[Issue]: user-buffer p2p + coll rmtAddr nullptr error

[visualxu]

How is this issue impacting you?

Application crash

Share Your Debug Logs

No response

Steps to Reproduce the Issue

Hi, I found that rmtAddr may be nullptr when ncclSend/Recv and ncclCollectives (such as ncclAllreduce/Allgather...) are called in sequence(using the same user-buffer).
The reason is: the peerRmtAddrs in IPC_COLLECTIVES and P2P are different (deviceMemory or hostMemory), so when I call p2p first and then collectives, the variable needUpdate is false and peerRmtAddrs is not filled.
https://github.com/NVIDIA/nccl/blob/master/src/transport/p2p.cc#L911
https://github.com/NVIDIA/nccl/blob/master/src/transport/p2p.cc#L927
Fix suggestion: When regRecord->regIpcAddrs.devPeerRmtAddrs is first allocated, hostRmtAddrs need to be copied to devPeerRmtAddrs.

Thanks.

NCCL Version

v2.24.3-1 cuda12

Your platform details

No response

Error Message & Behavior

No response

[stephenmsachs]

Thanks a lot for reporting this issue. We are looking into it and come back to you soon.