| # Experiment 1 Cell A — Agent-as-Tool + Direct Python tool calls (no MCP). | |
| # | |
| # This cell measures the baseline cost of the tool set without the MCP | |
| # JSON-RPC layer, so (Cell B latency) - (Cell A latency) isolates MCP | |
| # transport overhead. See docs/experiment1_capture_plan.md for the full | |
| # design and docs/execution_plan.md for the 5-cell experimental grid. | |
| # | |
| # STATUS (2026-04-24): runner wired via scripts/aat_runner.py (#104). | |
| # Cell A smoke evidence is the follow-up in Task 8 of the plan; the | |
| # canonical Insomnia Llama-3.1-8B capture for #25 is also a follow-up. | |
| # | |
| # Usage once unblocked: | |
| # sbatch --mail-type=BEGIN,END,FAIL --mail-user=$MAIL_USER \ | |
| # scripts/run_experiment.sh configs/aat_direct.env | |
| EXPERIMENT_NAME="aat_direct" | |
| EXPERIMENT_CELL="A" | |
| EXPERIMENT_FAMILY="exp1_mcp_overhead" | |
| SCENARIO_SET_NAME="smartgrid_multi_domain" | |
| SCENARIOS_GLOB="data/scenarios/multi_*.json" | |
| MODEL_ID="openai/Llama-3.1-8B-Instruct" | |
| ORCHESTRATION="agent_as_tool" | |
| MCP_MODE="direct" | |
| TRIALS=3 | |
| ENABLE_SMARTGRID_SERVERS=0 # Direct path bypasses MCP servers entirely | |
| CONTRIBUTING_EXPERIMENTS="exp1_mcp_overhead" | |
| SCENARIO_DOMAIN_SCOPE="multi_domain" | |
| MODEL_PROVIDER="vllm" | |
| SERVING_STACK="insomnia_vllm" | |
| QUANTIZATION_MODE="fp16" | |
| # 32768 matches the repo's canonical benchmark-context lane (configs/example_*.env, | |
| # configs/experiment2/*.env). Smoke configs (configs/aat_*_smoke.env) stay at | |
| # 8192. See #135 for context: Cell A run 8979314 hit a replay context-window | |
| # exceeded path at 8192 → 8193 input tokens; 32768 is the proven replay headroom. | |
| MAX_MODEL_LEN=32768 | |
| TEMPERATURE=0.0 | |
| MAX_TOKENS=0 | |
| LAUNCH_VLLM=1 | |
| VLLM_MODEL_PATH="models/Llama-3.1-8B-Instruct" | |
| VLLM_PORT=8000 | |
| VLLM_ENABLE_AUTO_TOOL_CHOICE=1 | |
| VLLM_TOOL_CALL_PARSER="llama3_json" | |
| ENABLE_WANDB=1 | |
| WANDB_ENTITY="assetopsbench-smartgrid" | |
| WANDB_PROJECT="assetopsbench-smartgrid" | |
| WANDB_MODE="online" | |
| # Agent-as-Tool dispatch now defaults to scripts/aat_runner.py. Set | |
| # AAT_RUNNER_TEMPLATE only for custom parity/variant smoke commands. | |
| AAT_PARALLEL_TOOL_CALLS=false | |
| # Torch profiler: captures one replay pass per run while vLLM is still live. | |
| # Trace lands in profiling/traces/<RUN_ID>_torch/. Override: TORCH_PROFILE=0 sbatch ... | |
| TORCH_PROFILE="${TORCH_PROFILE:-1}" | |