Skip to content

Commit cc3c2f0

Browse files
liym27liym27
committed
add test for pir sequence parallel
1 parent f6fc7ff commit cc3c2f0

File tree

3 files changed

+202
-2
lines changed

3 files changed

+202
-2
lines changed

‎paddlenlp/trainer/auto_trainer.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -40,7 +40,7 @@
4040
has_length,
4141
speed_metrics,
4242
)
43-
from .utils.ckpt_converter import CheckpointConverter
43+
# from .utils.ckpt_converter import CheckpointConverter
4444
from .utils.helper import distributed_file, distributed_isfile # nested_truncate,
4545

4646
try:

‎paddlenlp/trainer/trainer.py

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1302,14 +1302,16 @@ def _maybe_log_save_evaluate(self, tr_loss, model, epoch, ignore_keys_for_eval,
13021302
logs: Dict[str, float] = {}
13031303

13041304
# all_gather + mean() to get average loss over all processes
1305-
tr_loss_scalar = self._get_item_from_loss(self._nested_gather(tr_loss).mean())
1305+
avg_loss = self._nested_gather(tr_loss).mean()
1306+
tr_loss_scalar = self._get_item_from_loss(avg_loss)
13061307

13071308
# reset tr_loss to zero
13081309
tr_loss.subtract_(tr_loss)
13091310

13101311
logs["loss"] = round(tr_loss_scalar / (self.state.global_step - self._globalstep_last_logged), 8)
13111312
logs["learning_rate"] = float("{0:.3e}".format(self._get_learning_rate()))
13121313
logs["global_step"] = int(self.state.global_step)
1314+
logs["loss_md5"] = avg_loss._md5sum()
13131315

13141316
divisor = 2**30
13151317
# TODO(@gexiao): replace these codes with unified APIs in Paddle

‎scripts/distribute/ci_case_auto.sh

Lines changed: 198 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -55,6 +55,8 @@ function llama_case_list_auto() {
5555

5656
llama_align_dygraph_dy2st_auto_bs2_bf16_DP2-MP1-PP1
5757
llama_convert_hybrid_ckpt_to_auto_parallel_bs2_fp32_DP2-MP1-PP1
58+
llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP1-SP
59+
llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP2-SP
5860
}
5961

6062
function llm_gpt_case_list_auto() {
@@ -971,6 +973,187 @@ function llama_dy2st_auto_bs4_bf16_DP1-MP1-PP4-SD2-VPP3_split_bw() {
971973
echo "=========== $FUNCNAME run end ==========="
972974
}
973975

976+
function llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP1-SP() {
977+
echo "=========== $FUNCNAME run begin ==========="
978+
export PYTHONPATH=$root_path/:$PYTHONPATH
979+
export PYTHONPATH=/paddle/Paddle/build_gpu/python/:$PYTHONPATH
980+
export FLAGS_call_stack_level=3
981+
export FLAGS_enable_pir_api=1
982+
export FLAGS_dynamic_static_unified_comm=1
983+
984+
export NVIDIA_TF32_OVERRIDE=0
985+
export FLAGS_cudnn_deterministic=1
986+
export FLAGS_embedding_deterministic=1
987+
988+
task_name="llama_align_dygraph_dy2st_pir_auto_bs2_bf16_dp2mp2pp1_sp"
989+
case_out_dir="output/$task_name"
990+
case_log_dir="output/$task_name""_log"
991+
992+
for to_static in "0" "1"; do
993+
rm -rf $case_out_dir
994+
rm -rf $case_log_dir
995+
python -u -m paddle.distributed.launch \
996+
--gpus "0,1,2,3" \
997+
--log_dir $case_log_dir \
998+
run_pretrain_auto.py \
999+
--model_type "llama" \
1000+
--model_name_or_path "facebook/llama-7b" \
1001+
--tokenizer_name_or_path "facebook/llama-7b" \
1002+
--input_dir "./data" \
1003+
--output_dir $case_out_dir \
1004+
--split 949,50,1 \
1005+
--weight_decay 0.01 \
1006+
--warmup_ratio 0.01 \
1007+
--max_grad_norm 0.0 \
1008+
--learning_rate 3e-05 \
1009+
--min_learning_rate 3e-06 \
1010+
--max_steps 10 \
1011+
--logging_steps 10 \
1012+
--eval_steps 1000 \
1013+
--save_steps 50000 \
1014+
--continue_training 0 \
1015+
--do_train true \
1016+
--do_eval false \
1017+
--do_predict false \
1018+
--disable_tqdm true \
1019+
--skip_profile_timer true \
1020+
--save_total_limit 2 \
1021+
--device gpu \
1022+
--disable_tqdm true \
1023+
--dataloader_num_workers 1 \
1024+
--enable_auto_parallel 1 \
1025+
--per_device_train_batch_size 1 \
1026+
--gradient_accumulation_steps 1 \
1027+
--per_device_eval_batch_size 2 \
1028+
--recompute false \
1029+
--bf16 1\
1030+
--fp16_opt_level "O2" \
1031+
--amp_custom_black_list "reduce_sum" "c_softmax_with_cross_entropy" \
1032+
--amp_custom_white_list "lookup_table" "lookup_table_v2" \
1033+
--amp_master_grad 1 \
1034+
--fuse_attention_ffn false \
1035+
--fuse_attention_qkv false \
1036+
--fuse_sequence_parallel_allreduce false \
1037+
--use_flash_attention 0 \
1038+
--use_fused_rope false \
1039+
--use_fused_rms_norm 0 \
1040+
--max_seq_length 4096 \
1041+
--sep_parallel_degree 1 \
1042+
--sequence_parallel true \
1043+
--pipeline_parallel_degree 1 \
1044+
--sharding_parallel_degree 1 \
1045+
--tensor_parallel_degree 2 \
1046+
--virtual_pp_degree 1 \
1047+
--sharding "" \
1048+
--to_static ${to_static} \
1049+
--num_hidden_layers 4 \
1050+
>>${log_path}/$FUNCNAME 2>&1
1051+
loss=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss: ' '{print $2}' | awk -F ',' '{print $1}'`
1052+
loss_md5=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss_md5: ' '{print $2}' | awk -F ',' '{print $1}'`
1053+
ips=-1
1054+
mem=-1
1055+
echo "result: to_static=$to_static loss=$loss ips=$ips mem=$mem"
1056+
loss_base=9.34753036
1057+
loss_md5_base=8ea72495fba4e1b9ba004b4431e27218
1058+
ips_base=-1
1059+
mem_base=-1
1060+
check_result $FUNCNAME ${loss_base} ${loss} ${ips_base} ${ips} ${mem_base} ${mem}
1061+
check_md5_result $FUNCNAME ${loss_md5_base} ${loss_md5}
1062+
done
1063+
echo "=========== $FUNCNAME run end ==========="
1064+
}
1065+
1066+
function llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP2-SP() {
1067+
echo "=========== $FUNCNAME run begin ==========="
1068+
export PYTHONPATH=$root_path/:$PYTHONPATH
1069+
export PYTHONPATH=/paddle/Paddle/build_gpu/python/:$PYTHONPATH
1070+
export FLAGS_call_stack_level=3
1071+
export FLAGS_enable_pir_api=1
1072+
export FLAGS_dynamic_static_unified_comm=1
1073+
1074+
export NVIDIA_TF32_OVERRIDE=0
1075+
export FLAGS_cudnn_deterministic=1
1076+
export FLAGS_embedding_deterministic=1
1077+
1078+
task_name="llama_align_dygraph_dy2st_pir_auto_bs2_bf16_dp2mp2pp2_sp"
1079+
case_out_dir="output/$task_name"
1080+
case_log_dir="output/$task_name""_log"
1081+
1082+
for to_static in "0" "1"; do
1083+
rm -rf $case_out_dir
1084+
rm -rf $case_log_dir
1085+
python -u -m paddle.distributed.launch \
1086+
--gpus "0,1,2,3,4,5,6,7" \
1087+
--log_dir $case_log_dir \
1088+
run_pretrain_auto.py \
1089+
--model_type "llama" \
1090+
--model_name_or_path "facebook/llama-7b" \
1091+
--tokenizer_name_or_path "facebook/llama-7b" \
1092+
--input_dir "./data" \
1093+
--output_dir $case_out_dir \
1094+
--split 949,50,1 \
1095+
--weight_decay 0.01 \
1096+
--warmup_ratio 0.01 \
1097+
--max_grad_norm 0.0 \
1098+
--learning_rate 3e-05 \
1099+
--min_learning_rate 3e-06 \
1100+
--max_steps 10 \
1101+
--logging_steps 10 \
1102+
--eval_steps 1000 \
1103+
--save_steps 50000 \
1104+
--continue_training 0 \
1105+
--do_train true \
1106+
--do_eval false \
1107+
--do_predict false \
1108+
--disable_tqdm true \
1109+
--skip_profile_timer true \
1110+
--save_total_limit 2 \
1111+
--device gpu \
1112+
--disable_tqdm true \
1113+
--dataloader_num_workers 1 \
1114+
--enable_auto_parallel 1 \
1115+
--per_device_train_batch_size 1 \
1116+
--gradient_accumulation_steps 1 \
1117+
--per_device_eval_batch_size 2 \
1118+
--recompute false \
1119+
--bf16 1\
1120+
--fp16_opt_level "O2" \
1121+
--amp_custom_black_list "reduce_sum" "c_softmax_with_cross_entropy" \
1122+
--amp_custom_white_list "lookup_table" "lookup_table_v2" \
1123+
--amp_master_grad 1 \
1124+
--fuse_attention_ffn false \
1125+
--fuse_attention_qkv false \
1126+
--fuse_sequence_parallel_allreduce false \
1127+
--use_flash_attention 0 \
1128+
--use_fused_rope false \
1129+
--use_fused_rms_norm 0 \
1130+
--max_seq_length 4096 \
1131+
--sep_parallel_degree 1 \
1132+
--sequence_parallel true \
1133+
--pipeline_parallel_degree 2 \
1134+
--sharding_parallel_degree 1 \
1135+
--tensor_parallel_degree 2 \
1136+
--virtual_pp_degree 1 \
1137+
--sharding "" \
1138+
--to_static ${to_static} \
1139+
--num_hidden_layers 4 \
1140+
>>${log_path}/$FUNCNAME 2>&1
1141+
loss=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss: ' '{print $2}' | awk -F ',' '{print $1}'`
1142+
loss_md5=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss_md5: ' '{print $2}' | awk -F ',' '{print $1}'`
1143+
ips=-1
1144+
mem=-1
1145+
echo "result: to_static=$to_static loss=$loss loss_md5=$loss_md5 ips=$ips mem=$mem"
1146+
loss_base=9.25199432
1147+
loss_md5_base=83531e98ee11cd271db175150ab254bb
1148+
ips_base=-1
1149+
mem_base=-1
1150+
check_result $FUNCNAME ${loss_base} ${loss} ${ips_base} ${ips} ${mem_base} ${mem}
1151+
check_md5_result $FUNCNAME ${loss_md5_base} ${loss_md5}
1152+
done
1153+
echo "=========== $FUNCNAME run end ==========="
1154+
}
1155+
1156+
9741157
function llama_align_dygraph_dy2st_auto_bs2_bf16_DP2-MP1-PP1() {
9751158
echo "=========== $FUNCNAME run begin ==========="
9761159
export PYTHONPATH=$root_path/:$PYTHONPATH
@@ -1886,6 +2069,21 @@ EOF
18862069

18872070
############ case end ############
18882071

2072+
function check_md5_result() {
2073+
echo -e "$1" >> ${log_path}/result.log
2074+
2075+
if [ $# -ne 3 ]; then
2076+
echo -e "\033[31m $1 parameter transfer failed: $@ \033[0m" | tee -a ${log_path}/result.log
2077+
exit -1
2078+
fi
2079+
2080+
echo -e "loss_md5_base: $2 loss_md5: $3" | tee -a ${log_path}/result.log
2081+
if [ $2 != $3 ];then
2082+
echo -e "\033[31m $1 loss_md5 diff check failed! \033[0m" | tee -a ${log_path}/result.log
2083+
exit -1
2084+
fi
2085+
}
2086+
18892087
function check_result() {
18902088
echo -e "$1" >> ${log_path}/result.log
18912089
if [ $? -ne 0 ];then

0 commit comments

Comments
 (0)