llama.cpp Fixes MTP Performance Bug, Boosting Qwen Token Generation by 10%

solyarisoftware · x · 2026-07-29

llama.cpp has fixed a performance bug related to Multi-Token Prediction (MTP). Previously, when loading MTP GGUFs with --fit, the NextN/MTP block was not counted as part of ngpulayers.

This caused the engine to offload one layer too few, leaving layer 0 running on the CPU and accidentally disabling the fused Gated Delta Net GPU kernel, which degraded token-generation performance. The fix from PR #26177 is included in release b10152. The contributor measured a 10% faster token generation speed with Qwen3.6-35B-A3B after the front layers were properly kept on the GPU.

Related event: llama.cpp One-Line Fix Boosts MTP Inference Speed by 10%(2 posts)→

Original post →

More from Infra

Infra channel →