!3872 fix bad perf in 16 cards scenario

Merge pull request !3872 from luoyang/mymaster

official/cv/yolov5/scripts/run_distribute_train.sh

@@ -50,7 +50,7 @@ export RANK_SIZE=8
 export RANK_TABLE_FILE=$RANK_TABLE_FILE
 
 cpus=`cat /proc/cpuinfo| grep "processor"| wc -l`
-avg=`expr $cpus \/ $RANK_SIZE`
+avg=`expr $cpus \/ $DEVICE_NUM`
 gap=`expr $avg \- 1`
 
 for((i=0; i<${DEVICE_NUM}; i++))