A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Misc] Reduce initialization time of auto_tune (#23682)
Signed-off-by: Weida Hong <wdhongtw@google.com>
Weida Hong committed
Sep 24, 2025 at 01:34 UTC
24e822274562790a8e7cd7fde2972be719eb38cb
1 file changed
+6
-1
benchmarks/auto_tune/auto_tune.sh
+6
-1
@@ -103,10 +103,15 @@ start_server() {
103
VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
104
vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
105
fi
106
+ local server_pid=$!
107
108
# wait for 10 minutes...
109
server_started=0
110
for i in {1..60}; do
111
+ # This line checks whether the server is still alive or not,
112
+ # since that we should always have permission to send signal to the server process.
113
+ kill -0 $server_pid 2> /dev/null || break
114
+
115
RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
116
STATUS_CODE=$(echo "$RESPONSE" | tail -n 1)
117
if [[ "$STATUS_CODE" -eq 200 ]]; then
@@ -118,7 +123,7 @@ start_server() {
123
done
124
125
if (( ! server_started )); then
121
- echo "server did not start within 10 minutes. Please check server log at $vllm_log".
126
+ echo "server did not start within 10 minutes or crashed. Please check server log at $vllm_log".
127
return 1
128
else
129
return 0