deepseek-ai / deepseek-ai/DeepEP
Timeout message is printed in barrier_block
- Dominant language
- Cuda
- Stars
- 10.1k
- Forks
- 1.4k
- Avg merge
- 4d 1h
- Merged PRs (30d)
- 2
Description
When we train MoE LLM by going with Megatron empowered by DeepEP for EP communication, Megatron will hang randomly with below error messages printed in the output log:
> DeepEP timeout check failed: rank = 6, thread = 0, value = 0)
> DeepEP timeout check failed: rank = 6, thread = 1, value = 0)
> DeepEP timeout check failed: rank = 6, thread = 2, value = 0)
> DeepEP timeout check failed: rank = 6, thread = 3, value = 0)
> DeepEP timeout check failed: rank = 6, thread = 4, value = 0)
> DeepEP timeout check failed: rank = 6, thread = 5, value = 1024)
> DeepEP timeout check failed: rank = 6, thread = 6, value = 0)
> DeepEP timeout check failed: rank = 6, thread = 7, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 0, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 1, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 2, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 3, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 4, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 5, value = 1024)
> DeepEP timeout check failed: rank = 2, thread = 6, value = 0)
> DeepEP timeout check failed: rank = 2, thread = 7, value = 0)
In this case, with CP enabled, the EP communiction group is unfolded along CP dimention.
By checking DeepEP code, we understood that this message is printed by barrier_block which used to sync up NVL ranks within a node. As shown in the above message, it looks the rank with local id 5 failed to go into the barrier_block.
We can duplicate this situation, within a 2-nodes environment by training a customized small LLM, when get a RNIC port down. But in our large-scaled cluster, we didn't observe the alarms of port up/down events when this problem happened.
In our cluster environment, NVSHMEM 3.3.9 is used.
It would be highly appreciated if any other clue or thought. Many thanks.
Contributor guide
No contributing guide indexed for this repository
Assessment
This issue has not been assessed yet.