Dashboards, log locations, per-sample outputs, and how to stop services.
⬅ Main README · Prev: « Running experiments · Next: Configuration »
python -m spatial_agent.launch_managers.vllm_manager # [1] vLLM dashboard
python -m spatial_agent.launch_managers.gpu_server_manager # [1] GPU server dashboard
python -m spatial_agent.launch_managers.agent_manager # [1] Agent dashboardtail -f spatial_agent/logs/slurm_vllm/vllm-<served_name>_<job_id>.out
tail -f spatial_agent/logs/slurm_gpu_server/gpu-<chain>_<job_id>.out
tail -f spatial_agent/logs/slurm_agent/spatial-<benchmark>_<job_id>.out
tail -f spatial_agent/logs/slurm_agent/chain_<benchmark>_<exp>_<id>.logA run writes to spatial_agent/work_dir/<work_dir_name>/:
work_dir/<run>/
├── config.json # full resolved config for the run
├── predictions.jsonl # one JSON prediction per sample
├── results_summary.json # aggregate accuracy + per-question-type breakdown
├── report_pdf/<sample_id>.pdf # rendered per-sample PDF report
└── session-<sample_id>/ # one directory per sample
├── session_report.html # interactive HTML trace
├── trace.jsonl # raw event log (init, code cells, tool calls, …)
├── show_images/ # every image registered via show(...)
└── vlm_queries/ # isolated vlm.locate / vlm.ask_with_thinking transcripts
session_report.html is the most useful artifact for debugging — it shows every code cell, its stdout, every variable created, every show() image, and the final answer.
python -m spatial_agent.launch_managers.vllm_manager # [3] Stop Server
python -m spatial_agent.launch_managers.gpu_server_manager # [3] Stop GPU Server(s)
python -m spatial_agent.launch_managers.agent_manager # [4] Stop Experiment(s)scancel -u $USER # cancel all SLURM jobs
pkill -f "spatial_agent.launch_managers" # kill background chain managersNext: Configuration »