Llama.cpp deep dive: Heterogeneous GPU setup boosts speed by 70% and enables 262k context

fintip · reddit · 2026-08-20

After 3 days of benchmarking on a heterogeneous 40GB VRAM setup (RTX 4090 laptop + RX 7900 XTX eGPU), the author optimized llama.cpp parameters to increase generation speed from 16 to 27 t/s and unlock the full 262k context. Key findings include: enabling speculative decoding (MTP + ngram) significantly boosts generation; the 7900 XTX excels at decoding tasks; TB4 bandwidth overhead is manageable; and a bug in llama.cpp's multi-GPU MTP implementation was identified and reported. The optimized launch command is provided.

Original post →

More from Infra

Infra channel →