vllm_omni.utils ¶

Modules:

Name	Description
`audio`	Audio utility functions shared across models and entrypoints.
`audio_resample`	Streaming resampling, shared by the engine and the entrypoints.
`cuda_stream`	Owned CUDA streams for graphs whose library workspaces must stay isolated.
`custom_voice_io`	Shared helpers for custom voice profile tools.
`device_copy`	Host-to-device copies that do not block the host on queued CUDA/NPU work.
`forced_aligner`	Forced-aligner config + word-timestamp decoding for TTS.
`mm_outputs`	Utilities for handling multimodal outputs / building multimodal output
`profiler`
`qwen3_force_align_processor`	Qwen3 forced-aligner text/timestamp processor.
`seeded_exponential`	Batched per-request `exponential_(generator=...)` that is bit-identical to torch.
`speaker_cache`	Process-wide thread-safe LRU cache for speaker extraction artifacts.
`speakers`	Speaker-name helpers shared by serving and model-side voice selection.
`torch_compile`	Helpers for models that call `torch.compile` directly, outside vLLM's compile wrapper.
`tracking_parser`