Skip to content

Commit f5b73df

Browse files
waliwali777waliwali777
committed
merge test
1 parent d203f2a commit f5b73df

File tree

1 file changed

+76
-72
lines changed

1 file changed

+76
-72
lines changed

‎scripts/distribute/ci_case_auto.sh‎

Lines changed: 76 additions & 72 deletions
Original file line numberDiff line numberDiff line change
@@ -668,78 +668,82 @@ function llama_align_dygraph_dy2st_pir_auto_bs2_bf16_DP2-MP2-PP1-SP() {
668668
case_log_dir="output/$task_name""_log"
669669

670670
for to_static in "0" "1"; do
671-
rm -rf $case_out_dir
672-
rm -rf $case_log_dir
673-
python -u -m paddle.distributed.launch \
674-
--gpus "0,1,2,3" \
675-
--log_dir $case_log_dir \
676-
run_pretrain_auto.py \
677-
--model_type "llama" \
678-
--model_name_or_path "facebook/llama-7b" \
679-
--tokenizer_name_or_path "facebook/llama-7b" \
680-
--input_dir "./data" \
681-
--output_dir $case_out_dir \
682-
--split 949,50,1 \
683-
--weight_decay 0.01 \
684-
--warmup_ratio 0.01 \
685-
--max_grad_norm 0.0 \
686-
--learning_rate 3e-05 \
687-
--min_learning_rate 3e-06 \
688-
--max_steps 10 \
689-
--logging_steps 10 \
690-
--eval_steps 1000 \
691-
--save_steps 50000 \
692-
--continue_training 0 \
693-
--do_train true \
694-
--do_eval false \
695-
--do_predict false \
696-
--disable_tqdm true \
697-
--skip_profile_timer true \
698-
--save_total_limit 2 \
699-
--device gpu \
700-
--disable_tqdm true \
701-
--dataloader_num_workers 1 \
702-
--enable_auto_parallel 1 \
703-
--per_device_train_batch_size 1 \
704-
--gradient_accumulation_steps 1 \
705-
--per_device_eval_batch_size 2 \
706-
--recompute false \
707-
--bf16 1\
708-
--fp16_opt_level "O2" \
709-
--amp_custom_black_list "reduce_sum" "c_softmax_with_cross_entropy" \
710-
--amp_custom_white_list "lookup_table" "lookup_table_v2" \
711-
--amp_master_grad 1 \
712-
--fuse_attention_ffn false \
713-
--fuse_attention_qkv false \
714-
--fuse_sequence_parallel_allreduce false \
715-
--use_flash_attention 0 \
716-
--use_fused_rope false \
717-
--use_fused_rms_norm 0 \
718-
--max_seq_length 4096 \
719-
--sep_parallel_degree 1 \
720-
--sequence_parallel true \
721-
--pipeline_parallel_degree 1 \
722-
--sharding_parallel_degree 1 \
723-
--tensor_parallel_degree 2 \
724-
--virtual_pp_degree 1 \
725-
--sharding "" \
726-
--to_static ${to_static} \
727-
--num_hidden_layers 4 \
728-
>>${log_path}/$FUNCNAME 2>&1
729-
loss=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss: ' '{print $2}' | awk -F ',' '{print $1}'`
730-
loss_md5=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss_md5: ' '{print $2}' | awk -F ',' '{print $1}'`
731-
ips=-1
732-
mem=-1
733-
echo "result: to_static=$to_static loss=$loss ips=$ips mem=$mem"
734-
loss_base=9.16783295
735-
loss_md5_base=8ea72495fba4e1b9ba004b4431e27218
736-
if [ $IS_A100 -ne 0 ];then
737-
loss_base=9.38009949
738-
fi
739-
ips_base=-1
740-
mem_base=-1
741-
check_result $FUNCNAME ${loss_base} ${loss} ${ips_base} ${ips} ${mem_base} ${mem}
742-
# check_md5_result $FUNCNAME ${loss_md5_base} ${loss_md5}
671+
for use_recompute in "1" "0"; do
672+
rm -rf $case_out_dir
673+
rm -rf $case_log_dir
674+
python -u -m paddle.distributed.launch \
675+
--gpus "0,1,2,3" \
676+
--log_dir $case_log_dir \
677+
run_pretrain_auto.py \
678+
--model_type "llama" \
679+
--model_name_or_path "facebook/llama-7b" \
680+
--tokenizer_name_or_path "facebook/llama-7b" \
681+
--input_dir "./data" \
682+
--output_dir $case_out_dir \
683+
--split 949,50,1 \
684+
--weight_decay 0.01 \
685+
--warmup_ratio 0.01 \
686+
--max_grad_norm 0.0 \
687+
--learning_rate 3e-05 \
688+
--min_learning_rate 3e-06 \
689+
--max_steps 10 \
690+
--logging_steps 10 \
691+
--eval_steps 1000 \
692+
--save_steps 50000 \
693+
--continue_training 0 \
694+
--do_train true \
695+
--do_eval false \
696+
--do_predict false \
697+
--disable_tqdm true \
698+
--skip_profile_timer true \
699+
--save_total_limit 2 \
700+
--device gpu \
701+
--disable_tqdm true \
702+
--dataloader_num_workers 1 \
703+
--enable_auto_parallel 1 \
704+
--per_device_train_batch_size 1 \
705+
--gradient_accumulation_steps 1 \
706+
--per_device_eval_batch_size 2 \
707+
--recompute ${use_recompute} \
708+
--bf16 1\
709+
--fp16_opt_level "O2" \
710+
--amp_custom_black_list "reduce_sum" "c_softmax_with_cross_entropy" \
711+
--amp_custom_white_list "lookup_table" "lookup_table_v2" \
712+
--amp_master_grad 1 \
713+
--fuse_attention_ffn false \
714+
--fuse_attention_qkv false \
715+
--fuse_sequence_parallel_allreduce false \
716+
--use_flash_attention 0 \
717+
--use_fused_rope false \
718+
--use_fused_rms_norm 0 \
719+
--max_seq_length 4096 \
720+
--sep_parallel_degree 1 \
721+
--sequence_parallel true \
722+
--pipeline_parallel_degree 1 \
723+
--sharding_parallel_degree 1 \
724+
--tensor_parallel_degree 2 \
725+
--virtual_pp_degree 1 \
726+
--sharding "" \
727+
--to_static ${to_static} \
728+
--num_hidden_layers 4 \
729+
>>${log_path}/$FUNCNAME 2>&1
730+
loss=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss: ' '{print $2}' | awk -F ',' '{print $1}'`
731+
loss_md5=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'loss_md5: ' '{print $2}' | awk -F ',' '{print $1}'`
732+
ips=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'interval_tokens_per_second_per_device: ' '{print $2}' | awk -F ',' '{print $1}'`
733+
mem=`cat $case_log_dir/workerlog.0 | grep 'global_step: 10' | awk -F 'max_memory_reserved: ' '{print $2}' | awk -F ',' '{print $1}'`
734+
echo "result: to_static=$to_static loss=$loss ips=$ips mem=$mem"
735+
loss_base=9.16783295
736+
loss_md5_base=8ea72495fba4e1b9ba004b4431e27218
737+
if [ $IS_A100 -ne 0 ];then
738+
loss_base=9.38009949
739+
fi
740+
ips=-1
741+
mem=-1
742+
ips_base=-1
743+
mem_base=-1
744+
check_result $FUNCNAME ${loss_base} ${loss} ${ips_base} ${ips} ${mem_base} ${mem}
745+
# check_md5_result $FUNCNAME ${loss_md5_base} ${loss_md5}
746+
done
743747
done
744748
echo "=========== $FUNCNAME run end ==========="
745749
}

0 commit comments

Comments
 (0)