vllm_omni.model_executor.models.common ¶

Modules:

Name	Description
`alias_free_activation`	Alias-free activation for BigVGAN-style speech decoders.
`cfg_pairing`	Model-neutral classifier-free-guidance request pairing for the vLLM v1 scheduler.
`fused_sampling`	Fused Gumbel selection after native nucleus probability reductions.
`ming`	Components shared by the Ming-family model packages.
`nucleus_ras_sampling`	Shared TTS sampling primitives: nucleus (top-p/top-k) and RAS.
`ops`	Common operators shared across models and hardware backends.
`qwen3_code_predictor`	Qwen3 Code Predictor -- optimized re-prefill, no KV cache.
`short_kv_attention`	Short causal attention for the scratch KV cache inside a code predictor.
`snake_activation`	Shared Snake/SnakeBeta activations for speech decoders.
`talker_first_audio`	Shared eligibility; each model separately controls its first-audio opt-in.
`whisper_vq`	WhisperVQEncoder: HF WhisperEncoder + VQ codebook + inter-layer pooling.