RuntimeError: CUDA error: device-side assert triggered

诸神缄默不语-个人CSDN博文目录

我在运行PyG和transformers代码时两次遇到了这一问题，在此加以记录。

1. 运行PyG时遇到的错误

运行环境，报错信息和查找到的错误内容：
Linux系统
Python 3.8（使用anaconda管理的虚拟环境）
PyTorch 1.11+cudatoolkit 10.2（通过anaconda下载）
torch-scatter 2.0.9
torch-sparse 0.6.14
pyg-nightly 2.1.0.dev20220815

报错形式是重复多行（其中具体数字可能会产生改变）：

/opt/conda/conda-bld/pytorch_1646755853042/work/aten/src/ATen/native/cuda/Indexing.cu:703: indexSelectLargeIndex: block: [279,0,0], thread: [118,0,0] Assertion `srcIndex < srcSelectDimSize` failed.

最后给出的报错信息：

Traceback (most recent call last):
  File "try1.py", line 128, in 
    print(model(train_data.x_dict,train_data.edge_index_dict))
  File "env_path/lib/python3.8/site-packages/torch/fx/graph_module.py", line 630, in wrapped_call
    raise e.with_traceback(None)
RuntimeError: CUDA error: device-side assert triggered

这他娘的谁看得懂！

第一步的解决方案是把数据和模型从GPU上放到CPU上，然后再重新运行代码，就会给出正常的报错信息（而不是莫名其妙的信息）了。比如我的问题就是邻接矩阵（edge_index）中由于之前代码撰写的错误，出现了比节点数-1更大的索引：

Traceback (most recent call last):
  File "try1.py", line 146, in 
    print(model(train_data.x_dict,train_data.edge_index_dict))
  File "env_path/lib/python3.8/site-packages/torch/fx/graph_module.py", line 630, in wrapped_call
    raise e.with_traceback(None)
IndexError: index out of range in self

2. 运行transformers时遇到的错误

运行环境，报错信息和查找到的错误内容：
Linux系统
Python 3.8（使用anaconda管理的虚拟环境）
PyTorch 1.11+cudatoolkit 10.2（通过anaconda下载）
transformers 4.21.1

我是在使用AutoModelForSequenceClassification类，初始化的时候定义num_labels失误，导致数据中标签取值达到了这个数字，因此报错。
解决方式就是把num_labels改成正确的值。

报错信息：

Traceback (most recent call last):
  File "c1bert.py", line 109, in 
    optimizer.step()
  File "env_path/lib/python3.8/site-packages/torch/optim/lr_scheduler.py", line 65, in wrapper
    return wrapped(*args, **kwargs)
  File "env_path/lib/python3.8/site-packages/torch/optim/optimizer.py", line 88, in wrapper
    return func(*args, **kwargs)
  File "env_path/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "env_path/lib/python3.8/site-packages/torch/optim/adamw.py", line 145, in step
    F.adamw(params_with_grad,
  File "env_path/lib/python3.8/site-packages/torch/optim/_functional.py", line 155, in adamw
    param.addcdiv_(exp_avg, denom, value=-step_size)
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.
/opt/conda/conda-bld/pytorch_1646755853042/work/aten/src/ATen/native/cuda/Loss.cu:257: nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [4,0,0] Assertion `t >= 0 && t < n_classes` failed.