deepseek-ai / deepseek-ai/DeepEP
Internode Combine Timeout
Open
- Dominant language
- Cuda
- Stars
- 10.1k
- Forks
- 1.4k
- Avg merge
- 4d 1h
- Merged PRs (30d)
- 2
Description
Hello! I found under this configuration, the **internode combine** will timeout: (The variable names I use here are the same as the ones used in `test_internode.py`)
**num_sms = 48
nvl_chunk_size = 32
nvl_buffer_size = 80
rdma_chunk_size = 20
rdma_buffer_size = 128**
For my experiment setup:
**Two nodes, each with 8 * H20 GPU and 4 * NIC, ROCE**
Could you try the config in your environment to confirm or deny this. Thanks a lot!
Contributor guide
No contributing guide indexed for this repository
Assessment
This issue has not been assessed yet.