| # Fetch a public model, then run one pinned CPU measurement. | |
| set -euo pipefail | |
| if [[ $# -lt 2 || $# -gt 3 ]]; then | |
| echo 'usage: scripts/run_model.sh <hf_repo> <commit_sha> [Q4_K_M|Q5_K_M|Q8_0]' >&2 | |
| exit 2 | |
| fi | |
| MODEL="$1" | |
| REVISION="$2" | |
| QUANT="${3:-Q4_K_M}" | |
| IMAGE='ghcr.io/ggml-org/llama.cpp:full@sha256:5faf86f95747fbb40014a8b28c505d8ec0da3d983d4d368a6650bc056f252688' | |
| HARNESS_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" | |
| EVIDENCE="${EVIDENCE_DIR:-$PWD/evidence}" | |
| mkdir -p "$EVIDENCE" | |
| EVIDENCE="$(cd "$EVIDENCE" && pwd)" | |
| python3 "$HARNESS_DIR/fetch_inputs.py" --model "$MODEL" --revision "$REVISION" --quant "$QUANT" --evidence-dir "$EVIDENCE" | |
| docker run --rm \ | |
| --user "$(id -u):$(id -g)" \ | |
| --cpus="${HARNESS_CPUS:-8}" \ | |
| --memory="${HARNESS_MEMORY:-28g}" \ | |
| --network none \ | |
| --entrypoint /bin/bash \ | |
| -e LD_LIBRARY_PATH=/app \ | |
| -e HOME=/tmp \ | |
| -v "$EVIDENCE:/evidence" \ | |
| -v "$HARNESS_DIR:/harness:ro" \ | |
| "$IMAGE" -lc \ | |
| 'python3 /harness/harness/run_owned.py --inputs /evidence/derived-inputs.json --evidence-dir /evidence' | |