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] auto_tune: kill specific vllm process (#26304)
Signed-off-by: Karan Goel <karangoel@google.com>
Karan Goel committed
Oct 6, 2025 at 11:02 UTC
824a3f403f8cbb5b7984d33ca4741f67928c1d61
1 file changed
+4
-4
benchmarks/auto_tune/auto_tune.sh
+4
-4
@@ -74,7 +74,7 @@ start_server() {
74
local vllm_log=$4
75
local profile_dir=$5
76
77
- pkill -if vllm
77
+ pkill -if "vllm serve" || true
78
79
# Define the common arguments as a bash array.
80
# Each argument and its value are separate elements.
@@ -139,7 +139,7 @@ run_benchmark() {
139
echo "vllm_log: $vllm_log"
140
echo
141
rm -f $vllm_log
142
- pkill -if vllm
142
+ pkill -if "vllm serve" || true
143
144
echo "starting server..."
145
# Call start_server without a profile_dir to avoid profiling overhead
@@ -232,7 +232,7 @@ run_benchmark() {
232
233
echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
234
235
- pkill -if vllm
235
+ pkill -if "vllm serve" || true
236
sleep 10
237
echo "===================="
238
return 0
@@ -308,6 +308,6 @@ if (( $(echo "$best_throughput > 0" | bc -l) )); then
308
else
309
echo "No configuration met the latency requirements. Skipping final profiling run."
310
fi
311
-pkill -if vllm
311
+pkill -if "vllm serve" || true
312
echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
313
echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"