vllm_omni.model_executor.models.voxcpm2.lora ¶
Merge a native VoxCPM2 LoRA checkpoint at startup for all requests.
merge_voxcpm2_lora ¶
merge_voxcpm2_lora(
adapter_path: str,
*,
base_lm: Module,
residual_lm: Module,
tts: Module,
) -> int
Validate and merge one local native safetensors adapter.
Call once after base weight loading and before fusion or graph capture. Validate all checkpoint keys and tensors before changing weights.