| source ~/kvq-nightly/bin/activate | |
| OUT=~/bench/nightly_check; L=$OUT/tfa_bf16_short.log; : > $L | |
| vllm serve ~/models/gemma-4-31B-it --served-model-name gemma --port 8010 --dtype bfloat16 --max-model-len 4096 \ | |
| --gpu-memory-utilization 0.90 --seed 0 --kv-cache-dtype bfloat16 --attention-backend TRITON_FLASH_ATTN >> $L 2>&1 & | |
| P=$!; ok=0 | |
| for i in $(seq 1 60); do curl -sf localhost:8010/health >/dev/null && { ok=1; break; }; kill -0 $P 2>/dev/null || break; sleep 10; done | |
| if [ $ok = 1 ]; then echo SERVER_UP; python $OUT/img.py; else echo SERVER_FAILED; command grep -E "ValueError|Error:" $L | sed 's/.*\] //' | sort -u | tail -5 | cut -c1-400; fi | |
| kill $P 2>/dev/null; wait $P 2>/dev/null; pkill -f "vllm serve"; echo CONTROL_DONE | |