Cross-platform wrapper scripts for llama-server that simplify running .gguf models with intelligent configuration management.

Scripts: llms (UNIX/WSL) and llms.ps1 (PowerShell).

• Optimal Settings Retention: Automatically saves and loads optimal settings for each model.
• Flexible Discovery: Discover and run models using partial names.
• Priority-based Config: Seamlessly handles CLI arguments, environment variables, and config files.
• Dry-run Capability: Preview commands without execution.

1. Configure Model Directories: Create ~/.config/llms.ini (or %APPDATA%\llms.ini on Windows):

   ModelsDirs = /path/to/your/models

2. Run a Model:

   llms Mistral 64000  # First run: specify context size (saved automatically)
   llms Mistral        # Subsequent runs: settings remembered

3. List Models: llms list

llms <partial_name> [<context_size>] [llama-server args...] [--dry-run]

• Partial Name: Case-insensitive match against .gguf files.
• Context Size: Required for first run, optional thereafter.
• Arguments: Any llama-server flag (e.g., --mlock, --n-gpu-layers 30).
• Dry Run: Use --dry-run to preview the command without executing or saving config.

Companion .mmproj files (for multi-modal models) are detected and loaded automatically if they follow a specific naming convention.

Naming Convention: The companion file must be named {BaseName}.mmproj{Suffix}.gguf, where {BaseName} is a prefix of the main model's filename.

Example:

• Main Model: Qwen3-VL-30B-A3B-Thinking-UD-Q4.gguf
• Companion: Qwen3-VL-30B-A3B-Thinking-UD.mmproj-F16.gguf

The script will automatically find the companion file because Qwen3-VL-30B-A3B-Thinking-UD is the start of the main model's name.

• Persistent: llms Mistral --n-gpu-layers 30 (saves to .ini)
• Temporary (Global): LLMS_PORT=9090 llms Mistral (ENV only)

Note: Since version 1.3.0, both scripts ignore per-model ENV variables (e.g., LLMS_CTX_SIZE) to favor explicit CLI/config settings.

1. Script directory: ./llms.ini
2. User config: ~/.config/llms.ini (UNIX) or %USERPROFILE%\AppData\Local\llms.ini (Windows)

• Persisted (Per-Model): --mlock, --no-mmap, --jinja, --cont-batching, etc.
• Transient (Server-Wide): --no-webui, --verbose, --dry-run, --help.

Run Pester tests for PowerShell:

Invoke-Pester ./llms.tests.ps1

• "No model file found": Check ModelsDirs in llms.ini or run llms list.
• "Specify <context_size>": Model has no saved config yet. Run once with a numeric size.
• "llama-server not found": Ensure llama-server is in your system $PATH.