AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.8.4-1

Browse Source 
Fix hang in corner cases of alltoallv using point to point send/recv.
Harmonize error messages.
Fix missing NVTX section in the license.
Update README.
This commit is contained in:
Sylvain Jeaugey
2021-02-09 15:34:08 -08:00
parent 3996562690
commit 911d61f214
19 changed files with 53 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/bootstrap.cc
+3 -3
View File
@@ -72,7 +72,7 @@ static ncclResult_t bootstrapNetRecv(int fd, void* data, int size) {
int recvSize;
NCCLCHECK(socketRecv(fd, &recvSize, sizeof(int)));
if (recvSize > size) {
WARN("Message truncated : received %d bytes instead of %d\n", recvSize, size);
WARN("Message truncated : received %d bytes instead of %d", recvSize, size);
return ncclInternalError;
}
NCCLCHECK(socketRecv(fd, data, std::min(recvSize, size)));
@@ -244,7 +244,7 @@ static ncclResult_t remoteAlloc(void** ptr, int fd) {
void* ncclRemoteMemAllocationService(void* args) {
struct remAllocState* state = (struct remAllocState *) args;
if (cudaSetDevice(state->cudaDev) != cudaSuccess) {
WARN("[Rem Allocator] Failed to set CUDA device %d\n", state->cudaDev);
WARN("[Rem Allocator] Failed to set CUDA device %d", state->cudaDev);
}
// Prepare poll descriptor
@@ -490,7 +490,7 @@ ncclResult_t bootstrapRecv(void* commState, int peer, void* data, int size) {
ncclResult_t bootstrapClose(void* commState) {
struct extState* state = (struct extState*)commState;
if (state->unexpectedConnections != NULL) {
WARN("Unexpected connections are not empty.\n");
WARN("Unexpected connections are not empty");
return ncclInternalError;
}
close(state->extListenFd);