Improve MSCCL algorithms (#1023)

[ROCm/rccl commit: 0a53077c9c]
This commit is contained in:
Ziyue Yang
2024-01-04 06:51:34 +08:00
committed by GitHub
parent 0c1f773021
commit e3d45f9de4
11 changed files with 8601 additions and 6 deletions
+2 -3
View File
@@ -408,11 +408,10 @@ ncclResult_t mscclSetupKernel(const void* sendBuff, void* recvBuff, size_t count
uint8_t fullOpMask = (1<<MSCCL_RECV_COPY_SEND) |
(1<<MSCCL_RECV_REDUCE_SEND) |
(1<<MSCCL_RECV_REDUCE_COPY_SEND) |
(1<<MSCCL_RECV_REDUCE_COPY) |
(1<<MSCCL_LOCAL_COPY);
(1<<MSCCL_RECV_REDUCE_COPY);
//check if need full ops msccl kernel
if ((hostAlgo->typeMask & fullOpMask) || rcclParamMscclForceFullOps()) {
WARN("MSCCL: this version of MSCCL build doesn't support fill Ops");
WARN("MSCCL: this version of MSCCL build doesn't support full Ops");
return ncclInternalError;
}