Fix UDS connection failure when using ncclCommSplit. Issue #1185


[ROCm/rccl commit: 48bb7fec79]
This commit is contained in:
Sylvain Jeaugey committed 2024-02-26 02:52:39 -08:00
1 orang tua 2ab8a3a750
melakukan 5fa796af0c
6 files changed
+19 -10

No files matched your search

+4 -2
Melihat File
@@ -1660,6 +1660,9 @@ ncclResult_t ncclProxyInit(struct ncclComm* comm, struct ncclSocket* sock, union
comm->proxyState->listenSock = sock;
comm->proxyState->peerAddresses = peerAddresses;
comm->proxyState->peerAddressesUDS = peerAddressesUDS;
// UDS support
NCCLCHECK(ncclIpcSocketInit(&comm->proxyState->ipcSock, comm->rank, peerAddressesUDS[comm->rank], comm->abortFlag));
// Seed the random number generator for UDS filename generation
struct timeval time;
gettimeofday(&time,NULL);
@@ -1693,8 +1696,7 @@ ncclResult_t ncclProxyCreate(struct ncclComm* comm) {
ncclSetThreadName(comm->proxyState->thread, "NCCL Service %2d", comm->cudaDev);
// UDS support
INFO(NCCL_PROXY, "UDS: Creating service thread comm %p rank %d pidHash %lx", comm, comm->rank, comm->peerInfo[comm->rank].pidHash);
NCCLCHECK(ncclIpcSocketInit(&comm->proxyState->ipcSock, comm->rank, comm->peerInfo[comm->rank].pidHash, comm->abortFlag));
INFO(NCCL_PROXY, "UDS: Creating service thread comm %p rank %d", comm, comm->rank);
pthread_create(&comm->proxyState->threadUDS, NULL, ncclProxyServiceUDS, comm->proxyState);
ncclSetThreadName(comm->proxyState->threadUDS, "NCCL UDS Service %2d", comm->cudaDev);
}