PyTorch分布式训练初始化阶段出现SIGSEGV错误的原因咨询

错误日志 #

ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: -11) local_rank: 0 (pid: 3680358) of binary: /home/lifesci/ekeys/anaconda3/envs/minigpt4-4/bin/python
Traceback (most recent call last):
  File "/home/lifesci/ekeys/anaconda3/envs/minigpt4-4/bin/torchrun", line 33, in <module>
    sys.exit(load_entry_point('torch==2.0.1', 'console_scripts', 'torchrun')())
  File "/home/lifesci/ekeys/anaconda3/envs/minigpt4-4/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
    return f(*args, **kwargs)
  File "/home/lifesci/ekeys/anaconda3/envs/minigpt4-4/lib/python3.10/site-packages/torch/distributed/run.py", line 794, in main
    run(args)
  File "/home/lifesci/ekeys/anaconda3/envs/minigpt4-4/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
    elastic_launch(
  File "/home/lifesci/ekeys/anaconda3/envs/minigpt4-4/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/home/lifesci/ekeys/anaconda3/envs/minigpt4-4/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError: 
=========================================================
train.py FAILED
---------------------------------------------------------
Failures:
[1]:
  time      : 2023-10-21_17:36:57
  host      : gnode10.hanhai22.scc.ustc.edu.cn
  rank      : 1 (local_rank: 1)
  exitcode  : -11 (pid: 3680359)
  error_file: <N/A>
  traceback : Signal 11 (SIGSEGV) received by PID 3680359
[2]:
  time      : 2023-10-21_17:36:57
  host      : gnode10.hanhai22.scc.ustc.edu.cn
  rank      : 2 (local_rank: 2)
  exitcode  : -11 (pid: 3680360)
  error_file: <N/A>
  traceback : Signal 11 (SIGSEGV) received by PID 3680360
[3]:
  time      : 2023-10-21_17:36:57
  host      : gnode10.hanhai22.scc.ustc.edu.cn
  rank      : 3 (local_rank: 3)
  exitcode  : -11 (pid: 3680361)
  error_file: <N/A>
  traceback : Signal 11 (SIGSEGV) received by PID 3680361
---------------------------------------------------------
Root Cause (first observed failure):
[0]:
  time      : 2023-10-21_17:36:57
  host      : gnode10.hanhai22.scc.ustc.edu.cn
  rank      : 0 (local_rank: 0)
  exitcode  : -11 (pid: 3680358)
  error_file: <N/A>
  traceback : Signal 11 (SIGSEGV) received by PID 3680358
=========================================================

错误本质与常见原因 #

Exitcode -11对应SIGSEGV段错误，即进程尝试访问非法内存地址。在分布式训练初始化阶段触发这类错误，常见原因包括：

• 显存过载：多卡训练时，单卡分配的显存超过剩余容量，比如模型体积过大、batch size未按卡数缩放，初始化时瞬间占满显存导致崩溃。
• 初始化顺序错误：代码中在调用torch.distributed.init_process_group之前就执行了GPU操作（如模型移至GPU），多进程争抢GPU资源触发内存冲突。
• 依赖版本不兼容：即使切换了PyTorch和CUDA版本，若cuDNN、NCCL等配套库或GPU驱动与CUDA版本不匹配，底层CUDA操作会触发段错误。
• 系统资源冲突：GPU被其他进程占用，或服务器共享内存、PCIe链路存在资源竞争，导致分布式进程无法正常初始化CUDA上下文。
• 自定义算子问题：代码中使用的自定义CUDA算子或第三方库未做分布式兼容处理，多进程下出现内存访问错误。

排查步骤 #

• 单卡验证：先以单卡模式运行脚本，确认模型本身可正常初始化，排除基础代码问题。
• 降低显存负载：缩小batch size，或启用半精度训练（torch.float16），减少单卡显存占用后再尝试分布式。
• 修正初始化顺序：确保先执行分布式初始化，再绑定对应GPU，示例正确流程：

  import torch.distributed as dist
  dist.init_process_group(backend='nccl')
  local_rank = dist.get_rank()
  torch.cuda.set_device(local_rank)
  model = model.to(local_rank)
  

• 核对依赖版本：确认GPU驱动、CUDA、cuDNN、NCCL版本与当前PyTorch版本完全匹配。
• 清理GPU资源：用nvidia-smi查看GPU占用，终止其他占用进程后重启训练。
• 逐步扩容验证：先以torchrun --nproc_per_node=1启动单进程分布式训练，确认正常后再逐步增加进程数。

内容的提问来源于stack exchange，提问作者ekeys