Provide an optimized systemd service template with dynamic per-model GGUF optimization

オープン
#194 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

まだ誰も着手していません。

評価

難易度
2/5
見積もり時間
1〜3時間
初心者へのやさしさ
35/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
活発
技術スタック
bash
領域
devops

調査の方向性

The proposal is a systemd unit whose ExecStartPre bash script generates a preset .ini for llama-server --models-preset; check the repo's server docs and examples to confirm the preset keys and flags it uses. Before any file is added, the maintainer needs to agree on whether this belongs in the examples folder, and the script's gguf-parser/jq field names and cache-type choices need verifying. Done means a reviewed unit file and a README note that run cleanly on a sample models directory.

索引モデルが issue の本文から書いたものです。

説明

enhancement
Prerequisites
  • I am running the latest code. Mention the version if possible as well.
  • I carefully followed the README.md.
  • I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
  • I reviewed the Discussions, and have a new and useful enhancement to share.
Feature Description

Hello! I would like to share and propose a production-ready systemd service file for beellama.cpp that automatically generates optimal presets and handles dynamic per-model configurations at startup.

This configuration uses gguf-parser and jq to parse local models dynamically, matching them with multimodal projectors (mmproj) and configuring layers, context sizes, and KV cache types automatically.
Features of this Systemd Unit:

Inline Dynamic Preset Generation: Automatically scans the models directory before starting the server and writes an optimized configuration file (beellama_auto.ini).

Smart Architecture Detection: Extracts block counts (gpu-layers), maximum context lengths, and architecture types using gguf-parser and jq.

Automatic Multimodal Mapping: Automatically detects and pairs mmproj files with their corresponding base models based on filename matching.

LoRA & KV Cache Tuning: Automatically adjusts cache types (e.g., switches to q4_0 if LoRA ranks are detected, otherwise uses kvarn4).

Proposed Systemd Unit File:

[Unit]
Description=beellama.cpp Server with Dynamic Per-Model Optimization
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
Environment="MODELS_DIR=/run/media/user/Data/LLM"
Environment="PRESET_FILE=/etc/conf.d/beellama_auto.ini"

ExecStartPre=/bin/bash -c ' \
    set -Eeuo pipefail; \
    echo "=== beellama: inline dynamic preset generator ==="; \
    echo "# Auto-generated optimal preset" > "$PRESET_FILE"; \
    echo "" >> "$PRESET_FILE"; \
    echo "[*]" >> "$PRESET_FILE"; \
    echo "threads = 8" >> "$PRESET_FILE"; \
    echo "threads-batch = 8" >> "$PRESET_FILE"; \
    echo "flash-attn = on" >> "$PRESET_FILE"; \
    echo "" >> "$PRESET_FILE"; \
    while IFS= read -r -d "" MODEL_FILE; do \
        MODEL_NAME=$(basename "$MODEL_FILE" .gguf); \
        MODEL_DIR=$(dirname "$MODEL_FILE"); \
        MMPROJ_FILE=""; \
        CANDIDATE=$(find "$MODEL_DIR" -type f -iname "*mmproj*.gguf" -print -quit || true); \
        if [ -n "$CANDIDATE" ]; then \
            CLEAN_MODEL_NAME=$(echo "$MODEL_NAME" | sed "s/-Q[0-9].*//i" | sed "s/-IQ.*//i"); \
            if echo "$CANDIDATE" | grep -qi "$CLEAN_MODEL_NAME"; then \
                MMPROJ_FILE="$CANDIDATE"; \
            fi; \
        fi; \
        LAYERS="auto"; \
        CTX_SIZE="32768"; \
        CACHE_TYPE="kvarn4"; \
        ARCH_DETECTED="unknown"; \
        if command -v gguf-parser >/dev/null 2>&1 && command -v jq >/dev/null 2>&1; then \
            JSON_OUT=$(gguf-parser --model "$MODEL_FILE" --json 2>/dev/null || true); \
            if [ -n "$JSON_OUT" ]; then \
                DETECTED_LAYERS=$(echo "$JSON_OUT" | jq -r ".architecture.blockCount // empty" 2>/dev/null || true); \
                if [ -n "$DETECTED_LAYERS" ] && [ "$DETECTED_LAYERS" != "null" ]; then \
                    LAYERS="$DETECTED_LAYERS"; \
                fi; \
                DETECTED_CTX=$(echo "$JSON_OUT" | jq -r ".architecture.maximumContextLength // empty" 2>/dev/null || true); \
                if [ -n "$DETECTED_CTX" ] && [ "$DETECTED_CTX" != "null" ]; then \
                    if [ "$DETECTED_CTX" -lt 32768 ]; then \
                        CTX_SIZE="$DETECTED_CTX"; \
                    else \
                        CTX_SIZE="32768"; \
                    fi; \
                fi; \
                ARCH_DETECTED=$(echo "$JSON_OUT" | jq -r ".architecture.architecture // \"unknown\"" 2>/dev/null || true); \
                LORARANK=$(echo "$JSON_OUT" | jq -r ".architecture.attentionKeyValueLORARank // 0" 2>/dev/null || true); \
                if [ "$LORARANK" -gt 0 ] 2>/dev/null; then \
                    CACHE_TYPE="q4_0"; \
                fi; \
            fi; \
        fi; \
        echo "Model: $MODEL_NAME [Arch: $ARCH_DETECTED] -> Layers: $LAYERS, Ctx: $CTX_SIZE, Cache: $CACHE_TYPE"; \
        echo "[$MODEL_NAME]" >> "$PRESET_FILE"; \
        echo "model = $MODEL_FILE" >> "$PRESET_FILE"; \
        echo "gpu-layers = $LAYERS" >> "$PRESET_FILE"; \
        echo "ctx-size = $CTX_SIZE" >> "$PRESET_FILE"; \
        echo "cache-type-k = $CACHE_TYPE" >> "$PRESET_FILE"; \
        echo "cache-type-v = $CACHE_TYPE" >> "$PRESET_FILE"; \
        if [ -n "$MMPROJ_FILE" ]; then \
            echo "mmproj = $MMPROJ_FILE" >> "$PRESET_FILE"; \
        fi; \
        echo "" >> "$PRESET_FILE"; \
    done < <(find "$MODELS_DIR" -type f -name "*.gguf" ! -iname "*mmproj*" -print0); \
    echo "Preset generation complete."'
    
ExecStart=/usr/bin/llama-server --models-preset /etc/conf.d/beellama_auto.ini --host 0.0.0.0 --port 8080 --models-max 1 --cache-reuse 1024 --cache-ram 16384
ExecReload=/bin/kill -s HUP $MAINPID
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target

Feel free to adapt or integrate this into the documentation or examples folder if others might find it useful!

Motivation

Key benefits for users:

  • Zero-Touch Configuration: Automatically scans the designated models directory, eliminating the need to manually write or update .ini preset files whenever new models are added or removed.
  • Smart Resource Optimization: Leverages gguf-parser and jq to dynamically extract hardware-friendly metrics per model (such as exact architecture block counts for optimal GPU layer offloading and context limits).
  • Seamless Multimodal Support: Automatically detects and pairs mmproj files with their respective base models based on filename structure.
  • Production Reliability: Standardizes deployment using standard systemd patterns (Restart=on-failure, graceful HUP reloading, and proper environment separation) making it ideal for self-hosted LLM setups.
Possible Implementation

No response

主要言語
C++
スター
1.2k
フォーク
80
平均マージ
4日 20時間
マージ済み PR(30日)
2

環境構築

このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

Anbeeld/beellama.cpp のほかの issue

Anbeeld/beellama.cpp の issue をすべて見る

似ている issue

C++ の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。