scripts/bench.sh

#!/usr/bin/env bash
# Sweep llama-bench across thread counts on this machine to pick the best
# --threads for a given Gemma 4 GGUF. Results saved under scripts/bench-results/.
#
#   ./bench.sh /models/gemma-4-E4B-it-Q4_K_M.gguf
#   THREADS_LIST="2 4 8" ./bench.sh <gguf-path-inside-container>
#
# Requires the GGUF present in ./models (pre-download with download-models.sh
# --weights). Uses the prebuilt llama.cpp image (CPU).
set -euo pipefail
cd "$(dirname "$0")/.."

MODEL="${1:?usage: bench.sh /models/<file>.gguf  (path is inside the container)}"
THREADS_LIST="${THREADS_LIST:-1 2 4 8}"
IMAGE="${LLAMA_IMAGE:-ghcr.io/ggml-org/llama.cpp:full}"
OUT="scripts/bench-results"; mkdir -p "$OUT"
STAMP="$(date +%Y%m%d-%H%M%S)"
REPORT="$OUT/bench-$STAMP.md"

echo "# llama-bench $(uname -n) $STAMP" | tee "$REPORT"
echo "model: $MODEL" | tee -a "$REPORT"
echo | tee -a "$REPORT"

for t in $THREADS_LIST; do
  echo "## threads=$t" | tee -a "$REPORT"
  docker run --rm -v "$PWD/models:/models" "$IMAGE" \
    llama-bench -m "$MODEL" -t "$t" -p 512 -n 128 2>/dev/null \
    | tee -a "$REPORT" || echo "  (run failed for t=$t)" | tee -a "$REPORT"
  echo | tee -a "$REPORT"
done

echo "Saved: $REPORT"