NVIDIA / NVIDIA/apex

Segmentation fault (core dumped) and RuntimeError: cuda runtime error (98) : unrecognized error code at apex/contrib/csrc/optimizers/fused_adam_cuda_kernel.cu:226

Open
#537 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
9k
Forks
1.5k
Avg merge
2d 4h
Merged PRs (30d)
3

Description

THCudaCheck FAIL file=apex/contrib/csrc/optimizers/fused_adam_cuda_kernel.cu line=226 error=98 : unrecognized error code

Traceback (most recent call last):
  File "finetune.py", line 261, in <module>
    optimizer.step()
  File "/home/ai/anaconda3/lib/python3.6/site-packages/apex/contrib/optimizers/fp16_optimizer.py", line 154, in step
    grad_norms=norm_groups)
  File "/home/ai/anaconda3/lib/python3.6/site-packages/apex/contrib/optimizers/fused_adam.py", line 180, in step
    group['weight_decay'])
RuntimeError: cuda runtime error (98) : unrecognized error code at apex/contrib/csrc/optimizers/fused_adam_cuda_kernel.cu:226

If you suspect this is an IPython bug, please report it at:
    https://github.com/ipython/ipython/issues
or send an email to the mailing list at ipython-dev@python.org

You can print a more detailed traceback right now with "%tb", or use "%debug"
to interactively debug it.

Extra-detailed tracebacks for bug-reporting purposes can be enabled via:
    %config Application.verbose_crash=True

Segmentation fault (core dumped)

I have to use the completely FP16 training for one model with multi-gpu, but amp failed to use with DataParal in mode 'O2' or 'O3'. And I have no choice but to use FP16_Optimizer in apex.contrib. When I use FusedLayerNorm, it calls the segmentation fault too.

environment:
cuda 10.0
pytorch 1.2.0
gcc 5.4.0

The apex is installed with

pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext"  --global-option="--deprecated_fused_adam" ./

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the failure in apex/contrib/csrc/optimizers/fused_adam_cuda_kernel.cu at line 226 using the reported CUDA 10.0, PyTorch 1.2.0, and build flags. Then investigate the FP16_Optimizer and FusedLayerNorm paths mentioned in the traceback, including multi-GPU DataParallel. Done means identifying and resolving the crash or documenting a confirmed compatibility cause.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.