llama.cpp ROCm regression: prompt processing drops from ~1200 to ~460 t/s on RX 9070XT
uber-linny · reddit · 2026-10-31
A Reddit user with an RX 9070XT reports a likely llama.cpp performance regression. On the Lemonade B1356 Fast ROCm build, prompt processing for Qwen3.5 27B IQ3S (3.4375 bpw, ngl 99, fa on) hits 1049–1214 t/s across pp512–pp4096; on the B1357 build the same setup drops to 460–495 t/s, roughly half speed. Self-compiling llama.cpp on top of the fast B1356 ROCm build restores the slow speeds, suggesting a regression introduced upstream rather than a packaging issue. The author is asking developers to identify and fix the cause.
More from Infra
- YOLOv2 Face Detection Hits 30fps on a K210 Chip via Custom C++ Firmware — alexcovo_eth · 2026-10-31
- Playable Environments: Millisecond Cloud Branching as the Next Test Bed for Agents — matei_zaharia · 2026-10-31
- Apple silicon is so overpowered only local LLM runners and 8K editors notice — vivekhaldar · 2026-10-31
- Ming-TTS inference ported to Zig/MLX hits 18x speedup on M4/M5 chips — s_busso · 2026-10-31
- Wafer hits #3 on OpenRouter by daily token volume at 1.9T — ycombinator · 2026-10-31
- Blogger argues $30K local Kimi K3.1 setup is worth it, despite taking 4.76 years to match GPT-6.1 Terra output — bookwormengr · 2026-10-31