == expected_ff, f"checkpoint has {expected_ff}"

try:
    Model(clip_model_type=t, top_feedforward=ff)
except ValueError as e:
    if 'top_feedforward' in str(e):
        ff = torch.load(path)['config']['top_feedforward']
        Model(clip_model_type=t, top_feedforward=ff)
    else:
        raise

When it happens

Trigger: Instantiating Model with top_feedforward=2048 while the checkpoint for that clip_model_type was trained with top_feedforward=1024; changing the hyperparameter in config without retraining; loading someone else's checkpoint with your own defaults.

Common situations: Hyperparameter tuning changed top_feedforward but old checkpoints are still loaded; shared checkpoint files across experiments with different heads; defaults in code drifted from the exported config.

Related errors

• head checkpoint was trained against a different backbone —…
• head registry hash != — refusing to start
• unexpected head param layout
• Unknown model type: . Choices
• All weights must be positive

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/a1c6c8ffa0b99812. Report an issue: GitHub.

Appendix: source

Thrown at clip/model.py:123

        self._device = "cuda:0" if self.use_gpu else "cpu"
        if self.use_gpu:
            self._prepare_model_for_gpu()

        self.clip_model.logit_scale.to(self._device)

        if clip_model_type in self.TWITTER_MODELS:
            map_location = None if self.use_gpu else torch.device("cpu")
            local_model_path = io_utils.maybe_download_file(
                self.TWITTER_MODELS[clip_model_type]
            )
            print(f"loading model from: {local_model_path}")
            checkpoint_contents = torch.load(
                local_model_path, map_location=map_location
            )

            config = checkpoint_contents["config"]
            if config["top_feedforward"] != top_feedforward:
                raise ValueError(
                    f"The value of top_feedforward specified ({top_feedforward}) does not match that in the "
                    f"checkpoint for {clip_model_type} ({config['top_feedforward']})."
                )

            if self._multi_gpu:
                image_state_dict = checkpoint_contents[
                    "image_encoder_state_dict_multi_gpu"
                ]
                text_state_dict = checkpoint_contents[
                    "text_encoder_state_dict_multi_gpu"
                ]
            else:
                image_state_dict = checkpoint_contents["image_encoder_state_dict"]
                text_state_dict = checkpoint_contents["text_encoder_state_dict"]

            self.image_encoder.load_state_dict(image_state_dict, strict=True)
            self.text_encoder.load_state_dict(text_state_dict, strict=True)

View on GitHub (pinned to 24c60942c5)