#!/bin/sh
# Run the first one-variable benchmark sweep and preserve its raw evidence.
set -eu

if [ "$#" -ne 2 ]; then
    echo "usage: $0 GEM_BINARY MODEL_GGUF" >&2
    exit 1
fi

gem=$1
model=$2
root=$(CDPATH='' cd -- "$(dirname -- "$0")/.." && pwd)
stamp=$(date +%Y%m%d-%H%M%S)
out=${GEM_BENCH_DIR:-"artifacts/$stamp"}

mkdir -p "$out"
"$gem" print-config --model "$model" > "$out/config.txt"
"$gem" bench --dry-run --model "$model" > "$out/baseline-command.txt"
sw_vers > "$out/macos.txt"
system_profiler SPHardwareDataType SPDisplaysDataType > "$out/hardware.txt"

llama_dir=${GEM_LLAMA_CPP_DIR:-"$root/../qwen38-llama-bench/llama.cpp"}
if [ -d "$llama_dir/.git" ]; then
    git -C "$llama_dir" rev-parse HEAD > "$out/llama-commit.txt"
    git -C "$llama_dir" status --short > "$out/llama-status.txt"
fi

for flash in auto on off; do
    "$gem" bench --model "$model" --flash-attn "$flash" --repetitions 3 \
        > "$out/flash-$flash.json"
done

for flash in on off; do
    "$root/scripts/analyze-benchmarks.sh" "$out/flash-auto.json" \
        "$out/flash-$flash.json" "$out/flash-$flash-comparison.json"
done

echo "gem: wrote benchmark evidence to $out"