llama.cpp users find prompt-speed hack

This is exactly the kind of performance folklore that matters in local inference: one obscure flag can make a bigger difference than swapping models.

• –llama.cpp is a serious local inference engine with broad hardware support, so small configuration changes can have outsized real-world impact
• –The reported gain centers on prompt processing, which is often the bottleneck that makes large local models feel sluggish in coding and chat workflows
• –GitHub discussion around batch-size versus ubatch-size shows the flag maps more to device-level computation than application-level batching, which explains why experimentation can pay off
• –This looks more like a model-and-hardware-specific tuning win than a universal rule, especially given recent Qwen prompt reprocessing issues reported by users on HIP/AMD setups