llama.cpp b9200 Release

·2026.05.18 08:46

Key point

llama.cpp b9200 includes an optimization that improves MTP prompt processing speed.

llama.cpp b9200 includes an MTP prompt processing optimization.

Instead of copying logits for every token in the batch, it only uses the pre-norm needed for MTP.
This change reduces memory traffic and speeds up PP (prompt processing).
Based on community comments, testing is currently underway.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.