Switches BBS example to use mbsize=3 and gas=2 to fit in 16GB of memory. (#341)

tests/model/BingBertSquad/deepspeed_bsz24_fp16_config.json

 {
   "train_batch_size": 24,
-  "train_micro_batch_size_per_gpu": 6,
+  "train_micro_batch_size_per_gpu": 3,
   "steps_per_print": 1,
   "optimizer": {
     "type": "Adam",

tests/model/BingBertSquad/deepspeed_bsz24_fp16_zero2_config.json

 {
   "train_batch_size": 24,
-  "train_micro_batch_size_per_gpu": 6,
+  "train_micro_batch_size_per_gpu": 3,
   "steps_per_print": 1,
   "optimizer": {
     "type": "Adam",

tests/model/BingBertSquad/deepspeed_bsz24_fp32_config.json

 {
   "train_batch_size": 24,
-  "train_micro_batch_size_per_gpu": 6,
+  "train_micro_batch_size_per_gpu": 3,
   "steps_per_print": 1,
   "optimizer": {
     "type": "Adam",

tests/model/BingBertSquad/run_BingBertSquad.sh

@@ -121,7 +121,7 @@ echo "deepspeed: ${enable_deepspeed}"
 echo "other_args: ${other_args}"
 
 EFFECTIVE_BATCH_SIZE=${batch_size}
-MAX_GPU_BATCH_SIZE=6
+MAX_GPU_BATCH_SIZE=3
 PER_GPU_BATCH_SIZE=$((EFFECTIVE_BATCH_SIZE/num_gpus))
 if [[ $PER_GPU_BATCH_SIZE -lt $MAX_GPU_BATCH_SIZE ]]; then
        GRAD_ACCUM_STEPS=1

tests/model/BingBertSquad/run_BingBertSquad_sanity.sh

@@ -122,7 +122,7 @@ echo "deepspeed: ${enable_deepspeed}"
 echo "other_args: ${other_args}"
 
 EFFECTIVE_BATCH_SIZE=${batch_size}
-MAX_GPU_BATCH_SIZE=6
+MAX_GPU_BATCH_SIZE=3
 PER_GPU_BATCH_SIZE=$((EFFECTIVE_BATCH_SIZE/num_gpus))
 if [[ $PER_GPU_BATCH_SIZE -lt $MAX_GPU_BATCH_SIZE ]]; then
        GRAD_ACCUM_STEPS=1