feat(local-llm): B0.5 — llama-swap + lokale Modellauswahl in Diagnostic

Mehrere lokale Modelle, on-demand geladen/geswappt, in Diagnostic waehlbar.
Design: das Brain schickt den Modellnamen (aus local_llm.json) im llm_request
mit -> Adapter -> llama-swap laedt/swappt. Keine separate Gamebox-Config noetig.

- xtts: `llama`-Container -> `llama-swap` (unified-cuda), config.yaml mit
  qwen3-8b (Standard) + qwen3-4b; Auto-Download via -hf, Cache /models geteilt
  (qwen3-8b schon da). Adapter -> llama-swap:8080, Timeout 600s (Erst-Download).
- adapter: `model` aus dem Request an llama-swap durchreichen (Fallback env).
- brain: router.load_config liest localLlmModel; local_llm_chat(model=...);
  agent gibt cfg-Modell mit; bridge reicht model durch (_local_llm + Route).
- diagnostic: /api/local-models-list (aus /shared/config/local_models.json,
  seeded), local-llm-config um localLlmModel erweitert; Dropdown "Lokales
  Modell" im Settings-Block + Erst-Download-Hinweis.

BLIND gebaut (Gamebox nicht testbar hier): llama-swap CLI/Config-Pfad beim
ersten Start via `docker logs aria-llama-swap` pruefen. Live-Lade-Status
(Adapter->Diagnostic) ist B0.5-2 (Folgeschritt).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-07-11 14:11:59 +02:00
co-authored by Claude Opus 4.8
parent 3ddcf665f0
commit 8ae20a9bd8
9 changed files with 174 additions and 46 deletions
+20 -24
View File
@@ -90,22 +90,19 @@ services:
# Container-Restarts.
restart: unless-stopped
# ─── Lokales LLM (Plan B, B0) — llama.cpp-Server (GPU) ────────
# Serviert Qwen3-8B (GGUF Q4_K_M) OpenAI-kompatibel auf :8081, NUR im
# Compose-Netz (kein RVS direkt) — die Bruecke macht der llm-adapter.
# ─── Lokales LLM (Plan B, B0.5) — llama-swap (GPU) ────────────
# llama-swap laedt/swappt mehrere Modelle on-demand (nur eins passt gleich-
# zeitig in die 12 GB). Welches geladen wird, bestimmt das `model`-Feld im
# Request — das Brain schickt es aus local_llm.json mit. Erster Load eines
# Modells zieht das GGUF via -hf von HF (Cache unter /models, persistent).
# OpenAI-kompatibel auf :8080, nur im Compose-Netz; die Bruecke macht der
# llm-adapter. Modell-Liste: ./llama-swap/config.yaml.
#
# AUTO-DOWNLOAD: llama.cpp zieht das GGUF beim ersten Start selbst von
# Hugging Face (-hf <repo>:<quant>) und cached es unter /models (persistent
# via Bind-Mount -> kein Re-Download bei Restart). Kein manuelles Ablegen
# noetig. Modell wechseln = LLM_HF_REPO/LLM_HF_QUANT in der .env aendern +
# Container neu. (Alternativ lokale Datei: command auf -m /models/x.gguf.)
#
# VRAM auf der RTX 3060 (12 GB): whisper-small (~1-2) + f5tts (~1-2) +
# qwen3-8b-q4 (~6) ~= 9-10 GB. Passt, aber knapp — bei OOM: LLM_CTX kleiner
# oder Quant auf Q4_K_S/IQ4_XS wechseln.
llama:
image: ghcr.io/ggml-org/llama.cpp:server-cuda
container_name: aria-llama
# BLIND GEBAUT (kein Gamebox-Test hier): beim ersten Start
# `docker logs -f aria-llama-swap` pruefen. Image bundelt llama-server.
llama-swap:
image: ghcr.io/mostlygeek/llama-swap:unified-cuda
container_name: aria-llama-swap
deploy:
resources:
reservations:
@@ -114,13 +111,11 @@ services:
count: 1
capabilities: [gpu]
volumes:
- ./models:/models # HF-Download-Cache (persistent)
- ./models:/models # HF-Download-Cache (persistent)
- ./llama-swap/config.yaml:/app/config.yaml:ro # Modell-Liste
environment:
- LLAMA_CACHE=/models # llama.cpp legt -hf-Downloads hier ab
command: >
-hf ${LLM_HF_REPO:-Qwen/Qwen3-8B-GGUF}:${LLM_HF_QUANT:-Q4_K_M}
--host 0.0.0.0 --port 8081
-ngl 99 -c ${LLM_CTX:-8192} --jinja
- LLAMA_CACHE=/models # llama-server legt -hf-Downloads hier ab
command: ["--config", "/app/config.yaml", "--listen", "0.0.0.0:8080"]
restart: unless-stopped
# ─── Local-LLM-Adapter — RVS <-> llama.cpp (Plan B, B0) ──────
@@ -130,14 +125,15 @@ services:
build: ./llm-adapter
container_name: aria-llm-adapter
depends_on:
- llama
- llama-swap
environment:
- RVS_HOST=${RVS_HOST}
- RVS_PORT=${RVS_PORT:-443}
- RVS_TLS=${RVS_TLS:-true}
- RVS_TLS_FALLBACK=${RVS_TLS_FALLBACK:-true}
- RVS_TOKEN=${RVS_TOKEN}
- LLAMA_URL=http://llama:8081
- LLAMA_URL=http://llama-swap:8080
- LLM_MODEL=${LLM_MODEL:-qwen3-8b}
- LLM_TIMEOUT_SEC=${LLM_TIMEOUT_SEC:-60}
# Erster Load eines Modells kann ein GGUF ziehen (mehrere GB) — grosszuegig.
- LLM_TIMEOUT_SEC=${LLM_TIMEOUT_SEC:-600}
restart: unless-stopped