RVC-Boss/GPT-SoVITS · error · RuntimeError

SoVits V3/4模型不支持流式推理模式

Error message

SoVits V3/4模型不支持流式推理模式

What it means

RuntimeError raised in the streaming decode branch when the loaded vits_model has no decode_streaming capability but streaming was requested. SoVITS v3/v4 use a different vocoder architecture whose decoder cannot do chunked streaming generation, so only v1/v2 (25hz/non-25hz classic) models support streaming; requesting it with v3/v4 hits the else-branch and raises.

Source

Thrown at GPT_SoVITS/TTS_infer_pack/TTS.py:1442

                            # token_padding_length = int(phones.shape[-1]*2)-_semantic_tokens.shape[-1]
                            # if token_padding_length>0:
                            #     _semantic_tokens = F.pad(_semantic_tokens, (0, token_padding_length), "constant", 486)
                            # else:
                            #     token_padding_length = 0

                            audio_chunk, latent, latent_mask = self.vits_model.decode_streaming(
                                                    _semantic_tokens.unsqueeze(0), 
                                                    phones, refer_audio_spec, 
                                                    speed=speed_factor,
                                                    sv_emb=sv_emb,
                                                    result_length=semantic_tokens.shape[-1]+overlap_len if not is_first_chunk else None,
                                                    overlap_frames=last_latent[:,:,-overlap_len*(2 if self.vits_model.semantic_frame_rate == "25hz" else 1):] \
                                                    if last_latent is not None else None,
                                                    padding_length=token_padding_length
                                                )
                            audio_chunk=audio_chunk.detach()[0, 0, :]
                        else:
                            raise RuntimeError(i18n("SoVits V3/4模型不支持流式推理模式"))
                        
                        if overlap_len>overlap_length:
                            audio_chunk=audio_chunk[-int((overlap_length+semantic_tokens.shape[-1])*upsample_rate):]

                        audio_chunk_ = audio_chunk
                        if is_first_chunk and not is_final:
                            is_first_chunk = False
                            audio_chunk_ = audio_chunk_[:-overlap_size]
                        elif is_first_chunk and is_final: 
                            is_first_chunk = False
                        elif not is_first_chunk and not is_final:
                            audio_chunk_ = self.sola_algorithm([last_audio_chunk, audio_chunk_], overlap_size)
                            audio_chunk_ = (
                                audio_chunk_[last_audio_chunk.shape[0]-overlap_size:-overlap_size] if not is_final \
                                    else audio_chunk_[last_audio_chunk.shape[0]-overlap_size:]
                                    )

                        last_latent = latent

View on GitHub (pinned to d523079fc0)

Solutions

1. Switch to a v1 or v2 SoVITS model if streaming/realtime output is a hard requirement.
2. Or disable streaming (return_finished_audio=True / stream=False) so inference falls back to full-clip generation even with v3/v4.
3. If you build a client, probe model capability (check model_version or hasattr(model, 'decode_streaming')) before requesting streaming, and fall back to non-streaming automatically.

Example fix

# before
for chunk in handler.run(..., stream=True):  # RuntimeError: SoVits V3/4模型不支持流式推理模式
    play(chunk)

# after
streaming_ok = getattr(handler.vits_model, "decode_streaming", None) is not None
if streaming_ok:
    for chunk in handler.run(..., stream=True):
        play(chunk)
else:
    play(handler.run(..., return_finished_audio=True))

streaming_supported = getattr(handler.vits_model, "decode_streaming", None) is not None
if stream_requested and not streaming_supported:
    stream_requested = False  # fall back to full-clip generation

def model_supports_streaming(handler) -> bool:
    """v1/v2 sovits expose decode_streaming; v3/v4 do not."""
    return getattr(handler.vits_model, "decode_streaming", None) is not None

try:
    yield from handler.run(..., stream=True)
except RuntimeError as e:
    if "不支持流式" in str(e):
        yield handler.run(..., return_finished_audio=True)
    else:
        raise

When it happens

Trigger: Calling the streaming inference API (stream=True / the generator variant of run) while a SoVITS v3 or v4 model is loaded — the code calls self.vits_model.decode_streaming only when the attribute exists, otherwise raises RuntimeError with this message.

Common situations: API server configured for streaming responses but user switched the sovits checkpoint to a v3/v4 lora or full model; new deployment copies a streaming example config while shipping v4 weights; client demands realtime chunked audio but the model cannot provide it.

Related errors

• SoVITS %s 底模缺失，无法加载相应 LoRA 权重
• {path_sovits}SoVITS {model_version}底模缺失，无法加载相应 LoRA 权重
• {path_sovits}SoVITS {model_version}底模缺失，无法加载相应 LoRA 权重
• 搜索音频长度 ({search_len}) 必须大于等于参考音频长度 ({ref_len})

AI-assisted analysis of RVC-Boss/GPT-SoVITS@d523079fc0 (2026-08-15). Data as JSON: /api/errors/5e55cc0399781b28. Report an issue: GitHub.