!10165 Performance optimization of Bert on GPU by the graph_kernel

From: @hanhuifeng2020
Reviewed-by: @gaoxiong1,@ryanww
Signed-off-by: @ryanww

akg

@@ -1 +1 @@
-Subproject commit 72b359ad457ed8f4f254c8a3bd2bde88967202fb
+Subproject commit 065ca5353077903828bebc1baedd4d1c0b052bb6

model_zoo/official/nlp/bert/run_pretrain.py

@@ -94,7 +94,8 @@ def _get_optimizer(args_opt, network):
 def _auto_enable_graph_kernel(device_target, graph_kernel_mode):
     """Judge whether is suitable to enable graph kernel."""
     return graph_kernel_mode in ("auto", "true") and device_target == 'GPU' and \
-        cfg.bert_network == 'base' and cfg.batch_size == 32 and cfg.optimizer == 'AdamWeightDecay'
+        cfg.bert_network == 'base' and (cfg.batch_size == 32 or cfg.batch_size == 64) and \
+        cfg.optimizer == 'AdamWeightDecay'
 
 
 def run_pretrain():
@@ -148,7 +149,8 @@ def run_pretrain():
         context.reset_auto_parallel_context()
         context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, gradients_mean=True,
                                           device_num=device_num)
-        _set_bert_all_reduce_split()
+        if args_opt.device_target == 'Ascend':
+            _set_bert_all_reduce_split()
     else:
         rank = 0
         device_num = 1

model_zoo/official/nlp/bert/scripts/run_distributed_pretrain_for_gpu.sh

@@ -27,18 +27,18 @@ DATA_DIR=$3
 SCHEMA_DIR=$4
 
 mpirun --allow-run-as-root -n $RANK_SIZE --output-filename log_output --merge-stderr-to-stdout \
-	python run_pretrain.py				\
+  python run_pretrain.py        \
-		--device_target="GPU"			\
+    --device_target="GPU"      \
-		--distribute="true"				\
+    --distribute="true"        \
-		--epoch_size=$EPOCH_SIZE		\
+    --epoch_size=$EPOCH_SIZE    \
-		--enable_save_ckpt="true"		\
+    --enable_save_ckpt="true"    \
-		--enable_lossscale="false"		\
+    --enable_lossscale="false"    \
-		--do_shuffle="true"				\
+    --do_shuffle="true"        \
-		--enable_data_sink="true"		\
+    --enable_data_sink="true"    \
-		--data_sink_steps=1				\
+    --data_sink_steps=20        \
-		--load_checkpoint_path=""			\
+    --load_checkpoint_path=""      \
-		--save_checkpoint_steps=10000	\
+    --save_checkpoint_steps=10000  \
-		--save_checkpoint_num=1			\
+    --save_checkpoint_num=1      \
-		--data_dir=$DATA_DIR			\
+    --data_dir=$DATA_DIR      \
-		--schema_dir=$SCHEMA_DIR > log.txt 2>&1 &
+    --schema_dir=$SCHEMA_DIR > log.txt 2>&1 &
 

model_zoo/official/nlp/bert/scripts/run_standalone_pretrain_for_gpu.sh

@@ -39,7 +39,7 @@ python run_pretrain.py  \
     --enable_lossscale="false" \
     --do_shuffle="true" \
     --enable_data_sink="true" \
-    --data_sink_steps=1 \
+    --data_sink_steps=20 \
     --load_checkpoint_path="" \
     --save_checkpoint_path="" \
     --save_checkpoint_steps=10000 \