b10311

ggml-org/llama.cppb10311Aug 7, 2026by github-actions[bot]same day as b10310

What changed since b10310

low risk

b10311 is a single bug fix: Qwen3-TTS no longer reads the utterance twice during generation.

Bug fixes

  • MTMD Qwen3-TTS no longer re-feeds the text stream during generation, so the talker stops reading the utterance twice before emitting codec_eos.
  • The streaming overlay now uses the single tts_pad row that matches the non-streaming prefill layout instead of the streaming prompt layout.

Upgrade notes

This is a drop-in patch with no config, API, or model-format changes; only users of Qwen3-TTS via the MTMD pipeline are affected. Rebuild or pull the new binaries if you observed duplicated speech output. The ARM64 HWCAP/FP16 changes from b10310 remain in place.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

This release fixes a bug in the Multi-Modal Text-to-Speech (MTMD) pipeline for Qwen3-TTS. The issue caused the model to feed the text stream twice during generation, leading to the talker reading the utterance twice.

Key Highlights

  • Fixed MTMD (Qwen3-TTS) feeding text stream twice during generation (#26706)
  • Corrected the streaming overlay to match the non-streaming prefill layout

Release Notes (excerpt)

<details open> mtmd: stop feeding the text stream again during Qwen3-TTS generation (#26706) The reference implementation has two mutually exclusive prompt layouts. In non streaming mode the prefill carries the whole utterance text plus tts_eos summed with codec_pad, and the trailing text hidden collapses to a single tts_pad row. In streaming mode the prefill carries only the first text token and the trailing rows stream the rest of the text followed by tts_eos.…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories