audio-processing, laravel-support, skript, readme

scripts/start.sh

@@ -5,7 +5,7 @@
 #   ./scripts/start.sh --no-preload    nur starten
 #   ./scripts/start.sh --model glm-4.7-flash   anderes Chat-Modell vorladen
 #   ./scripts/start.sh --show-key      API-Key im Klartext ausgeben
-#   ./scripts/start.sh --pull          Images vorher aktualisieren
+#   ./scripts/start.sh --pull          Basis-Images aktualisieren und neu bauen
 set -euo pipefail
 
 cd "$(dirname "$(readlink -f "$0")")/.."
@@ -88,8 +88,12 @@ free_gb=$(df -BG --output=avail models | tail -1 | tr -dc '0-9')
 
 # ------------------------------------------------------------------- Start
 step "Starte Container"
-(( PULL )) && docker compose pull
-docker compose up -d --remove-orphans
+if (( PULL )); then
+  docker compose pull --ignore-buildable
+  docker compose build --pull
+fi
+# baut das llama-swap-Image (ROCm + whisper/sd-server), falls es noch fehlt
+docker compose up -d --build --remove-orphans
 
 step "Warte auf Bereitschaft"
 for i in $(seq 1 60); do
@@ -112,16 +116,24 @@ ok "Caddy bereit (Auth aktiv)"
 # ---------------------------------------------------------------- Preload
 if (( PRELOAD )); then
   step "Lade Modelle vor (erster Start lädt ggf. von Hugging Face herunter)"
-  preload() {  # $1 = Modell, $2 = Endpoint, $3 = JSON-Body
-    local t0=$SECONDS code
-    printf '  %-22s … ' "$1"
-    code=$(curl -s -o /dev/null -m 3600 -w '%{http_code}' -H 'Content-Type: application/json' \
-      "http://127.0.0.1:$BACKEND_PORT$2" -d "$3" || true)
+  preload() {  # $1 = Modell, $2 = Endpoint, Rest = curl-Argumente für den Body
+    local m=$1 ep=$2 t0=$SECONDS code; shift 2
+    printf '  %-22s … ' "$m"
+    code=$(curl -s -o /dev/null -m 3600 -w '%{http_code}' "http://127.0.0.1:$BACKEND_PORT$ep" "$@" || true)
     if [[ "$code" == "200" ]]; then echo "${G}ok${N} ($((SECONDS - t0)) s)"; else echo "${R}HTTP $code${N}"; fi
   }
-  preload qwen2.5-coder-1.5b /v1/completions '{"model":"qwen2.5-coder-1.5b","prompt":"x","max_tokens":1}'
-  preload nomic-embed        /v1/embeddings  '{"model":"nomic-embed","input":"x"}'
-  preload "$MODEL"           /v1/chat/completions "{\"model\":\"$MODEL\",\"messages\":[{\"role\":\"user\",\"content\":\"hi\"}],\"max_tokens\":1}"
+  json=(-H 'Content-Type: application/json' -d)
+  preload qwen2.5-coder-1.5b /v1/completions "${json[@]}" '{"model":"qwen2.5-coder-1.5b","prompt":"x","max_tokens":1}'
+  preload nomic-embed        /v1/embeddings  "${json[@]}" '{"model":"nomic-embed","input":"x"}'
+  if [[ -f models/whisper/ggml-large-v3-turbo-q8_0.bin ]]; then
+    wav=$(mktemp --suffix=.wav)
+    python3 -c 'import wave,sys; w=wave.open(sys.argv[1],"wb"); w.setnchannels(1); w.setsampwidth(2); w.setframerate(16000); w.writeframes(b"\0\0"*16000)' "$wav"
+    preload whisper /v1/audio/transcriptions -F model=whisper -F "file=@$wav"
+    rm -f "$wav"
+  else
+    warn "Whisper-Modell fehlt (models/whisper/) – siehe README"
+  fi
+  preload "$MODEL"           /v1/chat/completions "${json[@]}" "{\"model\":\"$MODEL\",\"messages\":[{\"role\":\"user\",\"content\":\"hi\"}],\"max_tokens\":1}"
 fi
 
 # ------------------------------------------------------------------ Status