Add: CodeNeedle Benchmark. Optimize model handling.

This commit is contained in:
2026-06-17 19:14:08 +02:00
parent 599822036f
commit ce64ef56e4
11 changed files with 799 additions and 57 deletions
+2 -1
View File
@@ -3,7 +3,8 @@
# ==========================================
# --- LLM ENGINE & MODEL SELECTION ---
MODEL_LIST=qwen2.5-coder:14b,qwen3:14b,deepseek-coder-v2:16b
MODEL_LIST=gemma4:e4b,qwen2.5-coder:14b,qwen3:14b,deepseek-coder-v2:16b
OPENCLAW_DEFAULT_MODEL=ollama/qwen3-general
OLLAMA_KEEP_ALIVE=-1
# KV cache quantization for large contexts (recommended on desktop GPUs): q8_0 or q4_0
# OLLAMA_KV_CACHE_TYPE=q8_0
+34 -1
View File
@@ -48,4 +48,37 @@ Wir entwickeln eine lokale, containerisierte AI-Workspace-Infrastruktur namens *
1. **OpenClaw Integration:** Die `openclaw.json` muss so konfiguriert werden, dass der Agent den `ghostnet-router` (SSE-Endpunkt) als primäre Tool-Quelle nutzt.
2. **Modell-Management:** DNS-Problematik in der Ollama-Container-Registry beheben (Registry-Auflösung schlägt aktuell fehl).
3. **Skill Expansion:** Entwicklung weiterer MCP-Bridges (z.B. für lokales Filesystem-Management oder Datenbank-Abfragen), die einfach am Router "angepluggt" werden können.
4. **UX-Automation:** Ein `ghostnet.sh` Wrapper, der die Vendor-Erkennung automatisiert und den Stack konsistent hochfährt.
4. **UX-Automation:** Ein `ghostnet.sh` Wrapper, der die Vendor-Erkennung automatisiert und den Stack konsistent hochfährt.
# Stand: 17.06.2026
## 📋 Handover & Context Prompt for the Next AI Agent
**Context:**
Das Repo hat inzwischen mehrere betriebliche Verbesserungen erhalten, die nicht nur die Laufzeit des Stacks, sondern auch Modellprofilierung und Benchmarking betreffen. OpenClaw laeuft weiterhin containerisiert gegen den lokalen Ollama-Stack und den MCP-Router, aber Konfiguration und Benchmark-Ausgaben sind jetzt deutlich sauberer organisiert.
**Current Architecture & Operational Notes:**
1. **Default-Modellsteuerung:** `agents.defaults.model` in `openclaw/config/openclaw.json` nutzt `${OPENCLAW_DEFAULT_MODEL}`. OpenClaw expandiert `${VAR}` direkt in Config-Strings zur Laufzeit. Compose setzt deshalb einen Fallback, damit fehlende Variablen den Config-Load nicht brechen.
2. **Default-Modellprofil:** Das generelle OpenClaw-Defaultprofil ist jetzt `ollama/qwen3-general`. Das bestehende `qwen3-coding`-Profil bleibt fuer codinglastige Workflows erhalten.
3. **Custom-Modelfiles:** Ollama baut alle Dateien unter `ollama/modelfiles/*.modelfile` automatisch beim Containerstart; der Modellname entspricht jeweils dem Dateinamen ohne Endung.
4. **Benchmark-Artefakte:** Laufzeitdaten liegen unter `.cache/benchmarks/existing-suite/runs/$RUN_ID`, waehrend kuratierte Ergebnisse unter `docs/benchmarks/existing-suite/$RUN_ID` landen.
5. **Benchmark-Reports:** Per-Model werden `report.html` und `RESULTS.md` erzeugt, fuer Suites zusaetzlich `SUITE_COMPARISON.html`. Die Erzeugung laeuft ueber `scripts/lib/report-generator.sh`.
**Benchmarking State:**
* `scripts/benchmark-stack.sh` und `scripts/benchmark-model-suite.sh` verwenden jetzt 600s Timeout fuer robustere lokale Modelllaeufe.
* `scripts/benchmark-codeneedle-suite.sh` bleibt bei 600s und dient als separater Recall-Benchmark gegen denselben Ollama-Endpunkt.
* Fuer stabile Vergleichslaeufe ist die Ollama-Konfiguration auf `OLLAMA_KEEP_ALIVE=-1`, `OLLAMA_KV_CACHE_TYPE=q8_0`, `OLLAMA_MAX_LOADED_MODELS=1`, `OLLAMA_NUM_PARALLEL=1` und `SHM_SIZE=16gb` ausgerichtet.
* Die Benchmark-Skripte trennen jetzt konsequent zwischen Temp-Artefakten und finaler Ergebnisdoku, damit `docs/` nicht mehr mit Rohlogs zugemuellt wird.
**Current Repository State:**
* **Model Routing:** OpenClaw-Defaultmodell ist per `.env` und Compose ueberschreibbar, ohne JSON manuell anzupassen.
* **Profiles:** `qwen3-general` ist das allgemeine Defaultprofil fuer Chats; `qwen3-coding`, `qwen25-coding` und `deepseek-coding` bleiben spezialisierte Profile.
* **Benchmark Presentation:** Benchmark-Ergebnisse stehen jetzt als HTML-Dashboard, Markdown-Lab-Notiz und Suite-Vergleich bereit.
* **Operational Validation:** Compose-Render und OpenClaw-Config-Expansion fuer `OPENCLAW_DEFAULT_MODEL` wurden direkt gegen das Upstream-Image verifiziert.
**Next Tasks / Where to Continue:**
1. **DNS/Ollama Registry:** Die externe Modellauflosung fuer Registry-Pulls im Containernetz weiter haerten.
2. **Additional Profiles:** Weitere nicht-codinglastige oder domänenspezifische Modelfiles auf Basis bestehender Profile ergänzen.
3. **Benchmark Evolution:** Adaptive Timeouts, Warm-up-Runs und historische Vergleichsreports sind sinnvolle naechste Ausbaustufen.
4. **OpenClaw Tooling:** Weitere MCP-Bridges und agentenspezifische Tool-Profile sauber in die bestehende Router-Struktur integrieren.
+63 -5
View File
@@ -10,6 +10,7 @@ Clone the repository and prepare your environment configuration:
`cp .env.example .env`
Open `.env` and set your `GPU_TYPE` (amd, nvidia, or intel) and your desired models.
Set `OPENCLAW_DEFAULT_MODEL` to the model OpenClaw should use by default.
### 2. Boot the Stack
Use the integrated management script, which automatically selects the correct vendor-specific compose file:
@@ -45,6 +46,7 @@ Use `ollama list` inside the container or on the host to see the exact model nam
Recommended custom profiles for this stack:
```text
qwen3-general -> Qwen3 14B, universal profile for OpenClaw chats
qwen3-coding -> Qwen3 14B, balanced default for agentic coding
qwen25-coding -> Qwen2.5-Coder 14B, code-first edit and repair profile
deepseek-coding -> DeepSeek-Coder-V2 16B, long-context repo analysis profile
@@ -78,15 +80,71 @@ If you change the host binding, use `TOOL_ROUTER_BIND` and `TOOL_ROUTER_HOST_POR
## Benchmarking Custom Models
The repository includes a reproducible benchmark script for comparing local models through the OpenClaw agent path.
The repository now ships two complementary benchmark entry points:
* `./scripts/benchmark-stack.sh` and `./scripts/benchmark-model-suite.sh` exercise the full OpenClaw agent path.
* `./scripts/benchmark-codeneedle-suite.sh` runs the upstream CodeNeedle recall benchmark against the same local Ollama instance.
The benchmark flow is split intentionally into runtime artifacts and curated results:
* Runtime artifacts go to `.cache/benchmarks/existing-suite/runs/$RUN_ID/`.
* Final benchmark outputs go to `docs/benchmarks/existing-suite/$RUN_ID/`.
Generated result sets now include more than raw `summary.txt` files:
* `report.html` per model for a quick interactive dashboard.
* `RESULTS.md` per model for git-friendly lab notes.
* `SUITE_COMPARISON.html` for side-by-side multi-model comparison.
These reports are generated automatically at the end of the benchmark scripts through `scripts/lib/report-generator.sh`.
Current timeout defaults were raised for slower local models and more reliable result capture:
* `benchmark-stack.sh`: 600s
* `benchmark-model-suite.sh`: 600s
* `benchmark-codeneedle-suite.sh`: 600s
The current Ollama tuning used for benchmark stability is:
* `OLLAMA_KEEP_ALIVE=-1`
* `OLLAMA_KV_CACHE_TYPE=q8_0`
* `OLLAMA_MAX_LOADED_MODELS=1`
* `OLLAMA_NUM_PARALLEL=1`
* `SHM_SIZE=16gb`
OpenClaw path examples:
```bash
./scripts/benchmark-stack.sh --model ollama/qwen3-coding:latest --lines 140 --timeout-sec 300 --no-start
./scripts/benchmark-stack.sh --model ollama/qwen25-coding:latest --lines 140 --timeout-sec 300 --no-start
./scripts/benchmark-stack.sh --model ollama/deepseek-coding:latest --lines 140 --timeout-sec 300 --no-start
./scripts/benchmark-stack.sh --model ollama/qwen3-coding:latest --lines 140 --timeout-sec 600 --no-start
./scripts/benchmark-model-suite.sh --lines 140 --timeout-sec 600
```
This benchmark is useful for latency, stability, and tool-path comparison under identical conditions.
CodeNeedle path example:
```bash
./scripts/benchmark-codeneedle-suite.sh \
--model qwen3-coding:latest \
--model qwen25-coding:latest \
--model deepseek-coding:latest
```
`benchmark-codeneedle-suite.sh` clones the upstream CodeNeedle repository into `.cache/benchmarks/codeneedle/upstream`, builds its Dockerfile with the configured container engine, and benchmarks the standard `http_server` and `jquery` corpora against `http://127.0.0.1:11434`.
Use `--runtime-root`, `--results-root`, or the legacy `--artifact-root` overrides if you want to keep artifacts somewhere else for a run.
Typical curated benchmark outputs look like this:
```text
docs/benchmarks/existing-suite/model-suite-TIMESTAMP/
├── qwen3-coding/
│ ├── summary.txt
│ ├── assistant-output.txt
│ ├── report.html
│ └── RESULTS.md
├── qwen25-coding/
│ └── ...
└── SUITE_COMPARISON.html
```
## Technical Architecture & Key Highlights
This setup relies on unique architectural design patterns engineered to overcome container engines boundaries and system strictness:
+1
View File
@@ -64,6 +64,7 @@ services:
environment:
- OPENCLAW_GATEWAY_MODE=local
- OPENCLAW_GATEWAY_PASSWORD=${OPENCLAW_PASSWORD}
- OPENCLAW_DEFAULT_MODEL=${OPENCLAW_DEFAULT_MODEL:-ollama/qwen3-general}
- OPENCLAW_CONFIG_PATH=/home/node/.openclaw/openclaw.json
- OPENCLAW_STATE_DIR=/home/node/.openclaw
- OLLAMA_URL=http://127.0.0.1:11434
+1
View File
@@ -39,6 +39,7 @@ services:
environment:
- OPENCLAW_GATEWAY_MODE=local
- OPENCLAW_GATEWAY_PASSWORD=${OPENCLAW_PASSWORD}
- OPENCLAW_DEFAULT_MODEL=${OPENCLAW_DEFAULT_MODEL:-ollama/qwen3-general}
- OPENCLAW_CONFIG_PATH=/home/node/.openclaw/openclaw.json
- OPENCLAW_STATE_DIR=/home/node/.openclaw
- OLLAMA_URL=http://127.0.0.1:11434
+1
View File
@@ -37,6 +37,7 @@ services:
environment:
- OPENCLAW_GATEWAY_MODE=local
- OPENCLAW_GATEWAY_PASSWORD=${OPENCLAW_PASSWORD}
- OPENCLAW_DEFAULT_MODEL=${OPENCLAW_DEFAULT_MODEL:-ollama/qwen3-general}
- OPENCLAW_CONFIG_PATH=/home/node/.openclaw/openclaw.json
- OPENCLAW_STATE_DIR=/home/node/.openclaw
- OLLAMA_URL=http://127.0.0.1:11434
+22
View File
@@ -0,0 +1,22 @@
# Basis-Modell aus der Registry laden
FROM qwen3:14b
# Allgemeines Profil fuer breit nutzbare OpenClaw-Chats ohne Coding-Fokus
PARAMETER num_ctx 12288
PARAMETER num_predict 4096
PARAMETER temperature 0.3
# Dem Modell seine primaere Identitaet verpassen
SYSTEM """
Du bist ein vielseitiger, zuverlaessiger KI-Assistent innerhalb des OpenClaw-Frameworks.
Deine Aufgaben:
- Beantworte allgemeine Fragen klar, praezise und hilfreich.
- Hilf bei Analyse, Recherche, Planung, Zusammenfassungen und alltaeglichen Aufgaben.
- Wenn der Nutzer codingbezogene Hilfe will, unterstuetze dabei sauber und sachlich, aber ohne Coding-Fokus als Grundhaltung.
Arbeitsregeln:
- Simuliere keine Terminal-Ausgaben, keine "Let me check..."-Monologe und keine erfundenen Befehlslogs.
- Nutze verfuegbare Tools/Befehle tatsaechlich und berichte nur echte Ergebnisse.
- Wenn etwas nicht fertig ist, nenne den konkreten naechsten Schritt statt langer Zwischenkommentare.
- Antworte standardmaessig auf Deutsch, ausser der Nutzer fordert explizit etwas anderes.
"""
+1 -1
View File
@@ -9,7 +9,7 @@
},
"agents": {
"defaults": {
"model": "ollama/qwen3-coding",
"model": "${OPENCLAW_DEFAULT_MODEL}",
"workspace": "/home/node/.openclaw/workspace",
"sandbox": {
"mode": "all",
+393
View File
@@ -0,0 +1,393 @@
#!/usr/bin/env bash
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
PROJECT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
CODE_NEEDLE_REPO_URL="https://github.com/alexziskind1/codeneedle.git"
usage() {
cat <<'EOF'
Usage: ./scripts/benchmark-codeneedle-suite.sh [options]
Vergleicht mehrere lokale Ollama-Modelle mit CodeNeedle gegen ein OpenAI-
kompatibles Endpoint-Schema. Das Skript klont CodeNeedle zur Laufzeit unter
.cache/benchmarks, baut das Upstream-Dockerfile mit Podman oder Docker und
fuehrt die Standard-Corpora http_server und jquery gegen den lokalen Ollama-
Server aus.
Optionen:
--model <name> Modellname in Ollama; mehrfach nutzbar
(default: qwen3-coding:latest, qwen25-coding:latest,
deepseek-coding:latest)
--corpus <name> CodeNeedle-Corpus; mehrfach nutzbar
(default: http_server, jquery)
--base-url <url> OpenAI-kompatible Basis-URL ohne /v1-Suffix
(default: http://127.0.0.1:11434)
--api-key <value> API-Key fuer den Endpoint (default: not-needed)
--temperature <n> Sampling-Temperatur (default: 0)
--max-tokens <n> Antwortbudget je Query (default: 8000)
--timeout-sec <n> Timeout je Query in Sekunden (default: 600)
--sample-k <n> Anzahl der getesteten Funktionen je Corpus (default: 16)
--sample-seed <n> Seed fuer die Stichprobe (default: 42)
--skip-preflight Ueberspringt CodeNeedle-Preflight-Probe pro Lauf
--runtime-root <path> Zielordner fuer temporaere Laufdaten (default: auto)
--results-root <path> Zielordner fuer kuratierte Resultate (default: auto)
--artifact-root <path> Legacy-Alias; nutzt denselben Pfad fuer Runtime und Resultate
--repo-dir <path> Ablageort fuer den Upstream-Clone (default: auto)
--image-tag <tag> Name des lokal gebauten Runner-Images (default: ghostnet-codeneedle:latest)
--container-engine <e> podman oder docker (default: aus .env/.env.example oder auto)
--refresh-upstream Upstream-Clone vor dem Lauf mit git pull aktualisieren
--rebuild-image Image vor dem Lauf neu bauen
--keep-runtime Runtime-Dateien nach dem Export nicht loeschen
-h, --help Hilfe anzeigen
EOF
}
need_cmd() {
if ! command -v "$1" >/dev/null 2>&1; then
echo "ERROR: Missing required command '$1'." >&2
exit 1
fi
}
read_env_var() {
local key="$1"
local file line
for file in "$PROJECT_DIR/.env" "$PROJECT_DIR/.env.example"; do
if [ ! -f "$file" ]; then
continue
fi
line="$(grep -E "^${key}=" "$file" | tail -n 1 || true)"
if [ -n "$line" ]; then
printf '%s\n' "${line#*=}"
return 0
fi
done
return 1
}
slugify() {
printf '%s' "$1" | tr '/:@ ' '----' | tr -cd '[:alnum:]_.-'
}
MODELS=(
"qwen3-coding:latest"
"qwen25-coding:latest"
"deepseek-coding:latest"
)
CORPORA=(
"http_server"
"jquery"
)
BASE_URL="http://127.0.0.1:11434"
API_KEY="not-needed"
TEMPERATURE="0"
MAX_TOKENS=8000
TIMEOUT_SEC=600
SAMPLE_K=16
SAMPLE_SEED=42
SKIP_PREFLIGHT=0
RUNTIME_ROOT=""
RESULTS_ROOT=""
LEGACY_ARTIFACT_ROOT=""
REPO_DIR=""
IMAGE_TAG="ghostnet-codeneedle:latest"
CONTAINER_ENGINE="${CONTAINER_ENGINE:-}"
REFRESH_UPSTREAM=0
REBUILD_IMAGE=0
KEEP_RUNTIME=0
while [ $# -gt 0 ]; do
case "$1" in
--model)
if [ "${MODELS[0]:-}" = "qwen3-coding:latest" ] && [ "${#MODELS[@]}" -eq 3 ]; then
MODELS=()
fi
MODELS+=("$2")
shift 2
;;
--corpus)
if [ "${CORPORA[0]:-}" = "http_server" ] && [ "${#CORPORA[@]}" -eq 2 ]; then
CORPORA=()
fi
CORPORA+=("$2")
shift 2
;;
--base-url)
BASE_URL="$2"
shift 2
;;
--api-key)
API_KEY="$2"
shift 2
;;
--temperature)
TEMPERATURE="$2"
shift 2
;;
--max-tokens)
MAX_TOKENS="$2"
shift 2
;;
--timeout-sec)
TIMEOUT_SEC="$2"
shift 2
;;
--sample-k)
SAMPLE_K="$2"
shift 2
;;
--sample-seed)
SAMPLE_SEED="$2"
shift 2
;;
--skip-preflight)
SKIP_PREFLIGHT=1
shift
;;
--runtime-root)
RUNTIME_ROOT="$2"
shift 2
;;
--results-root)
RESULTS_ROOT="$2"
shift 2
;;
--artifact-root)
LEGACY_ARTIFACT_ROOT="$2"
shift 2
;;
--repo-dir)
REPO_DIR="$2"
shift 2
;;
--image-tag)
IMAGE_TAG="$2"
shift 2
;;
--container-engine)
CONTAINER_ENGINE="$2"
shift 2
;;
--refresh-upstream)
REFRESH_UPSTREAM=1
shift
;;
--rebuild-image)
REBUILD_IMAGE=1
shift
;;
--keep-runtime)
KEEP_RUNTIME=1
shift
;;
-h|--help)
usage
exit 0
;;
*)
echo "ERROR: Unknown option '$1'." >&2
usage >&2
exit 1
;;
esac
done
need_cmd git
need_cmd jq
if [ -z "$CONTAINER_ENGINE" ]; then
CONTAINER_ENGINE="$(read_env_var CONTAINER_ENGINE || true)"
fi
if [ -z "$CONTAINER_ENGINE" ]; then
if command -v podman >/dev/null 2>&1; then
CONTAINER_ENGINE="podman"
elif command -v docker >/dev/null 2>&1; then
CONTAINER_ENGINE="docker"
else
echo "ERROR: Neither podman nor docker is available." >&2
exit 1
fi
fi
case "$CONTAINER_ENGINE" in
podman)
ENGINE_CMD=(podman)
MOUNT_SUFFIX=':Z'
;;
docker)
ENGINE_CMD=(docker)
MOUNT_SUFFIX=''
;;
*)
echo "ERROR: Invalid container engine '$CONTAINER_ENGINE'. Use 'podman' or 'docker'." >&2
exit 1
;;
esac
need_cmd "${ENGINE_CMD[0]}"
RUN_ID="codeneedle-suite-$(date +%Y%m%d-%H%M%S)"
if [ -n "$LEGACY_ARTIFACT_ROOT" ]; then
RUNTIME_ROOT="$LEGACY_ARTIFACT_ROOT"
if [ -z "$RESULTS_ROOT" ]; then
RESULTS_ROOT="$LEGACY_ARTIFACT_ROOT"
KEEP_RUNTIME=1
fi
fi
if [ -z "$RUNTIME_ROOT" ]; then
RUNTIME_ROOT="$PROJECT_DIR/.cache/benchmarks/$RUN_ID"
fi
if [ -z "$RESULTS_ROOT" ]; then
RESULTS_ROOT="$PROJECT_DIR/docs/benchmarks/codeneedle/$RUN_ID"
fi
if [ -z "$REPO_DIR" ]; then
REPO_DIR="$PROJECT_DIR/.cache/benchmarks/codeneedle/upstream"
fi
mkdir -p "$RUNTIME_ROOT" "$RESULTS_ROOT"
if [ ! -d "$REPO_DIR/.git" ]; then
mkdir -p "$(dirname "$REPO_DIR")"
git clone --depth 1 "$CODE_NEEDLE_REPO_URL" "$REPO_DIR"
elif [ "$REFRESH_UPSTREAM" -eq 1 ]; then
git -C "$REPO_DIR" pull --ff-only
fi
UPSTREAM_REF="$(git -C "$REPO_DIR" rev-parse HEAD)"
if [ "$REBUILD_IMAGE" -eq 1 ] || ! "${ENGINE_CMD[@]}" image inspect "$IMAGE_TAG" >/dev/null 2>&1; then
"${ENGINE_CMD[@]}" build -t "$IMAGE_TAG" -f "$REPO_DIR/Dockerfile" "$REPO_DIR"
fi
SUMMARY_TABLE="$RESULTS_ROOT/summary.tsv"
RUN_METADATA_FILE="$RESULTS_ROOT/run-metadata.txt"
printf 'corpus\tmodel\trun_rc\tpassed\ttotal\terrored\tprimary_matched\tprimary_total\thallucinated\tbonus\n' > "$SUMMARY_TABLE"
cat > "$RUN_METADATA_FILE" <<EOF
run_id=$RUN_ID
container_engine=$CONTAINER_ENGINE
image_tag=$IMAGE_TAG
upstream_repo=$CODE_NEEDLE_REPO_URL
upstream_ref=$UPSTREAM_REF
base_url=$BASE_URL
temperature=$TEMPERATURE
max_tokens=$MAX_TOKENS
timeout_sec=$TIMEOUT_SEC
sample_k=$SAMPLE_K
sample_seed=$SAMPLE_SEED
skip_preflight=$SKIP_PREFLIGHT
repo_dir=$REPO_DIR
runtime_root=$RUNTIME_ROOT
results_root=$RESULTS_ROOT
EOF
echo "CodeNeedle-Repo: $REPO_DIR @ $UPSTREAM_REF"
echo "Container-Engine: $CONTAINER_ENGINE"
echo "Image: $IMAGE_TAG"
echo "Base URL: $BASE_URL"
echo "Runtime-Wurzel: $RUNTIME_ROOT"
echo "Kuratierte Resultate: $RESULTS_ROOT"
echo
for corpus in "${CORPORA[@]}"; do
for model in "${MODELS[@]}"; do
corpus_slug="$(slugify "$corpus")"
model_slug="$(slugify "$model")"
runtime_dir="$RUNTIME_ROOT/$corpus_slug/$model_slug"
results_dir="$RESULTS_ROOT/$corpus_slug/$model_slug"
dump_file="$runtime_dir/results.json"
stdout_file="$runtime_dir/stdout.log"
stderr_file="$runtime_dir/stderr.log"
run_cmd=(
"${ENGINE_CMD[@]}" run --rm --network host
-v "$REPO_DIR:/app$MOUNT_SUFFIX"
-v "$runtime_dir:/runtime$MOUNT_SUFFIX"
-w /app
"$IMAGE_TAG"
python bench.py run
--corpus "$corpus"
--model "$model"
--base-url "$BASE_URL"
--api-key "$API_KEY"
--temperature "$TEMPERATURE"
--max-tokens "$MAX_TOKENS"
--timeout "$TIMEOUT_SEC"
-k "$SAMPLE_K"
--seed "$SAMPLE_SEED"
--dump /runtime/results.json
)
if [ "$SKIP_PREFLIGHT" -eq 1 ]; then
run_cmd+=(--skip-preflight)
fi
mkdir -p "$runtime_dir" "$results_dir"
echo "=== CodeNeedle fuer $model auf $corpus ==="
if "${run_cmd[@]}" >"$stdout_file" 2>"$stderr_file"; then
run_rc=0
else
run_rc=$?
fi
cp "$stdout_file" "$results_dir/stdout.log"
cp "$stderr_file" "$results_dir/stderr.log"
passed=""
total=""
errored=""
primary_matched=""
primary_total=""
hallucinated=""
bonus=""
if [ -f "$dump_file" ]; then
cp "$dump_file" "$results_dir/results.json"
passed="$(jq '[.results[] | select(.passed == true and (.error == null or .error == ""))] | length' "$dump_file")"
total="$(jq '.results | length' "$dump_file")"
errored="$(jq '[.results[] | select(.error != null and .error != "")] | length' "$dump_file")"
primary_matched="$(jq '[.results[].primary_matched] | add // 0' "$dump_file")"
primary_total="$(jq '[.results[].primary_total] | add // 0' "$dump_file")"
hallucinated="$(jq '[.results[].hallucinated] | add // 0' "$dump_file")"
bonus="$(jq '[.results[].bonus_matched] | add // 0' "$dump_file")"
fi
cat > "$results_dir/summary.txt" <<EOF
corpus=$corpus
model=$model
run_rc=$run_rc
passed=$passed
total=$total
errored=$errored
primary_matched=$primary_matched
primary_total=$primary_total
hallucinated=$hallucinated
bonus=$bonus
EOF
printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n' \
"$corpus" "$model" "$run_rc" "$passed" "$total" "$errored" \
"$primary_matched" "$primary_total" "$hallucinated" "$bonus" >> "$SUMMARY_TABLE"
if [ "$KEEP_RUNTIME" -eq 0 ] && [ "$runtime_dir" != "$results_dir" ]; then
rm -rf "$runtime_dir"
fi
done
done
echo
echo "CodeNeedle-Vergleich gespeichert unter: $RESULTS_ROOT"
if command -v column >/dev/null 2>&1; then
column -ts $'\t' "$SUMMARY_TABLE"
else
cat "$SUMMARY_TABLE"
fi
+75 -22
View File
@@ -5,6 +5,9 @@ set -euo pipefail
PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
cd "$PROJECT_DIR"
# Source report generator
source "$PROJECT_DIR/scripts/lib/report-generator.sh" 2>/dev/null || true
usage() {
cat <<'EOF'
Usage: ./scripts/benchmark-model-suite.sh [options]
@@ -14,8 +17,11 @@ und schreibt Artefakte plus eine Vergleichszusammenfassung.
Optionen:
--lines <n> Erwartete Zeilenanzahl LINE_1..LINE_n (default: 140)
--timeout-sec <n> Timeout je Lauf in Sekunden (default: 300)
--artifact-root <p> Zielordner fuer alle Benchmark-Artefakte (default: auto)
--timeout-sec <n> Timeout je Lauf in Sekunden (default: 600)
--runtime-root <p> Zielordner fuer temporaere Laufdaten (default: auto)
--results-root <p> Zielordner fuer kuratierte Resultate (default: auto)
--artifact-root <p> Legacy-Alias; nutzt denselben Pfad fuer Runtime und Resultate
--keep-runtime Runtime-Dateien der Einzelruns nicht loeschen
--session-prefix <s> Prefix fuer Session-Keys (default: suite:<timestamp>)
-h, --help Hilfe anzeigen
EOF
@@ -46,8 +52,11 @@ extract_metric() {
}
EXPECTED_LINES=140
TIMEOUT_SEC=300
ARTIFACT_ROOT=""
TIMEOUT_SEC=600
RUNTIME_ROOT=""
RESULTS_ROOT=""
LEGACY_ARTIFACT_ROOT=""
KEEP_RUNTIME=0
SESSION_PREFIX="suite:$(date +%Y%m%d-%H%M%S)"
while [ "$#" -gt 0 ]; do
@@ -60,10 +69,22 @@ while [ "$#" -gt 0 ]; do
TIMEOUT_SEC="$2"
shift 2
;;
--artifact-root)
ARTIFACT_ROOT="$2"
--runtime-root)
RUNTIME_ROOT="$2"
shift 2
;;
--results-root)
RESULTS_ROOT="$2"
shift 2
;;
--artifact-root)
LEGACY_ARTIFACT_ROOT="$2"
shift 2
;;
--keep-runtime)
KEEP_RUNTIME=1
shift
;;
--session-prefix)
SESSION_PREFIX="$2"
shift 2
@@ -80,11 +101,25 @@ while [ "$#" -gt 0 ]; do
esac
done
if [ -z "$ARTIFACT_ROOT" ]; then
ARTIFACT_ROOT="$PROJECT_DIR/storage/workspace/benchmarks/model-suite-$(date +%Y%m%d-%H%M%S)"
RUN_ID="model-suite-$(date +%Y%m%d-%H%M%S)"
if [ -n "$LEGACY_ARTIFACT_ROOT" ]; then
RUNTIME_ROOT="$LEGACY_ARTIFACT_ROOT"
if [ -z "$RESULTS_ROOT" ]; then
RESULTS_ROOT="$LEGACY_ARTIFACT_ROOT"
KEEP_RUNTIME=1
fi
fi
mkdir -p "$ARTIFACT_ROOT"
if [ -z "$RUNTIME_ROOT" ]; then
RUNTIME_ROOT="$PROJECT_DIR/.cache/benchmarks/$RUN_ID"
fi
if [ -z "$RESULTS_ROOT" ]; then
RESULTS_ROOT="$PROJECT_DIR/docs/benchmarks/existing-suite/$RUN_ID"
fi
mkdir -p "$RUNTIME_ROOT" "$RESULTS_ROOT"
MODELS=(
"ollama/qwen3-coding:latest"
@@ -92,27 +127,36 @@ MODELS=(
"ollama/deepseek-coding:latest"
)
SUMMARY_TABLE="$ARTIFACT_ROOT/summary.tsv"
SUMMARY_TABLE="$RESULTS_ROOT/summary.tsv"
printf 'model\tbenchmark_rc\tvalidation\trequest_exit\twall_ms\tprovider_tps\tend_to_end_tps\tload_sec\tkv_tokens\n' > "$SUMMARY_TABLE"
for model in "${MODELS[@]}"; do
slug="$(slugify_model "$model")"
run_dir="$ARTIFACT_ROOT/$slug"
runtime_dir="$RUNTIME_ROOT/$slug"
results_dir="$RESULTS_ROOT/$slug"
session_key="$SESSION_PREFIX:$slug"
run_args=(
--model "$model"
--lines "$EXPECTED_LINES"
--timeout-sec "$TIMEOUT_SEC"
--session-key "$session_key"
--runtime-dir "$runtime_dir"
--results-dir "$results_dir"
--no-start
)
if [ "$KEEP_RUNTIME" -eq 1 ]; then
run_args+=(--keep-runtime)
fi
echo "=== Benchmark fuer $model ==="
benchmark_rc=0
if ! ./scripts/benchmark-stack.sh \
--model "$model" \
--lines "$EXPECTED_LINES" \
--timeout-sec "$TIMEOUT_SEC" \
--session-key "$session_key" \
--artifact-dir "$run_dir" \
--no-start; then
if ./scripts/benchmark-stack.sh "${run_args[@]}"; then
benchmark_rc=0
else
benchmark_rc=$?
fi
summary_file="$run_dir/summary.txt"
summary_file="$results_dir/summary.txt"
validation="$(extract_value 'Validation' "$summary_file")"
request_exit="$(extract_value 'Request exit code' "$summary_file")"
@@ -128,5 +172,14 @@ for model in "${MODELS[@]}"; do
done
echo
echo "Vergleich gespeichert unter: $ARTIFACT_ROOT"
column -ts $'\t' "$SUMMARY_TABLE"
echo "Runtime-Wurzel: $RUNTIME_ROOT"
echo "Kuratierte Resultate: $RESULTS_ROOT"
column -ts $'\t' "$SUMMARY_TABLE"
# Generate suite comparison report
if [ -f "$SUMMARY_TABLE" ]; then
echo
echo "📊 Generating suite comparison report..."
generate_suite_comparison_report "$SUMMARY_TABLE" "$RESULTS_ROOT" 2>/dev/null || true
echo "✅ Reports generated in: $RESULTS_ROOT"
fi
+206 -27
View File
@@ -5,6 +5,9 @@ set -euo pipefail
PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
cd "$PROJECT_DIR"
# Source report generator
source "$PROJECT_DIR/scripts/lib/report-generator.sh" 2>/dev/null || true
RED='\033[0;31m'
GREEN='\033[0;32m'
YELLOW='\033[1;33m'
@@ -27,6 +30,37 @@ need_cmd() {
command -v "$1" >/dev/null 2>&1 || fail "Benoetigtes Kommando fehlt: $1"
}
read_env_var() {
local key="$1"
local env_file="$PROJECT_DIR/.env"
local line=""
if [ -f "$env_file" ]; then
line="$(grep -E "^${key}=" "$env_file" | tail -n 1 || true)"
fi
if [ -z "$line" ] && [ -f "$PROJECT_DIR/.env.example" ]; then
line="$(grep -E "^${key}=" "$PROJECT_DIR/.env.example" | tail -n 1 || true)"
fi
if [ -n "$line" ]; then
printf '%s' "${line#*=}"
fi
}
resolve_abs_path() {
local raw_path="$1"
if [[ "$raw_path" = /* ]]; then
printf '%s' "$raw_path"
return
fi
if command -v realpath >/dev/null 2>&1; then
realpath -m "$PROJECT_DIR/$raw_path"
return
fi
(cd "$PROJECT_DIR" && printf '%s/%s\n' "$PWD" "${raw_path#./}")
}
usage() {
cat <<'EOF'
Usage: ./scripts/benchmark-stack.sh [options]
@@ -37,9 +71,12 @@ validiert den Output und zeigt Metriken (Laufzeit, Token, KV-/VRAM-Metriken).
Optionen:
--model <name> OpenClaw-Modell (default: ollama/qwen3-coding:latest)
--lines <n> Erwartete Zeilenanzahl LINE_1..LINE_n (default: 120)
--timeout-sec <n> Timeout fuer den Agent-Call in Sekunden (default: 480)
--timeout-sec <n> Timeout fuer den Agent-Call in Sekunden (default: 600)
--session-key <key> Session-Key fuer den Lauf (default: auto)
--artifact-dir <path> Zielverzeichnis fuer Ergebnisdateien (default: auto)
--runtime-dir <path> Zielverzeichnis fuer temporaere Laufdaten (default: auto)
--results-dir <path> Zielverzeichnis fuer kuratierte Resultate (default: auto)
--artifact-dir <path> Legacy-Alias; nutzt denselben Pfad fuer Runtime und Resultate
--keep-runtime Runtime-Dateien nach dem Export nicht loeschen
--no-start Stack nicht starten, nur vorhandenen Stack nutzen
--no-reload Modell vor Lauf nicht mit 'ollama stop' entladen
-h, --help Hilfe anzeigen
@@ -132,13 +169,81 @@ extract_unquoted_metric() {
fi
}
resolve_workspace_media_path() {
local media_path="$1"
local normalized="$media_path"
normalized="${normalized#/workspace/}"
normalized="${normalized#/home/node/.openclaw/workspace/}"
if [ -z "$normalized" ] || [ "$normalized" = "$media_path" -a [[ "$media_path" = /* ]] ]; then
return 1
fi
printf '%s/%s\n' "$HOST_WORKSPACE_ABS" "$normalized"
}
extract_tool_call_text() {
local raw_text="$1"
local extracted
extracted="$(printf '%s\n' "$raw_text" | sed '/^```json$/d; /^```$/d; /^```$/d' | jq -rs '
map(select(type == "object" and (.name == "write" or .name == "edit"))) as $calls
| if ($calls | length) == 0 then ""
else
reduce $calls[] as $call ({content: ""};
if $call.name == "write" then
.content = ($call.arguments.content // .content)
elif $call.name == "edit" then
.content = (
if (($call.arguments.edits // []) | length) > 0 then
($call.arguments.edits[-1].newText // .content)
else
.content
end
)
else
.
end
)
| .content
end
' 2>/dev/null || true)"
if [ -n "$extracted" ]; then
printf '%s' "$extracted"
return 0
fi
return 1
}
extract_media_payload_text() {
local media_path=""
local host_path=""
while IFS= read -r media_path; do
[ -n "$media_path" ] || continue
host_path="$(resolve_workspace_media_path "$media_path" || true)"
if [ -n "$host_path" ] && [ -f "$host_path" ]; then
cat "$host_path"
return 0
fi
done < <(jq -r '.result.payloads[]? | .mediaUrl?, (.mediaUrls[]?)' "$JSON_OUT_FILE" 2>/dev/null | awk 'NF && !seen[$0]++')
return 1
}
MODEL="ollama/qwen3-coding:latest"
EXPECTED_LINES=120
TIMEOUT_SEC=480
TIMEOUT_SEC=600
SESSION_KEY=""
ARTIFACT_DIR=""
RUNTIME_DIR=""
RESULTS_DIR=""
LEGACY_ARTIFACT_DIR=""
START_STACK=1
FORCE_RELOAD=1
KEEP_RUNTIME=0
while [ "$#" -gt 0 ]; do
case "$1" in
@@ -158,10 +263,22 @@ while [ "$#" -gt 0 ]; do
SESSION_KEY="$2"
shift 2
;;
--artifact-dir)
ARTIFACT_DIR="$2"
--runtime-dir)
RUNTIME_DIR="$2"
shift 2
;;
--results-dir)
RESULTS_DIR="$2"
shift 2
;;
--artifact-dir)
LEGACY_ARTIFACT_DIR="$2"
shift 2
;;
--keep-runtime)
KEEP_RUNTIME=1
shift
;;
--no-start)
START_STACK=0
shift
@@ -195,19 +312,44 @@ need_cmd sed
need_cmd awk
need_cmd timeout
if [ -z "$ARTIFACT_DIR" ]; then
TS="$(date +%Y%m%d-%H%M%S)"
ARTIFACT_DIR="$PROJECT_DIR/storage/workspace/benchmarks/$TS"
fi
mkdir -p "$ARTIFACT_DIR"
RUN_ID="$(date +%Y%m%d-%H%M%S)"
HOST_WORKSPACE_RAW="$(read_env_var HOST_WORKSPACE_PATH)"
HOST_WORKSPACE_RAW="${HOST_WORKSPACE_RAW:-./storage/workspace}"
HOST_WORKSPACE_ABS="$(resolve_abs_path "$HOST_WORKSPACE_RAW")"
BENCH_OUTPUT_FILENAME="benchmark-output-${RUN_ID}.txt"
JSON_OUT_FILE="$ARTIFACT_DIR/agent-result.json"
ASSISTANT_TEXT_FILE="$ARTIFACT_DIR/assistant-output.txt"
OLLAMA_LOG_FILE="$ARTIFACT_DIR/ollama-since.log"
STATS_BEFORE_FILE="$ARTIFACT_DIR/container-stats-before.txt"
STATS_AFTER_FILE="$ARTIFACT_DIR/container-stats-after.txt"
REQUEST_ERR_FILE="$ARTIFACT_DIR/request-stderr.log"
SUMMARY_FILE="$ARTIFACT_DIR/summary.txt"
if [ -n "$LEGACY_ARTIFACT_DIR" ]; then
RUNTIME_DIR="$LEGACY_ARTIFACT_DIR"
if [ -z "$RESULTS_DIR" ]; then
RESULTS_DIR="$LEGACY_ARTIFACT_DIR"
KEEP_RUNTIME=1
fi
fi
if [ -z "$RUNTIME_DIR" ]; then
RUNTIME_DIR="$PROJECT_DIR/.cache/benchmarks/existing-suite/runs/$RUN_ID"
fi
if [ -z "$RESULTS_DIR" ]; then
RESULTS_DIR="$PROJECT_DIR/docs/benchmarks/existing-suite/$RUN_ID"
fi
mkdir -p "$RUNTIME_DIR" "$RESULTS_DIR"
JSON_OUT_FILE="$RUNTIME_DIR/agent-result.json"
ASSISTANT_TEXT_FILE="$RUNTIME_DIR/assistant-output.txt"
OLLAMA_LOG_FILE="$RUNTIME_DIR/ollama-since.log"
STATS_BEFORE_FILE="$RUNTIME_DIR/container-stats-before.txt"
STATS_AFTER_FILE="$RUNTIME_DIR/container-stats-after.txt"
REQUEST_ERR_FILE="$RUNTIME_DIR/request-stderr.log"
SUMMARY_FILE="$RUNTIME_DIR/summary.txt"
EXPORTED_SUMMARY_FILE="$RESULTS_DIR/summary.txt"
EXPORTED_ASSISTANT_FILE="$RESULTS_DIR/assistant-output.txt"
EXPORTED_JSON_FILE="$RESULTS_DIR/agent-result.json"
EXPORTED_OLLAMA_LOG_FILE="$RESULTS_DIR/ollama-since.log"
EXPORTED_STATS_BEFORE_FILE="$RESULTS_DIR/container-stats-before.txt"
EXPORTED_STATS_AFTER_FILE="$RESULTS_DIR/container-stats-after.txt"
EXPORTED_REQUEST_ERR_FILE="$RESULTS_DIR/request-stderr.log"
if [ -z "$SESSION_KEY" ]; then
SESSION_KEY="agent:main:benchmark:$(date +%s)"
@@ -221,7 +363,8 @@ echo "Model: $MODEL"
echo "Session key: $SESSION_KEY"
echo "Erwartete Zeilen: $EXPECTED_LINES"
echo "Timeout: ${TIMEOUT_SEC}s"
echo "Artefakte: $ARTIFACT_DIR"
echo "Runtime-Dateien: $RUNTIME_DIR"
echo "Kuratierte Resultate: $RESULTS_DIR"
if [ "$START_STACK" -eq 1 ]; then
echo
@@ -247,7 +390,7 @@ else
fi
echo "[4/6] Fuehre Benchmark-Anfrage aus ..."
PROMPT="Gib exakt ${EXPECTED_LINES} Zeilen aus. Jede Zeile muss exakt den Text LINE_<n> haben, mit aufsteigender Nummerierung von 1 bis ${EXPECTED_LINES}. Kein weiterer Text."
PROMPT="Gib exakt ${EXPECTED_LINES} Zeilen aus. Jede Zeile muss exakt den Text LINE_<n> haben, mit aufsteigender Nummerierung von 1 bis ${EXPECTED_LINES}. Kein weiterer Text. Nutze keine Tools, keine Dateien, keine Anhaenge und kein JSON. Antworte direkt als reiner Nachrichtentext. Falls du trotzdem Tools benutzt, schreibe exakt diese Zeilen in die Datei ${BENCH_OUTPUT_FILENAME} im Workspace und sonst nirgendwo hin."
SINCE_UTC="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
podman stats --no-stream --format 'table {{.Name}}\t{{.CPUPerc}}\t{{.MemUsage}}\t{{.MemPerc}}' ollama openclaw > "$STATS_BEFORE_FILE" 2>&1 || true
@@ -257,6 +400,7 @@ REQUEST_RC=0
if ! timeout "${TIMEOUT_SEC}s" podman exec openclaw openclaw agent \
--session-key "$SESSION_KEY" \
--model "$MODEL" \
--thinking off \
--message "$PROMPT" \
--json > "$JSON_OUT_FILE" 2> "$REQUEST_ERR_FILE"; then
REQUEST_RC=$?
@@ -290,8 +434,19 @@ fi
echo "[5/6] Validiere Antwortformat ..."
ASSISTANT_TEXT=""
RAW_ASSISTANT_TEXT=""
if [ "$HAS_VALID_JSON" -eq 1 ]; then
ASSISTANT_TEXT="$(jq -r '.result.payloads[0].text // ""' "$JSON_OUT_FILE")"
RAW_ASSISTANT_TEXT="$(jq -r '[.result.payloads[]?.text // empty] | join("\n")' "$JSON_OUT_FILE")"
ASSISTANT_TEXT="$RAW_ASSISTANT_TEXT"
if [ -n "$RAW_ASSISTANT_TEXT" ]; then
TOOL_CALL_TEXT="$(extract_tool_call_text "$RAW_ASSISTANT_TEXT" || true)"
if [ -n "$TOOL_CALL_TEXT" ]; then
ASSISTANT_TEXT="$TOOL_CALL_TEXT"
fi
fi
if [ -z "$ASSISTANT_TEXT" ]; then
ASSISTANT_TEXT="$(extract_media_payload_text || true)"
fi
fi
printf '%s' "$ASSISTANT_TEXT" > "$ASSISTANT_TEXT_FILE"
@@ -426,12 +581,12 @@ fi
echo "- journal_gpu_related_hits_since_start: $JOURNAL_HITS"
echo
echo "Artefakte"
echo "- agent_json: $JSON_OUT_FILE"
echo "- assistant_text: $ASSISTANT_TEXT_FILE"
echo "- ollama_log_since_start: $OLLAMA_LOG_FILE"
echo "- stats_before: $STATS_BEFORE_FILE"
echo "- stats_after: $STATS_AFTER_FILE"
echo "- request_stderr: $REQUEST_ERR_FILE"
echo "- agent_json: $EXPORTED_JSON_FILE"
echo "- assistant_text: $EXPORTED_ASSISTANT_FILE"
echo "- ollama_log_since_start: $EXPORTED_OLLAMA_LOG_FILE"
echo "- stats_before: $EXPORTED_STATS_BEFORE_FILE"
echo "- stats_after: $EXPORTED_STATS_AFTER_FILE"
echo "- request_stderr: $EXPORTED_REQUEST_ERR_FILE"
} | tee "$SUMMARY_FILE"
echo
@@ -441,6 +596,30 @@ echo
echo "Container-Stats nach Lauf:"
cat "$STATS_AFTER_FILE"
cp "$SUMMARY_FILE" "$EXPORTED_SUMMARY_FILE"
cp "$ASSISTANT_TEXT_FILE" "$EXPORTED_ASSISTANT_FILE"
cp "$JSON_OUT_FILE" "$EXPORTED_JSON_FILE"
cp "$OLLAMA_LOG_FILE" "$EXPORTED_OLLAMA_LOG_FILE"
cp "$STATS_BEFORE_FILE" "$EXPORTED_STATS_BEFORE_FILE"
cp "$STATS_AFTER_FILE" "$EXPORTED_STATS_AFTER_FILE"
if [ -s "$REQUEST_ERR_FILE" ]; then
cp "$REQUEST_ERR_FILE" "$EXPORTED_REQUEST_ERR_FILE"
fi
if [ "$KEEP_RUNTIME" -eq 0 ] && [ "$RUNTIME_DIR" != "$RESULTS_DIR" ]; then
rm -rf "$RUNTIME_DIR"
fi
# Generate professional reports
if [ -f "$EXPORTED_SUMMARY_FILE" ]; then
ok "Generating HTML report..."
generate_html_report "$MODEL_NAME" "$RESULTS_DIR" "$EXPORTED_SUMMARY_FILE" 2>/dev/null || true
ok "Generating Markdown summary..."
generate_markdown_summary "$MODEL_NAME" "$RESULTS_DIR" "$EXPORTED_SUMMARY_FILE" 2>/dev/null || true
fi
if [ "$REQUEST_RC" -eq 0 ] && [ "$VALIDATION_OK" -eq 1 ]; then
ok "Benchmark erfolgreich abgeschlossen"
exit 0