huggingface/transformers · error · ValueError

Calling `get_mtp_config` on a config without `num_mtp_layers

Error message

Calling `get_mtp_config` on a config without `num_mtp_layers`

What it means

ValueError from PreTrainedConfig.get_mtp_config: it deep-copies the decoder text config and builds the MTP (multi-token prediction) config, which requires num_mtp_layers to be defined. If the text config has no num_mtp_layers attribute (getattr returns None), MTP layers were never configured and building an MTP model is meaningless.

Source

Thrown at src/transformers/configuration_utils.py:1391

                    value = getattr(config_to_return, key)
                    delattr(config_to_return, key)
                    setattr(config_to_return, new_key, value)

        return config_to_return

    def get_mtp_config(self) -> "PreTrainedConfig":
        """
        Returns the mtp text config to be used to create the MTP model. Since the MTP layers are created by instantiating
        the same classes as the main model, we need to overwrite index-specific properties of the config such as `layer_types`
        or `mtp_layer_types` to create the correct mtp layers and avoid indexing issues in the layers (because MTP layers restart
        the indexing of layers at 0).
        """
        # Start from the text config
        text_config = copy.deepcopy(self.get_text_config(decoder=True))
        num_mtp_layers = getattr(text_config, "num_mtp_layers", None)
        # In this case, raise
        if num_mtp_layers is None:
            raise ValueError("Calling `get_mtp_config` on a config without `num_mtp_layers`")

        layer_types = getattr(text_config, "layer_types", None)
        mtp_layer_types = getattr(text_config, "mtp_layer_types", None)
        mlp_layer_types = getattr(text_config, "mlp_layer_types", None)
        mtp_mlp_layer_types = getattr(text_config, "mtp_mlp_layer_types", None)

        # Replace the index-based fields so that we can call `XXXDecoderLayer(config, layer_idx)` with the mtp config, and create
        # the correct layers
        if layer_types is not None:
            if mtp_layer_types is None:
                raise ValueError(
                    "Calling `get_mtp_config` on a config containing `layer_types` without `mtp_layer_types` is ambiguous"
                )
            text_config.layer_types = mtp_layer_types
        if mlp_layer_types is not None:
            if mtp_mlp_layer_types is None:
                raise ValueError(
                    "Calling `get_mtp_config` on a config containing `mlp_layer_types` without `mtp_mlp_layer_types` is ambiguous"

View on GitHub (pinned to a597f97485)

Solutions

1. Guard the call: only invoke get_mtp_config() when getattr(config, 'num_mtp_layers', None) is not None.
2. If MTP is intended, set num_mtp_layers (and the corresponding mtp layer-type fields) on the config.
3. Use the non-MTP model class for checkpoints without MTP weights.

Example fix

// before
mtp_cfg = config.get_mtp_config()  # ValueError on non-MTP checkpoint

// after
mtp_cfg = config.get_mtp_config() if getattr(config, "num_mtp_layers", None) else None

if getattr(config.get_text_config(decoder=True), "num_mtp_layers", None) is None:
    skip_mtp = True  # do not call get_mtp_config / build MTP model

def has_mtp_layers(config) -> bool:
    return getattr(config.get_text_config(decoder=True), "num_mtp_layers", None) is not None

try:
    mtp_config = config.get_mtp_config()
except ValueError as e:
    if "without `num_mtp_layers`" in str(e):
        mtp_config = None  # checkpoint has no MTP

When it happens

Trigger: Calling config.get_mtp_config() (directly or via MTP model construction) on a model whose config lacks num_mtp_layers, e.g. a checkpoint not trained with MTP (most standard LLMs).

Common situations: Generic code paths that unconditionally build MTP heads for models like DeepSeek-MTP/GLM-style architectures; loading a non-MTP checkpoint into MTP-aware code.

Related errors

• The embed_dim ({self.embed_dim}) is not a multiple of the nu
• The `{layer_types}` entries must be in {allowed_types} but g
• `num_hidden_layers` ({self.num_hidden_layers}) must be equal
• Calling `get_mtp_config` on a config containing `layer_types
• File not found: {audio}

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/ca879b6582ad9f71. Report an issue: GitHub.