llama.cpp Release b10311 Fixes Qwen3-TTS Text Streaming Bug
The llama.cpp project has released version b10311, which resolves a bug where the text stream was fed twice during Qwen3-TTS generation. This release fixes a prompt layout mismatch between streaming and non-streaming modes. This fix ensures that Qwen3-TTS, a multilingual text-to-speech model, does not repeat utterances or read text twice during speech generation. It improves the reliability and correctness of audio outputs when running Qwen3-TTS via llama.cpp. The issue occurred because the pipeline built a non-streaming prefill but applied a streaming overlay, causing the generator to process the utterance twice before emitting the end-of-stream codec token (codec_eos). The fix aligns the overlay to a single tts_pad row that matches the prefill.
## BACKGROUND
llama.cpp is a popular open-source LLM inference engine written in C/C++ that allows running models locally on consumer hardware. Qwen3-TTS is an open-source text-to-speech model developed by Alibaba's Qwen team that supports multilingual, expressive, and streaming speech generation.