fix redundant log when using distributed training using pytorch

Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/10959078

modelscope/exporters/torch_model_exporter.py

@@ -17,7 +17,7 @@ from modelscope.utils.regress_test_utils import (compare_arguments_nested,
                                                  numpify_tensor_nested)
 from .base import Exporter
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class TorchModelExporter(Exporter):

modelscope/models/base/base_torch_head.py

@@ -6,7 +6,7 @@ import torch
 from modelscope.models.base.base_head import Head
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class TorchHead(Head, torch.nn.Module):

modelscope/models/base/base_torch_model.py

@@ -10,7 +10,7 @@ from modelscope.utils.hub import parse_label_mapping
 from modelscope.utils.logger import get_logger
 from .base_model import Model
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class TorchModel(Model, torch.nn.Module):

modelscope/models/multi_modal/mplug/configuration_mplug.py

@@ -23,7 +23,7 @@ from transformers.utils import logging
 
 from modelscope.utils.constant import Tasks
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class MPlugConfig(PretrainedConfig):

modelscope/models/multi_modal/mplug/modeling_mplug.py

@@ -46,7 +46,7 @@ from modelscope.utils.constant import ModelFile
 
 transformers.logging.set_verbosity_error()
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 CONFIG_NAME = 'config.yaml'
 

modelscope/models/multi_modal/ofa/configuration_mmspeech.py

@@ -17,7 +17,7 @@ import warnings
 from transformers import PretrainedConfig
 from transformers.utils import logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class MMSpeechConfig(PretrainedConfig):

modelscope/models/multi_modal/ofa/configuration_ofa.py

@@ -17,7 +17,7 @@ import warnings
 from transformers import PretrainedConfig
 from transformers.utils import logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 OFA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'ofa-medium': 'https://huggingface.co/ofa-base/resolve/main/config.json',

modelscope/models/multi_modal/ofa/modeling_mmspeech.py

@@ -44,7 +44,7 @@ from .generate import utils
 from .modeling_ofa import (Embedding, OFADecoder, OFAModel, OFAPreTrainedModel,
                            _expand_mask, shift_tokens_right)
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 _CHECKPOINT_FOR_DOC = 'mmspeech-base'
 _CONFIG_FOR_DOC = 'MMSpeechConfig'

modelscope/models/multi_modal/ofa/modeling_ofa.py

@@ -38,7 +38,7 @@ from .resnet import ResNet
 from .utils.utils import DropPath
 from .vit import vit_base, vit_huge, vit_large, vit_large_336
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 _CHECKPOINT_FOR_DOC = 'ofa-base'
 _CONFIG_FOR_DOC = 'OFAConfig'

modelscope/models/multi_modal/ofa/tokenization_ofa.py

@@ -24,7 +24,7 @@ from transformers.utils import logging
 
 from modelscope.utils.constant import ModelFile
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 VOCAB_FILES_NAMES = {'vocab_file': 'vocab.json', 'merges_file': 'merges.txt'}
 

modelscope/models/multi_modal/ofa/tokenization_ofa_fast.py

@@ -23,7 +23,7 @@ from transformers.utils import logging
 from modelscope.utils.constant import ModelFile
 from .tokenization_ofa import OFATokenizer, OFATokenizerZH
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 VOCAB_FILES_NAMES = {
     'vocab_file': 'vocab.json',

modelscope/models/nlp/T5/backbone.py

@@ -41,7 +41,7 @@ from modelscope.utils.constant import Tasks
 from modelscope.utils.logger import get_logger
 from .configuration import T5Config
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 ###################################################

modelscope/models/nlp/T5/configuration.py

@@ -20,7 +20,7 @@ from transformers.onnx import OnnxSeq2SeqConfigWithPast
 
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class T5Config(PretrainedConfig):

modelscope/models/nlp/T5/text2text_generation.py

@@ -31,7 +31,7 @@ from modelscope.utils.logger import get_logger
 from .backbone import T5PreTrainedModel, T5Stack
 from .configuration import T5Config
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 # Warning message for FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
 __HEAD_MASK_WARNING_MSG = """

modelscope/models/nlp/bert/backbone.py

@@ -36,7 +36,7 @@ from modelscope.utils.logger import get_logger
 from modelscope.utils.nlp.utils import parse_labels_in_order
 from .configuration import BertConfig
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 _CONFIG_FOR_DOC = 'BertConfig'
 

modelscope/models/nlp/bert/configuration.py

@@ -22,7 +22,7 @@ from transformers.onnx import OnnxConfig
 
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class BertConfig(PretrainedConfig):

modelscope/models/nlp/bert/fill_mask.py

@@ -28,7 +28,7 @@ from modelscope.utils.constant import Tasks
 from .backbone import BertModel, BertPreTrainedModel
 from .configuration import BertConfig
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class BertPredictionHeadTransform(nn.Module):

modelscope/models/nlp/bert/text_classification.py

@@ -27,7 +27,7 @@ from modelscope.utils import logger as logging
 from modelscope.utils.constant import Tasks
 from .backbone import BertModel, BertPreTrainedModel
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 @MODELS.register_module(Tasks.text_classification, module_name=Models.bert)

modelscope/models/nlp/bert/text_ranking.py

@@ -12,7 +12,7 @@ from modelscope.utils.constant import Tasks
 from .backbone import BertModel
 from .text_classification import BertForSequenceClassification
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 @MODELS.register_module(Tasks.text_ranking, module_name=Models.bert)

modelscope/models/nlp/bert/token_classification.py

@@ -27,7 +27,7 @@ from modelscope.utils import logger as logging
 from modelscope.utils.constant import Tasks
 from .backbone import BertModel, BertPreTrainedModel
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 @MODELS.register_module(Tasks.token_classification, module_name=Models.bert)

modelscope/models/nlp/deberta_v2/backbone.py

@@ -33,7 +33,7 @@ from modelscope.utils import logger as logging
 from modelscope.utils.constant import Tasks
 from .configuration import DebertaV2Config
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 # Copied from transformers.models.deberta.modeling_deberta.ContextPooler

modelscope/models/nlp/deberta_v2/configuration.py

@@ -18,7 +18,7 @@ from transformers import PretrainedConfig
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class DebertaV2Config(PretrainedConfig):

modelscope/models/nlp/deberta_v2/tokenization_fast.py

@@ -28,7 +28,7 @@ if is_sentencepiece_available():
 else:
     DebertaV2Tokenizer = None
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 VOCAB_FILES_NAMES = {
     'vocab_file': 'spm.model',

modelscope/models/nlp/gpt3/configuration.py

@@ -17,7 +17,7 @@ import torch
 from transformers.configuration_utils import PretrainedConfig
 from transformers.utils import logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class GPT3Config(PretrainedConfig):

modelscope/models/nlp/gpt_moe/configuration.py

@@ -17,7 +17,7 @@ import torch
 from transformers.configuration_utils import PretrainedConfig
 from transformers.utils import logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class GPTMoEConfig(PretrainedConfig):

modelscope/models/nlp/palm_v2/configuration.py

@@ -19,7 +19,7 @@ from transformers.configuration_utils import PretrainedConfig
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class PalmConfig(PretrainedConfig):

modelscope/models/nlp/palm_v2/text_generation.py

@@ -760,7 +760,7 @@ class Translator(object):
 
     def __init__(self, model, dataset: str = 'cnn'):
         super().__init__()
-        self.logger = logging.get_logger(__name__)
+        self.logger = logging.get_logger()
         self.args = model.config
         self.args.dataset = dataset
         self.model = model.palm

modelscope/models/nlp/plug/configuration.py

@@ -21,7 +21,7 @@ from transformers import PretrainedConfig
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class PlugNLUConfig(PretrainedConfig):

modelscope/models/nlp/plug/distributed_plug.py

@@ -17,7 +17,7 @@ from modelscope.utils.torch_utils import set_random_seed_mpu
 from . import PlugModel
 from .configuration import PlugNLGConfig
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class DistributedPlug(TorchModel):

modelscope/models/nlp/ponet/backbone.py

@@ -36,7 +36,7 @@ from modelscope.utils.constant import Tasks
 from modelscope.utils.logger import get_logger
 from .configuration import PoNetConfig
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 is_pytorch_12plus = LooseVersion(torch.__version__) >= LooseVersion('1.12.0')
 

modelscope/models/nlp/ponet/configuration.py

@@ -18,7 +18,7 @@ from transformers import PretrainedConfig
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class PoNetConfig(PretrainedConfig):

modelscope/models/nlp/ponet/fill_mask.py

@@ -26,7 +26,7 @@ from modelscope.utils.constant import Tasks
 from modelscope.utils.logger import get_logger
 from .backbone import PoNetModel, PoNetPreTrainedModel
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class PoNetPredictionHeadTransform(nn.Module):

modelscope/models/nlp/ponet/tokenization.py

@@ -24,7 +24,7 @@ from transformers.tokenization_utils import BatchEncoding, EncodedInput
 from modelscope.utils.constant import ModelFile
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 VOCAB_FILES_NAMES = {'vocab_file': ModelFile.VOCAB_FILE}
 

modelscope/models/nlp/space/configuration.py

@@ -20,7 +20,7 @@
 from modelscope.models.nlp.structbert import SbertConfig
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class SpaceConfig(SbertConfig):

modelscope/models/nlp/space/model/tokenization_space.py

@@ -19,7 +19,7 @@ from transformers import BasicTokenizer, BertTokenizer, WordpieceTokenizer
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class SpaceTokenizer(BertTokenizer):

modelscope/models/nlp/structbert/adv_utils.py

@@ -18,7 +18,7 @@ from torch import nn
 
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 def _symmetric_kl_div(logits1, logits2, attention_mask=None):

modelscope/models/nlp/structbert/backbone.py

@@ -39,7 +39,7 @@ from modelscope.utils.logger import get_logger
 from modelscope.utils.nlp.utils import parse_labels_in_order
 from .configuration import SbertConfig
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class SbertEmbeddings(nn.Module):

modelscope/models/nlp/structbert/configuration.py

@@ -19,7 +19,7 @@ from transformers import PretrainedConfig
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class SbertConfig(PretrainedConfig):

modelscope/models/nlp/structbert/fill_mask.py

@@ -29,7 +29,7 @@ from modelscope.utils.constant import Tasks
 from .backbone import SbertModel, SbertPreTrainedModel
 from .configuration import SbertConfig
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class SbertPredictionHeadTransform(nn.Module):

modelscope/models/nlp/structbert/text_classification.py

@@ -29,7 +29,7 @@ from .adv_utils import compute_adv_loss
 from .backbone import SbertModel, SbertPreTrainedModel
 from .configuration import SbertConfig
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 @MODELS.register_module(

modelscope/models/nlp/structbert/token_classification.py

@@ -29,7 +29,7 @@ from .adv_utils import compute_adv_loss
 from .backbone import SbertModel, SbertPreTrainedModel
 from .configuration import SbertConfig
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 @MODELS.register_module(

modelscope/models/nlp/task_models/task_model.py

@@ -15,7 +15,7 @@ from modelscope.utils.constant import Fields, Tasks
 from modelscope.utils.file_utils import func_receive_dict_inputs
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 __all__ = ['EncoderDecoderTaskModelBase', 'SingleBackboneTaskModelBase']
 

modelscope/models/nlp/veco/backbone.py

@@ -26,7 +26,7 @@ from modelscope.utils import logger as logging
 from modelscope.utils.constant import Tasks
 from .configuration import VecoConfig
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 VECO_PRETRAINED_MODEL_ARCHIVE_LIST = []
 

modelscope/models/nlp/veco/configuration.py

@@ -21,7 +21,7 @@ from transformers import RobertaConfig
 
 from modelscope.utils import logger as logging
 
-logger = logging.get_logger(__name__)
+logger = logging.get_logger()
 
 
 class VecoConfig(RobertaConfig):

modelscope/pipelines/nlp/text_classification_pipeline.py

@@ -12,7 +12,7 @@ from modelscope.preprocessors import Preprocessor
 from modelscope.utils.constant import Fields, Tasks
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 @PIPELINES.register_module(

modelscope/preprocessors/base.py

@@ -13,7 +13,7 @@ from modelscope.utils.hub import read_config, snapshot_download
 from modelscope.utils.logger import get_logger
 from .builder import build_preprocessor
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 PREPROCESSOR_MAP = {
     # nlp

modelscope/preprocessors/nlp/text_classification_preprocessor.py

@@ -14,7 +14,7 @@ from modelscope.utils.logger import get_logger
 from .transformers_tokenizer import NLPTokenizer
 from .utils import labels_to_id, parse_text_and_label
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class TextClassificationPreprocessorBase(Preprocessor):

modelscope/preprocessors/nlp/text_generation_preprocessor.py

@@ -15,7 +15,7 @@ from modelscope.utils.logger import get_logger
 from .transformers_tokenizer import NLPTokenizer
 from .utils import parse_text_and_label
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 class TextGenerationPreprocessorBase(Preprocessor):

modelscope/preprocessors/nlp/token_classification_preprocessor.py

@@ -16,7 +16,7 @@ from modelscope.utils.type_assert import type_assert
 from .transformers_tokenizer import NLPTokenizer
 from .utils import parse_text_and_label
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 @PREPROCESSORS.register_module(

modelscope/trainers/hooks/checkpoint_hook.py

@@ -70,7 +70,7 @@ class CheckpointHook(Hook):
             os.makedirs(self.save_dir)
 
         if not hasattr(trainer, 'logger'):
-            self.logger = get_logger(__name__)
+            self.logger = get_logger()
         else:
             self.logger = trainer.logger
 

modelscope/trainers/hooks/lr_scheduler_hook.py

@@ -99,7 +99,7 @@ class PlateauLrSchedulerHook(LrSchedulerHook):
     def before_run(self, trainer):
         super().before_run(trainer)
         if not hasattr(trainer, 'logger'):
-            self.logger = get_logger(__name__)
+            self.logger = get_logger()
         else:
             self.logger = trainer.logger
 

modelscope/trainers/optimizer/child_tuning_adamw_optimizer.py

@@ -24,7 +24,7 @@ from torch.optim import Optimizer
 from modelscope.utils.logger import get_logger
 from .builder import OPTIMIZERS, default_group
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 __all__ = ['calculate_fisher', 'ChildTuningAdamW']
 

modelscope/utils/checkpoint.py

@@ -18,7 +18,7 @@ from modelscope.utils.config import JSONIteratorEncoder
 from modelscope.utils.constant import ConfigFields, ModelFile
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 storage = LocalStorage()
 

modelscope/utils/hub.py

@@ -14,7 +14,7 @@ from modelscope.utils.constant import (DEFAULT_MODEL_REVISION, ConfigFields,
                                        ModelFile)
 from .logger import get_logger
 
-logger = get_logger(__name__)
+logger = get_logger()
 
 
 def create_model_if_not_exist(

modelscope/utils/import_utils.py

@@ -21,8 +21,6 @@ from modelscope.utils.ast_utils import (INDEX_KEY, MODULE_KEY, REQUIREMENT_KEY,
 from modelscope.utils.error import *  # noqa
 from modelscope.utils.logger import get_logger
 
-logger = get_logger(__name__)
-
 if sys.version_info < (3, 8):
     import importlib_metadata
 else:

modelscope/utils/logger.py

@@ -1,5 +1,6 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
+import importlib
 import logging
 from typing import Optional
 
@@ -24,11 +25,27 @@ def get_logger(log_file: Optional[str] = None,
     if logger_name in init_loggers:
         return logger
 
+    # handle duplicate logs to the console
+    # Starting in 1.8.0, PyTorch DDP attaches a StreamHandler <stderr> (NOTSET)
+    # to the root logger. As logger.propagate is True by default, this root
+    # level handler causes logging messages from rank>0 processes to
+    # unexpectedly show up on the console, creating much unwanted clutter.
+    # To fix this issue, we set the root logger's StreamHandler, if any, to log
+    # at the ERROR level.
+    for handler in logger.root.handlers:
+        if type(handler) is logging.StreamHandler:
+            handler.setLevel(logging.ERROR)
+
     stream_handler = logging.StreamHandler()
     handlers = [stream_handler]
 
-    # TODO @wenmeng.zwm add logger setting for distributed environment
-    if log_file is not None:
+    if importlib.util.find_spec('torch') is not None:
+        from modelscope.utils.torch_utils import is_master
+        is_worker0 = is_master()
+    else:
+        is_worker0 = True
+
+    if is_worker0 and log_file is not None:
         file_handler = logging.FileHandler(log_file, file_mode)
         handlers.append(file_handler)
 
@@ -39,7 +56,11 @@ def get_logger(log_file: Optional[str] = None,
         handler.setLevel(log_level)
         logger.addHandler(handler)
 
-    logger.setLevel(log_level)
+    if is_worker0:
+        logger.setLevel(log_level)
+    else:
+        logger.setLevel(logging.ERROR)
+
     init_loggers[logger_name] = True
 
     return logger

modelscope/utils/test_utils.py

@@ -230,6 +230,8 @@ class DistributedTestCase(unittest.TestCase):
         tmp_env = copy.deepcopy(os.environ)
         tmp_env['PYTHONPATH'] = ':'.join(
             (tmp_env.get('PYTHONPATH', ''), script_dir)).lstrip(':')
+        # avoid distributed test hang
+        tmp_env['NCCL_P2P_DISABLE'] = '1'
         script_params = '--save_all_ranks=%s --save_file=%s' % (save_all_ranks,
                                                                 tmp_res_file)
         script_cmd = '%s %s %s' % (dist_start_cmd, tmp_run_file, script_params)