vllm_omni.model_executor.models.indextts2.configuration_indextts2 ¶

Bases: IndexTTS2Config

Configuration for the official IndexTTS 2.5 checkpoint.

IndexTTS 2.5 keeps the GPT/S2Mel backbone shapes used by IndexTTS 2, but switches text tokenization and speaker conditioning, and decodes semantic codes with EnhancedCodec. GPT latent conditioning is disabled by default upstream; the opt-in path is an experimental vLLM-Omni-specific latent variant with no official runnable reference output.

model_type = 'indextts2_5'

Bases: PretrainedConfig

model_type = 'indextts2'

output_sample_rate = int(
    self.s2mel["preprocess_params"].get("sr", 22050)
    if isinstance(self.s2mel, dict)
    else 22050
)

vocab_size = int(
    getattr(self, "vocab_size", 0)
    or self.gpt["number_mel_codes"]
)