如何修复 google colab 上的 cuda 运行时错误?
Posted
技术标签:
【中文标题】如何修复 google colab 上的 cuda 运行时错误?【英文标题】:How can I fix cuda runtime error on google colab? 【发布时间】:2021-10-12 07:42:19 【问题描述】:我正在尝试使用 BERT 和 pytorch 在 Hugging Face 页面之后执行命名实体识别示例:Token Classification with W-NUT Emerging Entities。
*** 上有a related question,但错误信息与我的情况不同。
cuda runtime error (710) : device-side assert triggered at /pytorch/aten/src/THC/generic/THCTensorMath.cu:29
我无法修复上述 cuda 运行时错误。
如何使用 运行时类型 GPU 在 google colab 上执行示例代码?
错误
trainer.train()
# Error Message
/usr/local/lib/python3.7/dist-packages/torch/autograd/__init__.py in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)
147 Variable._execution_engine.run_backward(
148 tensors, grad_tensors_, retain_graph, create_graph, inputs,
--> 149 allow_unreachable=True, accumulate_grad=True) # allow_unreachable flag
150
151
RuntimeError: cuda runtime error (710) : device-side assert triggered at /pytorch/aten/src/THC/generic/THCTensorMath.cu:29
代码
教程中介绍的原始数据和代码我没有更改,Token Classification with W-NUT Emerging Entities。
使用 W-NUT Emerging Entities 代码从浏览器访问令牌分类: custom_datasets.ipynb - Colaboratory
from pathlib import Path
import re
def read_wnut(file_path):
file_path = Path(file_path)
raw_text = file_path.read_text().strip()
raw_docs = re.split(r'\n\t?\n', raw_text)
token_docs = []
tag_docs = []
for doc in raw_docs:
tokens = []
tags = []
for line in doc.split('\n'):
token, tag = line.split('\t')
tokens.append(token)
tags.append(tag)
token_docs.append(tokens)
tag_docs.append(tags)
return token_docs, tag_docs
texts, tags = read_wnut('wnut17train.conll')
from sklearn.model_selection import train_test_split
train_texts, val_texts, train_tags, val_tags = train_test_split(texts, tags, test_size=.2)
unique_tags = set(tag for doc in tags for tag in doc)
tag2id = tag: id for id, tag in enumerate(unique_tags)
id2tag = id: tag for tag, id in tag2id.items()
from transformers import DistilBertTokenizerFast
tokenizer = DistilBertTokenizerFast.from_pretrained('distilbert-base-cased')
train_encodings = tokenizer(train_texts, is_split_into_words=True, return_offsets_mapping=True, padding=True, truncation=True)
val_encodings = tokenizer(val_texts, is_split_into_words=True, return_offsets_mapping=True, padding=True, truncation=True)
import numpy as np
def encode_tags(tags, encodings):
labels = [[tag2id[tag] for tag in doc] for doc in tags]
encoded_labels = []
for doc_labels, doc_offset in zip(labels, encodings.offset_mapping):
# create an empty array of -100
doc_enc_labels = np.ones(len(doc_offset),dtype=int) * -100
arr_offset = np.array(doc_offset)
# set labels whose first offset position is 0 and the second is not 0
doc_enc_labels[(arr_offset[:,0] == 0) & (arr_offset[:,1] != 0)] = doc_labels
encoded_labels.append(doc_enc_labels.tolist())
return encoded_labels
train_labels = encode_tags(train_tags, train_encodings)
val_labels = encode_tags(val_tags, val_encodings)
import torch
import os
#os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
torch.backends.cudnn.enabled = False
# check if CUDA is available
train_on_gpu = torch.cuda.is_available()
# torch.backends.cudnn.enabled
if not train_on_gpu:
print('CUDA is not available. Training on CPU ...')
else:
print('CUDA is available! Training on GPU ...')
class WNUTDataset(torch.utils.data.Dataset):
def __init__(self, encodings, labels):
self.encodings = encodings
self.labels = labels
def __getitem__(self, idx):
item = key: torch.tensor(val[idx]) for key, val in self.encodings.items()
item['labels'] = torch.tensor(self.labels[idx])
return item
def __len__(self):
return len(self.labels)
train_encodings.pop("offset_mapping") # we don't want to pass this to the model
val_encodings.pop("offset_mapping")
train_dataset = WNUTDataset(train_encodings, train_labels)
val_dataset = WNUTDataset(val_encodings, val_labels)
from transformers import DistilBertForTokenClassification
model = DistilBertForTokenClassification.from_pretrained('distilbert-base-cased', num_labels=len(unique_tags))
from transformers import DistilBertForSequenceClassification, Trainer, TrainingArguments, DistilBertForTokenClassification
from sklearn.metrics import precision_recall_fscore_support
import tensorflow as tf
def compute_metrics(pred):
labels = pred.label_ids
preds = pred.predictions.argmax(-1)
precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')
acc = accuracy_score(labels, preds)
return
'accuracy': acc,
'f1': f1,
'precision': precision,
'recall': recall
training_args = TrainingArguments(
output_dir='./results', # output directory
num_train_epochs=3, # total number of training epochs
per_device_train_batch_size=16, # batch size per device during training
per_device_eval_batch_size=64, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
logging_dir='./logs', # directory for storing logs
logging_steps=10,
)
model = DistilBertForTokenClassification.from_pretrained("distilbert-base-uncased")
trainer = Trainer(
model=model, # the instantiated ???? Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=train_dataset, # training dataset
eval_dataset=val_dataset, # evaluation dataset
compute_metrics=compute_metrics
)
trainer.train()
我做了什么
我检查了 cuda 和 GPU 相关设置。
#os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
torch.backends.cudnn.enabled = False
# check if CUDA is available
train_on_gpu = torch.cuda.is_available()
# torch.backends.cudnn.enabled
if not train_on_gpu:
print('CUDA is not available. Training on CPU ...')
else:
print('CUDA is available! Training on GPU ...')
#output
CUDA is available! Training on GPU ...
training_args.device
#output
device(type='cuda', index=0)
回复答案
当我注释掉该部分时,
#os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
#torch.backends.cudnn.enabled = False
当我没有重置运行时,错误消息变成了下面。
/usr/local/lib/python3.7/dist-packages/torch/autograd/__init__.py in _make_grads(outputs, grads)
49 if out.numel() != 1:
50 raise RuntimeError("grad can be implicitly created only for scalar outputs")
---> 51 new_grads.append(torch.ones_like(out, memory_format=torch.preserve_format))
52 else:
53 new_grads.append(None)
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
如果我重置运行时,消息是一样的。
RuntimeError: cuda runtime error (710) : device-side assert triggered at /pytorch/aten/src/THC/generic/THCTensorMath.cu:29
【问题讨论】:
【参考方案1】:也许问题出在这一行:
torch.backends.cudnn.enabled = False
您可以评论或删除它,然后重试。
【讨论】:
感谢您的回答。添加了结果和可用的相同代码 custom_datasets.ipynb - Colaboratory,可从浏览器获得。 custom_datasets.ipynb - Colaboratory(colab.research.google.com/github/huggingface/notebooks/blob/…)以上是关于如何修复 google colab 上的 cuda 运行时错误?的主要内容,如果未能解决你的问题,请参考以下文章
如何修复 Colab 上的“错误:pytorch3d 构建***失败”错误?
在 Google Colab 上从 CUDA 11.2 降级到 11.1 或 10.2(找不到包问题)
不使用多处理但在使用 PyTorch DataLoader 时在 google colab 上出现 CUDA 错误
Google Colab Fine Tuning BERT Base Cased with Transformers and PyTorch 中出现间歇性“RuntimeError: CUDA out