mhm0902 / flagperf Goto Github PK

export EXEC_IN_CONTAINER=True

# 先各服务器间root帐号的ssh信任关系和sudo免密配置
git clone https://github.com/FlagOpen/FlagPerf.git
cd FlagPerf/base/

cd FlagPerf/base/
vim configs/host.yaml

cd FlagPerf/base/
sudo python3 run.py

# 先各服务器间root帐号的ssh信任关系和sudo免密配置
git clone https://github.com/FlagOpen/FlagPerf.git
cd FlagPerf/training/
pip3 install -r requirements.txt

cd Flagperf/training/
vim run_benchmarks/config/cluster_conf.py

'''Cluster configs'''
#Hosts to run the benchmark. Each item is an IP address or a hostname.
HOSTS = ["10.1.2.3", "10.1.2.4", "10.1.2.5", "10.1.2.6"]
#ssh connection port
SSH_PORT = "22"

cd Flagperf/training/
vim run_benchmarks/config/test_conf.py

VENDOR = "nvidia" #选择本次运行的硬件
FLAGPERF_PATH="" # FlagPerf项目路径，如"/home/FlagPerf/training"
CASES={} # 本次运行的测例，按照对应模型readme准备好数据，修改模型对应的地址
#如运行"bert:pytorch_1.8:A100:1:8:1": "/raid/home_datasets_ckpt/bert/train/"，需要把:后面的路径替换为本地路径

python3 ./run_benchmarks/run.py
sudo python3 ./run_benchmarks/run.py

cd result/run2023XXXX/运行模型/
# ls
round1
# ls round1/
10.1.2.2_noderank0
# cd 10.1.2.2_noderank0/
# ls
cpu_monitor.log     pwr_monitor.log  rank2.out.log  rank5.out.log  start_pytorch_task.log
mem_monitor.log     rank0.out.log    rank3.out.log  rank6.out.log
nvidia_monitor.log  rank1.out.log    rank4.out.log  rank7.out.log


# tail -n 6 rank0.out.log
[PerfLog] {"event": "STEP_END", "value": {"loss": 2.679504871368408, "embedding_average": 0.916015625, "epoch": 1, "end_training": true, "global_steps": 3397, "num_trained_samples": 869632, "learning_rate": 0.000175375, "seq/s": 822.455385237589}, "metadata": {"file": "/workspace/flagperf/training/benchmarks/cpm/pytorch/run_pretraining.py", "lineno": 127, "time_ms": 1669034171032, "rank": 0}}
[PerfLog] {"event": "EVALUATE", "metadata": {"file": "/workspace/flagperf/training/benchmarks/cpm/pytorch/run_pretraining.py", "lineno": 127, "time_ms": 1669034171032, "rank": 0}}
[PerfLog] {"event": "EPOCH_END", "metadata": {"file": "/workspace/flagperf/training/benchmarks/cpm/pytorch/run_pretraining.py", "lineno": 127, "time_ms": 1669034171159, "rank": 0}}
[PerfLog] {"event": "TRAIN_END", "metadata": {"file": "/workspace/flagperf/training/benchmarks/cpm/pytorch/run_pretraining.py", "lineno": 136, "time_ms": 1669034171159, "rank": 0}}
[PerfLog] {"event": "FINISHED", "value": {"e2e_time": 1661.6114165782928, "training_sequences_per_second": 579.0933420700227, "converged": true, "final_loss": 3.066718101501465, "final_mlm_accuracy": 0.920166015625, "raw_train_time": 1501.713, "init_time": 148.937}, "metadata": {"file": "/workspace/flagperf/training/benchmarks/cpm/pytorch/run_pretraining.py", "lineno": 158, "time_ms": 1669034171646, "rank": 0}}

# 先各服务器间root帐号的ssh信任关系和sudo免密配置
git clone https://github.com/FlagOpen/FlagPerf.git
cd FlagPerf/inference/
pip3 install -r requirements.txt

cd Flagperf/inference/
vim configs/host.yaml

#必须修改项
FLAGPERF_PATH: "/home/FlagPerf/inference" #FlagPerf inference 路径
HOSTS: ["127.0.0.1"] # 机器地址
VENDOR = "nvidia" #测试机器对象，nvidia/kunlunxin/iluvatar
CASES:  #待测case，记得修改数据地址
    "resnet50:pytorch_1.13": "/raid/dataset/ImageNet/imagenet/val"

batch_size: 256
# 1 item(like 1 sequence, 1 image) flops
# Attention! For transformer decoder like bert, 1 token cause 2*param flops, so we need 2*length*params like 2*512*0.33B here
# format: a_1*a*2*...*a_nea_0,like 2*512*0.33e9(bert) or 4.12e9(resnet50)
flops: 4.12e9
fp16: true
compiler: tensorrt
num_workers: 8
log_freq: 30
repeat: 5
# skip validation(will also skip create_model, export onnx). Assert exist_onnx_path != null
no_validation: false
# set a real onnx_path to use exist, or set it to anything but null to avoid export onnx manually(like torch-tensorrt)
exist_onnx_path: null
# set a exist path of engine file like resnet50.trt/resnet50.plan/resnet50.engine
exist_compiler_path: null

VENDOR = "nvidia" #选择本次运行的硬件
FLAGPERF_PATH="" # FlagPerf项目路径，如"/home/FlagPerf/training"
CASES={} # 本次运行的测例，按照对应模型readme准备好数据，修改模型对应的地址
#如运行"bert:pytorch_1.8:A100:1:8:1": "/raid/home_datasets_ckpt/bert/train/"，需要把:后面的路径替换为本地路径

sudo python inference/run.py