feat: GPU-Split, Phase-Timer, llama.cpp-Update-Fixes

GPU-Setup:
- Coder → GPU 1 (device=1), Judge → GPU 2 (device=2), kein --tensor-split
- --chat-template-kwargs deprecated → --reasoning on (neue llama.cpp-API)
- Smoke-Test-Timeout 180s → 300s (neue fitting-Phase beim Start)

Fortschrittsanzeige:
- Phase-Timer [MM:SS] in Statuszeile während LLM-Inference (sendAndWait)
- currentActivity in Fix-Phase zeigt konkreten Blocker statt generischem Text

Dokumentation:
- README: GPU-Tabelle, VRAM-Abschätzung, --reasoning on, Timer-Beispiele
- BEDIENUNGSANLEITUNG: Ladezeit, Fortschrittsanzeige mit Timer-Beispielen

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Dieter Schlüter 2026-06-15 00:19:23 +02:00
commit 3f7ffe60c2
5 changed files with 122 additions and 72 deletions

View file

@ -22,7 +22,7 @@ fi
echo "[*] Starte llama.cpp-Server für Judge ..."
docker run -d \
--gpus '"device=1,2"' \
--gpus '"device=2"' \
--name "$CONTAINER_NAME" \
--restart unless-stopped \
-e HF_HOME="/hf_home" \
@ -34,7 +34,7 @@ docker run -d \
-c 262144 \
-n 16384 \
--jinja \
--chat-template-kwargs '{"enable_thinking":true}' \
--reasoning on \
--no-context-shift \
--temp 0.7 \
--top-p 0.80 \
@ -42,7 +42,6 @@ docker run -d \
--min-p 0.01 \
--repeat-penalty 1.05 \
--main-gpu 0 \
--tensor-split 0.5,0.5 \
-ngl 999 \
-fa on \
--kv-unified \
@ -57,7 +56,7 @@ docker run -d \
echo "[*] Warte auf Modell-Bereitschaft (Completion-Check, max. 180 s) ..."
MODEL_READY=0
for i in {1..90}; do
for i in {1..150}; do
HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" --max-time 10 \
-X POST "http://localhost:${HOST_PORT}/v1/chat/completions" \
-H "Content-Type: application/json" \