Llama-Launcher 1.3 Learns Your Hardware For Effortless AI Speed Boosts
A new update to llama-launcher brings hands-free performance tuning to users of the llama.cpp local inference engine. The tool is a lightweight graphical application that lets you build server launch commands by pointing and clicking, removing the need to type out long flag strings. Version 1.3 adds an auto-optimiser that tests flag combinations on your own hardware to find the fastest setup.
SolaryKryptic, the developer, refined the earlier sequential optimiser into a Bayesian-only method using Optuna’s TPE algorithm. The update makes it possible to squeeze more tokens per second from speculative decoding models like Gemma MTP without any manual tuning loops. The process runs benchmarks and perplexity checks in the background and surfaces the best-performing config.
Automatic performance tuning with Bayesian optimization
Key capabilities
- Detects CPU, GPU, and memory on launch.
- Browser for selecting .gguf model files.
- Optuna TPE search across core parameters.
- Live progress window with ETA and scores.
- PPL validation for cache-related changes.
- Baseline fallback if no trial wins.
- Real-time command preview and copy.
- Saves settings automatically between sessions.
Anyone running llama.cpp locally on Windows can use this to replace manual trial and error. The program handles everything from picking a GGUF file to testing thread counts, batch sizes, KV cache types, and speculative draft settings. It is especially helpful for people who experiment with MTP models and want to get the highest throughput without writing scripts. The optimiser respects hardware limits by reading your system specs upfront, so you don’t have to guess what commands your GPU and CPU can handle.
What you should know before running it
This version relies entirely on Bayesian (Optuna TPE) optimisation and no longer includes the old sequential method. The tool requires that you already have llama.cpp compiled and that llama-server.exe and llama-perplexity.exe are available. Because hardware detection uses WMI, the GUI currently works only on Windows. Future improvements are planned, and the developer is open to community suggestions via the repository.
"So far, I've seen upto a 15% improvement in speeds (as seen in the images) versus baseline commands with no tuning with Gemma 12B MTP during testing. — Source: Reddit
Project Links
