llama.cpp patch reports 25% Qwen3.5 prefill gain in MI100 test

The merged change raised Q5_K_M throughput from 422.29 to 529.34 tokens per second at a 512-token microbatch in the contributor's single-GPU test.

By · Published

A large graphics card with a circular fan and thick cables sits on a black test frame marked “RW,” with glowing lines and a geometric wall logo behind it.

The merged change raised Q5_K_M throughput from 422.29 to 529.34 tokens per second at a 512-token microbatch in the contributor's single-GPU test.