Pure-C TTS engine gains Metal and CUDA GPU backends with major speed gains

Developers have added two optional GPU backends — Apple Metal and NVIDIA CUDA — to qwen3-tts, a pure-C inference engine for the Qwen3-TTS model, while keeping the default CPU path unchanged. On Apple Silicon, the Metal backend reduced the 0.6B model's real-time factor from 1.5 to 0.60 on an M1 and 0.36 on an M2 Pro, with streaming first-audio latency of 314 milliseconds. The CUDA backend achieved an RTF of 0.44 on an RTX 4060-class GPU, scaling down to approximately 0.12 on an RTX 4090. Key performance gains came from keeping weights, KV cache, and activations resident on the device and encoding full decode steps into a single command buffer rather than offloading individual operations. Server-side request batching further improved throughput by roughly 2.8x on Metal and 3.35x on CUDA, allowing multiple concurrent users to be served in approximately the time of a single request.