| # Launch vLLM server for google/gemma-4-31B-it. | |
| # | |
| # Hardware: 2× L40S 48GB (TP=2). Note: Gemma shares the same GPU pair (0-1 by | |
| # default here) as Magistral in the paper's triadic-sweep node layout. They | |
| # cannot run simultaneously; stop one before starting the other. | |
| # Default port: 8009. Override with PORT env var. | |
| # | |
| # Key settings: | |
| # --tool-call-parser gemma4 — requires the custom jinja chat template | |
| # --chat-template — path to tool_chat_template_gemma4.jinja | |
| # tool_choice=auto — Gemma 4 returns empty tool_calls with 'required' | |
| # --max-num-seqs 64 — tuned to avoid KV-cache saturation under | |
| # concurrent multi-agent load | |
| # --no-scheduler-reserve-full-isl — avoids a vLLM 0.19.0 scheduler deadlock | |
| # | |
| # Note on max_tokens: Gemma 4 output is capped at 512 tokens in the runner | |
| # (compared to 1536 for other models) to limit KV-cache pressure. | |
| set -euo pipefail | |
| HF_HUB_CACHE="${HF_HUB_CACHE:-/data/hf_cache/hub}" | |
| export HF_HUB_CACHE | |
| VENV="${SAFEMASS_VENV:-$(dirname "$0")/../.venv}" | |
| PORT="${PORT:-8009}" | |
| LOGS="${LOGS_DIR:-/tmp/vllm_logs}" | |
| SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" | |
| TEMPLATE="${GEMMA_CHAT_TEMPLATE:-${SCRIPT_DIR}/../oasis/tool_chat_template_gemma4.jinja}" | |
| mkdir -p "$LOGS" | |
| if [[ ! -f "$TEMPLATE" ]]; then | |
| echo "ERROR: Gemma chat template not found at: $TEMPLATE" >&2 | |
| echo "Set GEMMA_CHAT_TEMPLATE env var or ensure oasis/tool_chat_template_gemma4.jinja exists." >&2 | |
| exit 1 | |
| fi | |
| CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1}" \ | |
| "${VENV}/bin/python3" -m vllm.entrypoints.openai.api_server \ | |
| --model google/gemma-4-31B-it \ | |
| --served-model-name google/gemma-4-31B-it \ | |
| --tensor-parallel-size 2 \ | |
| --max-model-len 16384 \ | |
| --gpu-memory-utilization 0.90 \ | |
| --port "$PORT" \ | |
| --host 127.0.0.1 \ | |
| --trust-remote-code \ | |
| --enable-auto-tool-choice \ | |
| --tool-call-parser gemma4 \ | |
| --chat-template "$TEMPLATE" \ | |
| --max-num-seqs 64 \ | |
| --no-scheduler-reserve-full-isl \ | |
| >> "${LOGS}/vllm_gemma_${PORT}.log" 2>&1 | |