SFT codes

[PinxueGuo]

请问现在版本的SFT codes是测试无误可以运行的吗？

在src/open-r1-multimodal/src/open_r1/sft.py中 的L48-L49会报error：

from open_r1.configs import SFTConfig
from open_r1.utils.callbacks import get_callbacks
确实没有这些module文件，我从Open-R1中抄codes过来解决这个报错后，还是会卡住无法训练，在第一个step时一直保持状态不动但不进行运算：

[HaozheZhao]

Hi, thanks for reporting the issue. I apologize for the inconvenience.

In #77 , I removed the two redundant import statements from src/open-r1-multimodal/src/open_r1/sft.py (lines 48–49) and made the necessary adjustments to the code. With these changes, the training process should now run correctly with a one-click start.

Please update to the latest version after we merge the PR and let me know if you encounter any further issues. Thank you again for your feedback!

[PinxueGuo]

感谢迅速的回应和更新。我对齐你PR代码后，仍然只是解决了前者报错，依然会卡住最后timeout报错：

  0%|                                                              | 0/14166 [00:00<?, ?it/s]
[rank2]:[E212 15:34:10.420036109 ProcessGroupNCCL.cpp:1484] [PG ID 0 PG GUID 0(default_pg) Rank 2] ProcessGroupNCCL's watchdog got stuck for 480 seconds without making progress in monitoring enqueued collectives. This typically indicates a NCCL/CUDA API (e.g., CudaEventDestroy) hang blocking the watchdog, and could be triggered by another thread holding the GIL inside a CUDA api (for example, CudaEventDestroy), or other deadlock-prone behaviors.If you suspect the watchdog is not actually stuck and a longer timeout would help, you can either increase the timeout (TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC) to a larger value or disable the heartbeat monitor (TORCH_NCCL_ENABLE_MONITORING=0).If either of aforementioned helps, feel free to file an issue to PyTorch about the short timeout or false positive abort; otherwise, please attempt to debug the hang.                                                                    
[rank2]:[F212 15:42:10.426977111 ProcessGroupNCCL.cpp:1306] [PG ID 0 PG GUID 0(default_pg) Rank 2] [PG ID 0 PG GUID 0(default_pg) Rank 2] Terminating the process after attempting to dump debug info, due to ProcessGroup NCCL watchdog hang.

或许可以上传一份你的env version like conda/pip list?

[PinxueGuo]

New hint.

1. 用2B模型 1gpu 1bs_per_device时 可以正常训练。
2. 2B 8gpu 1bs_per_device 依然会卡住（gpu mem没有跑满）。
3. 2B 1gpu 4bs_per_device 会GPU OOM。

怀疑可能是哪里导致的gpu mem占用过大导致？请问您有采取什么会影响gpu mem占用的设置吗？（zero3也试过 用处不大 只有在 第3种情况时有点区别 会跑几个step后再OOM）

[LiuRicky]

同样的问题，SFT时会在第一个steps卡住。
Details：8 gpu，bsz 2，显存非常有盈余。

[LiuRicky]

Dear All,
I fix this issue at #84

[PinxueGuo]

pip install trl==0.14.0 instead of pip install from latest github(trl-0.15.0.dev0) as in open-r1 setup script, solves this problem

[PinxueGuo]

Dear All, I fix this issue at #84

哈哈 🤝

[cnyvfang]

是不是只有这个版本行得通😂，我发现我使用了更新版本的trl也有多node时在第一个batch卡住然后nccl timeout的问题