| #!/bin/bash |
| |
| set -u |
| WS=/mnt/pvc/users/simon/agentptb/runs/d/workspace |
| CKPT=$1 |
| TAG=$2 |
| PORT=8500 |
| cd /root/work/a/prime-rl |
| export PRIME_API_KEY="$(cat $WS/.prime-api-key)" |
|
|
| nohup uv run inference --model.name "$CKPT" \ |
| --model.tool-call-parser qwen3_coder --model.reasoning-parser qwen3 \ |
| --server.port $PORT --parallel.tp 1 --parallel.dp 1 \ |
| --vllm-extra '{"limit_mm_per_prompt": {"image": 0, "video": 0}}' \ |
| > $WS/logs/inference-final-$TAG.log 2>&1 & |
| SPID=$! |
|
|
| ok=0 |
| for i in $(seq 1 60); do |
| curl -sf http://localhost:$PORT/health >/dev/null 2>&1 && { ok=1; break; } |
| sleep 10 |
| done |
| [ $ok -eq 1 ] || { echo "SERVER FAILED"; tail -5 $WS/logs/inference-final-$TAG.log; exit 1; } |
|
|
| OUT=$WS/evals/final-$TAG |
| mkdir -p "$OUT" |
|
|
| echo "=== tb2 89 ===" |
| uv run --no-sync eval terminal-bench-2-v1 -m "$CKPT" \ |
| --client.base-url http://127.0.0.1:$PORT/v1 \ |
| --env.agent.harness.id pi --env.agent.runtime.type broker \ |
| --env.agent.runtime.ready_timeout_seconds 1800 \ |
| --sampling.temperature 0.3 \ |
| -n 89 -r 1 -s -c 24 --no-push --rich false \ |
| -o "$OUT/tb2" > "$OUT/tb2.log" 2>&1 |
| uv run --no-sync python $WS/scripts/score.py "$OUT/tb2/traces.jsonl" || true |
|
|
| echo "=== swe 100 ===" |
| uv run --no-sync eval swebench-verified-v1 -m "$CKPT" \ |
| --client.base-url http://127.0.0.1:$PORT/v1 \ |
| --env.agent.harness.id pi --env.agent.runtime.type broker \ |
| --env.agent.runtime.ready_timeout_seconds 1800 \ |
| --sampling.temperature 0.3 \ |
| -n 100 -r 1 -s -c 64 --no-push --rich false \ |
| -o "$OUT/swe" > "$OUT/swe.log" 2>&1 |
| uv run --no-sync python $WS/scripts/score.py "$OUT/swe/traces.jsonl" || true |
|
|
| kill $SPID 2>/dev/null |
| echo "=== FINAL DONE $TAG ===" |
|
|