sgl-project/sglang · error · NotImplementedError

forward() is not supported in encoder_only mode. Use get_aud

Error message

forward() is not supported in encoder_only mode. Use get_audio_feature() instead.

What it means

The MiMo-v2 ASR outer wrapper refuses forward() when config.encoder_only is set, because in encoder-only mode there is no LM head / decode path; audio must be encoded via get_audio_feature().

Source

Thrown at python/sglang/srt/models/mimo_v2_asr.py:79

        return pattern.pad_input_tokens(input_ids, mm_inputs)

    def get_input_embeddings(self):
        if getattr(self.config, "encoder_only", False):
            return None
        return self.model.embed_tokens

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
        get_embedding: bool = False,
        pp_proxy_tensors: Optional[PPProxyTensors] = None,
    ) -> torch.Tensor:
        if getattr(self.config, "encoder_only", False):
            raise NotImplementedError(
                "forward() is not supported in encoder_only mode. "
                "Use get_audio_feature() instead."
            )

        hidden_states = general_mm_embed_routine(
            input_ids=input_ids,
            forward_batch=forward_batch,
            language_model=self.model,
            multimodal_model=self,
            positions=positions,
            pp_proxy_tensors=pp_proxy_tensors,
        )

        if not get_embedding:
            return self.logits_processor(
                input_ids, hidden_states, self.lm_head, forward_batch
            )
        return self.pooler(hidden_states, forward_batch)

View on GitHub (pinned to 0132848349)

Solutions

1. Call get_audio_feature() (the audio embedding path) instead of forward()
2. If you intended full ASR generation, remove "encoder_only": true from the config
3. Route through the multimodal/audio-specific serving path that already uses get_audio_feature

Example fix

# before
out = model.forward(input_ids, positions, forward_batch)
# after
audio_feats = model.get_audio_feature(input_audio, ...)

if getattr(model.config, 'encoder_only', False):
    feats = model.get_audio_feature(audio_input)
else:
    out = model.forward(...)

def is_encoder_only(model) -> bool:
    return bool(getattr(model.config, 'encoder_only', False))

try:
    model.forward(...)
except NotImplementedError as e:
    if 'encoder_only' in str(e):
        feats = model.get_audio_feature(audio_input)
    else:
        raise

When it happens

Trigger: Calling model.forward(...) (or a generic runner path) on MiMoV2ASR with config.encoder_only=True in the checkpoint config.

Common situations: Using the ASR model class standalone as an audio encoder (e.g. embedding service) but invoking the standard causal-LM forward path; misrouted pipelines.

Related errors

• unsupported audio item: loaded={loaded_type}, raw={raw_type}
• Multimodal data is corrupted or cannot be decoded: {e}
• Missing previous frame for delta payload
• kernel dispatch requires at least one tensor argument
• cos/sin shape does not cover image tokens and head_dim

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/820e1d760d273c42. Report an issue: GitHub.