dschlueter
4a31535b76
- Neue Befehle: /plan (Planungsmodus, nur PLAN.md), /cancel (Loop-Abbruch), /continue (Resume nach Unterbrechung), /discard (PLAN.md verwerfen) - contextWindow in models.json und llama.cpp-Servern: 131072 → 262144 - KV-Cache: q8_0 → q4_0 (weniger VRAM, passt zu 262k-Kontext auf 2× 3090) - parallel: 2 → 1 beim Coder (stabiler bei großem Kontext) - Optimize-Status mit ASCII-Fortschrittsbalken + Blocker-Preview - cancelRequested-Flag prüft nach jedem Loop-Schritt Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
90 lines
2.4 KiB
Bash
Executable file
90 lines
2.4 KiB
Bash
Executable file
#!/usr/bin/env bash
|
|
set -euo pipefail
|
|
|
|
HF_HOME="${HF_HOME:-/home/dschlueter/nvme2n1p7_home/huggingface}"
|
|
MODEL_REL_PATH="models/qwen3/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-IQ4_XS.gguf"
|
|
IMAGE="ghcr.io/ggml-org/llama.cpp:server-cuda"
|
|
CONTAINER_NAME="qwen36-27b-coder"
|
|
HOST_PORT=8001
|
|
CONTAINER_PORT=8000
|
|
MODEL_ALIAS="qwen3.5-coder"
|
|
|
|
echo "[*] Verwende HF_HOME = $HF_HOME"
|
|
if [ ! -f "$HF_HOME/$MODEL_REL_PATH" ]; then
|
|
echo "[!] Modell-Datei nicht gefunden: $HF_HOME/$MODEL_REL_PATH" >&2
|
|
exit 1
|
|
fi
|
|
|
|
if docker ps -a --format '{{.Names}}' | grep -q "^${CONTAINER_NAME}\$"; then
|
|
echo "[*] Stoppe existierenden Container $CONTAINER_NAME ..."
|
|
docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
|
|
fi
|
|
|
|
echo "[*] Starte llama.cpp-Server für Coder ..."
|
|
docker run -d \
|
|
--gpus '"device=1,2"' \
|
|
--name "$CONTAINER_NAME" \
|
|
--restart unless-stopped \
|
|
-e HF_HOME="/hf_home" \
|
|
-v "$HF_HOME:/hf_home:ro" \
|
|
-p "${HOST_PORT}:${CONTAINER_PORT}" \
|
|
"$IMAGE" \
|
|
-m "/hf_home/${MODEL_REL_PATH}" \
|
|
--alias "${MODEL_ALIAS}" \
|
|
-c 262144 \
|
|
-n 16384 \
|
|
--jinja \
|
|
--no-context-shift \
|
|
--temp 0.2 \
|
|
--top-p 0.95 \
|
|
--top-k 40 \
|
|
--min-p 0.01 \
|
|
--repeat-penalty 1.05 \
|
|
--main-gpu 0 \
|
|
--tensor-split 0.5,0.5 \
|
|
-ngl 999 \
|
|
-fa on \
|
|
--kv-unified \
|
|
--cache-type-k q4_0 \
|
|
--cache-type-v q4_0 \
|
|
--batch-size 1024 \
|
|
--ubatch-size 512 \
|
|
--parallel 1 \
|
|
--cont-batching \
|
|
--host 0.0.0.0 \
|
|
--port "$CONTAINER_PORT"
|
|
|
|
echo "[*] Warte auf HTTP ..."
|
|
HTTP_READY=0
|
|
for i in {1..90}; do
|
|
if curl -s "http://localhost:${HOST_PORT}/health" >/dev/null 2>&1 || \
|
|
curl -s "http://localhost:${HOST_PORT}/v1/models" >/dev/null 2>&1; then
|
|
HTTP_READY=1
|
|
break
|
|
fi
|
|
sleep 2
|
|
done
|
|
|
|
if [ "$HTTP_READY" -ne 1 ]; then
|
|
echo "[!] HTTP-Server wurde nicht rechtzeitig erreichbar." >&2
|
|
docker logs --tail 200 "$CONTAINER_NAME" || true
|
|
exit 1
|
|
fi
|
|
|
|
echo "[*] Teste Chat-Completion ..."
|
|
curl -s -X POST "http://localhost:${HOST_PORT}/v1/chat/completions" \
|
|
-H "Content-Type: application/json" \
|
|
-d "{
|
|
\"model\": \"${MODEL_ALIAS}\",
|
|
\"messages\": [
|
|
{ \"role\": \"system\", \"content\": \"Du bist ein präziser Coding-Assistent.\" },
|
|
{ \"role\": \"user\", \"content\": \"Antworte nur mit dem Wort: bereit\" }
|
|
],
|
|
\"max_tokens\": 8,
|
|
\"temperature\": 0.0,
|
|
\"stream\": false
|
|
}"
|
|
|
|
echo
|
|
echo "[*] Server bereit auf http://0.0.0.0:${HOST_PORT}"
|
|
echo "[*] Stoppen mit: docker rm -f ${CONTAINER_NAME}"
|