🦙 Llama Launcher

A powerful, Windows-native GUI launcher for llama.cpp's llama-server.exe. Streamline model management, server configuration, and local LLM inference with an intuitive interface, automatic GGUF scanning, and a robust caching system.

Platform: Windows 10/11 (64-bit) | Language: Delphi/Pascal

✨ Features

Category	Description
📦 GGUF Management	Automatic multi-threaded directory scanning, shard validation, and metadata extraction (architecture, layers, context, size, embedding, etc.)
💾 Smart Cache	Persistent `models-cache.txt` for instant metadata retrieval. Avoids repeated file parsing on startup
⚙️ Server Configuration	Full control over `llama-server` parameters: threads, context, GPU/CPU offloading, sampling (temp, top_k/p, min_p, repeat/presence penalty), flash-attn, KV cache types, and advanced flags
📋 Preset System	Per-model presets (auto-saved on launch) & generation presets (temperature, reasoning mode, seed, etc.) stored as `.ini` files
🌍 Multi-Language	25+ supported languages via INI-based translation system. Fallback to EN/FR if translation missing
📊 Real-time Logging	High-performance log buffer, memo display, auto-scroll, file export, and max-lines limiter
🔍 System Monitor	Live CPU/RAM/VRAM tracking in the status bar
🖥️ UX Enhancements	Tray icon with minimize-to-tray, drag & drop GGUF support, single-instance mutex, WebUI launcher, known arguments helper

📦 Requirements

Operating System: Windows 10 or Windows 11 (64-bit)
Backend Engine: llama-server.exe (from llama.cpp)
RAM/Storage: ~5 MB for the launcher itself (pre-compiled)
Permissions: Standard user rights (no admin required unless modifying system paths)

💡 The launcher is self-contained. All Delphi RTL and Indy components are statically linked or bundled.

📥 Installation & Setup

Download the latest release (LlamaLaunchers.zip) from the Releases page.
Extract the archive to your preferred directory (e.g., C:\Tools\LlamaLauncher\).
Place your llama-server.exe in a known directory (e.g., C:\llama.cpp\).
Launch LlamaLaunchers.exe.
Configure the executable path via 🔧 Settings → llama.cpp Path.
Select a .gguf model (via Browse, History, or Drag & Drop).
Click ▶ Start to launch the server.

📁 Folder Structure (Auto-created)

LlamaLauncher/
├── LlamaLaunchers.exe
├── presets/
│ ├── models-cache.txt # Cached GGUF metadata
│ ├── models/ # Per-model presets (.ini)
│ └── gen/ # Generation presets (.ini)
├── lang/ # Language translation files (.ini)
└── logs/ # Runtime logs (.log)

🖥️ Usage Guide

🔹 Model Selection & Scanning

Use the 📂 Browse button or drag & drop a .gguf file onto the window.
The app automatically validates shards (-00001-of-00003.gguf pattern), aggregates sizes, and extracts GGUF metadata.
Scanning is multi-threaded with a real-time progress bar.

🔹 Server Configuration

Performance: Threads, Context size, Device Mode (Auto/CPU/GPU), CPU/GPU layer offloading.
Generation: Temperature, Top-K, Top-P, Min-P, Repeat/Presence Penalty, Seed, Reasoning Mode.
Advanced: Flash Attention, KV Cache Types (fp16, q8_0, etc.), Rope Scaling, Batch/Ubatch sizes, Tools/Props flags, Extra CLI arguments.

🔹 Presets

Model Presets: Automatically saved when launching a model. Load via File → Load Preset.
Generation Presets: Manage sampling profiles via the 🧠 Gen Presets combobox. Save/Overwrite/Delete directly from the UI.

🔹 Tray & Multi-Instance

Minimizing hides the window to the system tray. Restore via tray icon → Restore.
The app enforces a single instance using a unique mutex. Opening a second instance brings the first to focus.

🌍 Language Support

Translations are stored in lang/.ini. The app auto-detects available languages and falls back to EN or FR if a translation is missing.

To add a new language:

Create lang/.ini in the app directory.
Follow the existing INI structure ([FormName.ComponentName] sections with Caption and Hint keys).
Restart the app.

🐛 Troubleshooting

Issue	Solution
`llama-server.exe` fails to launch or shows DLL errors	Install the Visual C++ 2015-2022 Redistributable (x64). Specifically, `VC14_redist.x64-14.50.35719.exe` (or a compatible version) is required for `llama-server.exe` to run correctly. Download it from Microsoft's official site if missing.
`GGUF not detected` or missing shards	Ensure all shard files (`-00001-of-NNNNN.gguf`) are present in the same directory. The scanner validates completeness.
Server fails to start	Verify `llama-server.exe` path, check `logs/llama_launcher_log_*.log` for CLI errors, and ensure required dependencies (CUDA/ROCm) are installed.
Cache outdated	Delete `presets\models-cache.txt` or click `🔄 Update` in the model info panel.
UI language missing	Ensure `lang/.ini exists. The app falls back to English/French automatically.`
High memory usage during scan	The scanner is multi-threaded but reads sequentially. Use `Fast Scan` mode in settings to skip tensor info parsing.

📜 License

This project is distributed under the MIT License. See LICENSE for details. llama.cpp and its ecosystem are developed by Georgi Gerganov and the open-source community under the Apache 2.0 License.

🤝 Credits & Acknowledgments

🧠 [llama.cpp](https://github.com/ggml-org/llama.cpp) - The core inference engine powering this launcher.

Built with ❤️ for the local AI community. Run powerful models locally, effortlessly.