Colab GPU T4运行时导入torchcodec崩溃问题求助

AI 小北

2026-6-12

在Colab T4 GPU运行时环境导入torchcodec导致运行时崩溃

问题详情

在Colab的GPU T4运行时环境中，仅执行import torchcodec就会导致运行时崩溃。原本在训练ASR模型过程中出现崩溃，逐步排查后，仅保留两行代码时，运行时在执行print语句前就已经崩溃。

测试代码

import torchcodec
#print("torchcodec imported successfully.")

运行时日志（含std::bad_alloc错误）

Oct 27, 2025, 4:24:43 PM    WARNING 0.00s - Note: Debugging will proceed. Set PYDEVD_DISABLE_FILE_VALIDATION=1 to disable this validation.
Oct 27, 2025, 4:24:43 PM    WARNING 0.00s - to python to disable frozen modules.
Oct 27, 2025, 4:24:43 PM    WARNING 0.00s - make the debugger miss breakpoints. Please pass -Xfrozen_modules=off
Oct 27, 2025, 4:24:43 PM    WARNING 0.00s - Debugger warning: It seems that frozen modules are being used, which may
Oct 27, 2025, 4:24:43 PM    WARNING kernel dfdda38e-6d14-4232-9b87-2445bf186294 restarted
Oct 27, 2025, 4:24:43 PM    INFO    AsyncIOLoopKernelRestarter: restarting kernel (1/5), keep random ports
Oct 27, 2025, 4:24:42 PM    WARNING what(): std::bad_alloc
Oct 27, 2025, 4:24:42 PM    WARNING terminate called after throwing an instance of 'std::bad_alloc'
Oct 27, 2025, 4:04:44 PM    WARNING 0.00s - Note: Debugging will proceed. Set PYDEVD_DISABLE_FILE_VALIDATION=1 to disable this validation.
Oct 27, 2025, 4:04:44 PM    WARNING 0.00s - to python to disable frozen modules.

解决建议

检查版本兼容性：确认torchcodec版本与Colab预装的PyTorch、CUDA版本匹配，版本不匹配易引发底层内存分配错误。
重新安装模块：卸载现有版本后从官方源重新安装，避免安装损坏：
```
pip uninstall -y torchcodec
pip install torchcodec --no-cache-dir
```
限制显存占用：导入torchcodec前设置PyTorch内存分配参数，避免一次性占用过多显存：
```
import torch
torch.cuda.set_per_process_memory_fraction(0.8)
import torchcodec
```
切换运行时环境：尝试切换到Colab CPU运行时验证是否为GPU环境特有问题，或调整PyTorch版本后再测试。