Change8

b8466

📦 llama-cppView on GitHub →
🐛 1 fixes🔧 2 symbols

Summary

This release fixes a critical bug related to out-of-bounds tensor reads during pooled embedding extraction for specific models like Qwen3-VL-Embedding, ensuring correct behavior for MEAN/CLS/LAST pooling.

🐛 Bug Fixes

  • Fixed tensor read out of bounds assertion when extracting pooled embeddings by using n_embd_out instead of n_embd_inp in encode() and decode() for MEAN/CLS/LAST pooling paths.

Affected Symbols