llm-server-profiles

Helper scripts and profiles for running llama.cpp server configs, plus an optional OpenCode skill.

What you get

• bin/llm-server: wrapper that loads a profile from ~/.config/llm-server/<name>.conf
• profiles/example.conf: starter profile
• skills/llm-server-profiles/SKILL.md: OpenCode skill for managing profiles

Install

./install.sh

The install script:

• Symlinks bin/llm-server into ~/.local/bin/llm-server
• Copies profiles/example.conf into ~/.config/llm-server (if missing)
• Symlinks the OpenCode skill into ~/.config/opencode/skills/llm-server-profiles (if present)

Usage

List profiles:

llm-server --list-profiles

Run with a profile:

llm-server -P s

If you see an error about --device, list available devices and update your profile:

llm-server --list-devices

Profile format

Profile files are bash snippets. Supported fields:

• MODEL_NAME (string)
• MODEL (string, path to gguf)
• USE_CASE (string)
• CTX_SIZE (number)
• LLMSERVER_ARGS (bash array of flags)

--device expects the exact ID shown by llm-server --list-devices (for example ROCm0), not always a numeric 0.

Example:

MODEL_NAME="Qwen3 Coder Next Q4_K_XL"
MODEL="/path/to/Qwen3-Coder-Next-Q4_K_XL.gguf"
USE_CASE="Speed profile for codebases under ~50k LOC"
CTX_SIZE=8192
LLMSERVER_ARGS=(
  # --device <id>  # use `llm-server --list-devices` (example: ROCm0)
  --gpu-layers auto
  --fit on
  --batch-size 2048
  --ubatch-size 512
  --threads 24
  --threads-batch 24
  --temp 0.2
  --top-p 0.9
  --top-k 40
  --repeat-penalty 1.05
)

Environment overrides

• LLM_SERVER_BIN: server binary to run (default: llama-server)
• LLM_SERVER_CONFIG_DIR: profile directory (default: ~/.config/llm-server)
• LLM_SERVER_MODEL: default model path if not set in profile
• LLM_SERVER_PORT: default port (default: 8080)
• LLM_SERVER_CTX_SIZE: default ctx size (default: 0, uses model default)

OpenCode skill

If you use OpenCode, the skill is in skills/llm-server-profiles/SKILL.md and is installed by install.sh.