Unsloth's Gemma 4 mmproj Breaks with Newer llama.cpp, Causing Silent Multimodal Failures
Top_Speaker_7785 · reddit · 2026-08-06
A developer found that Unsloth's Gemma 4 mmproj file became incompatible with newer llama.cpp builds, causing multimodal features (image analysis, voice transcription) to silently fail, outputting <unused49> garbage tokens. The root cause is third-party quantizers not staying in sync with llama.cpp's internal format. Switching to ggml-org's official GGUF fixed the issue. The post details the debugging process, root cause analysis, and fixes, and raises community questions.
More from coding & agent
- YC-Backed Axelrod Labs Launches: AI Agent Workforce to Automate Hotel Operations — ycombinator · 2026-08-06
- AI Agents Extracting Employee Knowledge Face Incentive Misalignment — manosaie · 2026-08-06
- Cloudflare Unveils Browser in Workers: 4x Less Memory for Millions of Concurrent Agents — dinasaur_404 · 2026-08-06
- AI Reshapes Dev Division: Frontend Roles to Merge with Product and Design — vista8 · 2026-08-06
- Stop Re-asking: A Practical Workflow for Organizing AI Coding Chats — Ok_Negotiation_2587 · 2026-08-06
- Cloudflare Kitesurf MCP Server in Action: Giving Agents a Browser — dinasaur_404 · 2026-08-06