Feature: Add Fun-CosyVoice3-0.5B-2512 native TensorRT support

Request type

New model or checkpoint support.

Problem and use case

Add a native CosyVoice3 family for reference-voice-conditioned text-to-speech. Users should be able to prepare a reference voice and generate speech through TensorRT-Model-Connect without a PyTorch or ONNX Runtime fallback for learned inference.

Proposed outcome

• Keep checkpoint mapping, TensorRT graphs, reference frontend, orchestration, and validation owned by families/cosyvoice3.
• Provide native CAMPPlus and speech-tokenizer reference preparation, speech-token LLM generation, conditioning, Flow, and HiFT waveform generation.
• Support FP32, single-GPU, batch-one offline synthesis with 24 kHz mono output and one prepared voice per bundle.
• Package component engines for native C++ audio generation, with explicit profile limits and errors for oversized requests or generation that reaches its limit without a stop token.
• Retain component validation and native runtime/tokenizer contracts, and distinguish synthesis smoke checks from numerical parity and audio-quality qualification.

Model and target details

• Checkpoint: FunAudioLLM/Fun-CosyVoice3-0.5B-2512.
• Pinned revision: 29e01c4e8d000f4bcd70751be16fa94bf3d85a18.
• Task: text-to-speech .
• Initial precision: FP32.
• Initial validation target: RTX 4070 Laptop GPU (8 GB), TensorRT 11.1 / CUDA 13.
• Initial runtime path: explicit component build, voice preparation, and packaging, followed by native C++ generate-audio. Generic trtmc build integration remains follow-up work.

Submission checks

• I searched open and closed issues and found no duplicate request.
• I removed secrets, private/internal evidence, personal paths, and restricted artifacts.

Start by examining the existing family structure and the proposed families/cosyvoice3 ownership boundary, then trace the explicit component build, voice preparation, packaging, and native C++ generate-audio path. Done means FP32 single-GPU batch-one synthesis works at 24 kHz mono with the pinned checkpoint, component validation and runtime/tokenizer contracts are retained, and profile-limit and stop-token errors are covered.