b9979

ggml-org/llama.cppb9979Jul 12, 2026by github-actions[bot]same day as b9978

What changed since b9978

medium risk

b9979 fixes silent prompt truncation on embedded NUL bytes in multimodal input, adding an explicit text length field.

Breaking changes

  • The mtmd_input_text struct now carries an explicit text_len field, so C/C++ callers that construct it directly must set the length or risk incorrect behavior.

New features

  • Multimodal text input now tracks an explicit length, allowing prompts containing NUL bytes to be tokenized in full.

Improvements

  • Prompt length handling in mtmd now mirrors llama_tokenize and the text-only path, making the two code paths consistent.

Bug fixes

  • Prompts with embedded NUL bytes are no longer silently cut off at the tokenizer boundary, which previously dropped all later messages and the assistant marker without any log output.

Upgrade notes

This is a one-commit follow-up to b9978 focused on the multimodal (mtmd) path. If you embed mtmd directly, update any code that initializes mtmd_input_text to populate the new text_len field; server and CLI users need no configuration changes. Rebuild against the new headers rather than mixing old binaries with new libraries.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Fixes a silent prompt truncation bug where prompts containing embedded NUL bytes were cut off at the tokenizer boundary, causing subsequent messages and markers to be dropped.

Key Highlights

  • Fix for NUL character handling
  • Message integrity preservation
  • Multimodal text input validation

Release Notes (excerpt)

<details open> mtmd: fix silent prompt truncation on embedded NUL (#25548) * mtmd: fix silent prompt truncation on embedded NUL mtmd_input_text carried the prompt as a bare const char* with no length, so a NUL byte in message content cut the prompt at the tokenizer boundary and dropped every later message plus the assistant marker, with no log. Add an explicit text_len and thread it through, matching llama_tokenize and the text only path. * cleanup ---------…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories