microsoft/VibeVoice · error · RuntimeError

Output embeddings (lm_head) are not defined for this model.

Error message

Output embeddings (lm_head) are not defined for this model. Create one before calling set_output_embeddings if needed.

What it means

The streaming inference wrapper defines get_output_embeddings() -> None and makes set_output_embeddings a hard RuntimeError: the model has no lm_head (vocabulary projection) because TTS generation ends in the diffusion head and an eos classifier, not token logits. Calling set_output_embeddings violates that design.

Source

Thrown at vibevoice/modular/modeling_vibevoice_streaming_inference.py:232

        
    def get_input_embeddings(self):
        return self.model.get_input_embeddings()
    
    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)
    
    def get_output_embeddings(self):
        """
        This model does not define an `lm_head` (vocabulary projection).
        """
        return None
    
    def set_output_embeddings(self, new_embeddings):
        """
        No-op because there is no `lm_head`. Provided only to satisfy optional API calls.
        To enable, first create `self.lm_head` then allow assignment.
        """
        raise RuntimeError("Output embeddings (lm_head) are not defined for this model. "
                           "Create one before calling set_output_embeddings if needed.")
    
    def set_speech_tokenizers(self, acoustic_tokenizer=None):
        """Set the speech tokenizers used for encoding and decoding speech."""
        self.model.set_speech_tokenizers(acoustic_tokenizer)
    
    def set_ddpm_inference_steps(self, num_steps=None):
        self.ddpm_inference_steps = num_steps or self.config.diffusion_head_config.ddpm_num_inference_steps

    def prepare_inputs_for_generation(
        self,
        input_ids: torch.LongTensor,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        **kwargs,
    ):

View on GitHub (pinned to 94da20d98b)

Solutions

Remove the set_output_embeddings call — this model has no vocabulary head by design.
If you added vocabulary tokens, resize via the text tokenizer/embedding path instead of the output head.
get_output_embeddings() returns None; treat None as 'no head' in generic code paths.
For logits-style output use forward_tts_lm, which returns tts_eos_classifier logits.

Example fix

# before
model.set_output_embeddings(new_head)  # RuntimeError

# after
# no lm_head exists; use the TTS end-of-stream logits instead
out = model.forward_tts_lm(input_ids=ids, ...)
print(out.logits)  # from tts_eos_classifier

Defensive patterns

Strategy: type-guard

Validate before calling

if model.get_output_embeddings() is None:
    # no lm_head by design; skip embedding resize/tie logic
    pass

Type guard

def has_lm_head(model) -> bool:
    return model.get_output_embeddings() is not None

Try / catch

try:
    model.set_output_embeddings(head)
except RuntimeError:
    # model has no vocabulary head; not an error condition
    pass

Prevention

Check get_output_embeddings() is not None before calling setters
Don't port HF CausalLM boilerplate onto the streaming wrapper
Use forward_tts_lm logits for stop-token decisions

When it happens

Trigger: Generic transformers utilities (resize_token_embeddings, tie_weights, generation utilities) or user code calling model.set_output_embeddings(new_embeddings) on the streaming inference model.

Common situations: Reusing HF causal-LM boilerplate that resizes embeddings after adding tokens; hooking the model into tooling that assumes a standard CausalLM interface; migration from VibeVoice (non-streaming) which does have output embeddings.

Related errors

AI-assisted analysis of microsoft/VibeVoice@94da20d98b (2026-08-15). Data as JSON: /api/errors/a7d78d56fd6b7dc6. Report an issue: GitHub.