Merge remote-tracking branch 'nccl/master' into develop

This commit is contained in:
Marzieh Berenjkoub
2026-01-20 13:01:49 -06:00
240 changed files with 16266 additions and 3578 deletions
+2 -2
View File
@@ -430,7 +430,7 @@ void Replayer::replay()
// no op or root
case rrAllToAll:
{
NCCL_CALL(ncclAllToAll(sbuffer, rbuffer, call.count, call.datatype, commMap[call.comm], streams[call.stream].first));
NCCL_CALL(ncclAlltoAll(sbuffer, rbuffer, call.count, call.datatype, commMap[call.comm], streams[call.stream].first));
break;
}
case rrAllGather:
@@ -506,7 +506,7 @@ void Replayer::replay()
log.read((char*)recvcounts.data(), size * sizeof(size_t));
log.read((char*)rdispls.data(), size * sizeof(size_t));
NCCL_CALL(ncclAllToAllv(sbuffer, sendcounts.data(), sdispls.data(), rbuffer, recvcounts.data(), rdispls.data(),
NCCL_CALL(ncclAlltoAllv(sbuffer, sendcounts.data(), sdispls.data(), rbuffer, recvcounts.data(), rdispls.data(),
call.datatype, commMap[call.comm], streams[call.stream].first));
HIP_CALL(hipStreamSynchronize(streams[call.stream].first)); // TODO: remove
break;
+2 -2
View File
@@ -303,9 +303,9 @@ ncclResult_t pncclScatter(const void* sendbuff, void* recvbuff,
*
* In-place operation will happen if sendbuff == recvbuff.
*/
ncclResult_t ncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
ncclResult_t ncclAlltoAll(const void* sendbuff, void* recvbuff, size_t count,
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
ncclResult_t pncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
ncclResult_t pncclAlltoAll(const void* sendbuff, void* recvbuff, size_t count,
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
/*
+4 -4
View File
@@ -787,10 +787,10 @@ ncclResult_t pncclScatter(const void* sendbuff, void* recvbuff,
@param[in] datatype Data buffer element datatype
@param[in] comm Communicator group object to execute on
@param[in] stream HIP stream to execute collective on */
ncclResult_t ncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
ncclResult_t ncclAlltoAll(const void* sendbuff, void* recvbuff, size_t count,
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
/*! @cond include_hidden */
ncclResult_t pncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
ncclResult_t pncclAlltoAll(const void* sendbuff, void* recvbuff, size_t count,
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
/*! @endcond */
@@ -812,11 +812,11 @@ ncclResult_t pncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
@param[in] datatype Data buffer element datatype
@param[in] comm Communicator group object to execute on
@param[in] stream HIP stream to execute collective on */
ncclResult_t ncclAllToAllv(const void *sendbuff, const size_t sendcounts[],
ncclResult_t ncclAlltoAllv(const void *sendbuff, const size_t sendcounts[],
const size_t sdispls[], void *recvbuff, const size_t recvcounts[],
const size_t rdispls[], ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
/*! @cond include_hidden */
ncclResult_t pncclAllToAllv(const void *sendbuff, const size_t sendcounts[],
ncclResult_t pncclAlltoAllv(const void *sendbuff, const size_t sendcounts[],
const size_t sdispls[], void *recvbuff, const size_t recvcounts[],
const size_t rdispls[], ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
/*! @endcond */
+1 -1
View File
@@ -32,7 +32,7 @@
#include "utils.h"
#include "rocm_smi/rocm_smi.h"
const char* ncclFuncStr[NCCL_NUM_FUNCTIONS+2] = { "Broadcast", "Reduce", "AllGather", "ReduceScatter", "AllReduce", "SendRecv", "AllToAllPivot" };
const char* ncclFuncStr[NCCL_NUM_FUNCTIONS+2] = { "Broadcast", "Reduce", "AllGather", "ReduceScatter", "AllReduce", "SendRecv", "AlltoAllPivot" };
const char* ncclAlgoStr[NCCL_NUM_ALGORITHMS] = { "Tree", "Ring", "CollNetDirect", "CollNetChain", "NVLS", "NVLSTree", "PAT" };
const char* ncclProtoStr[NCCL_NUM_PROTOCOLS] = { "LL", "LL128", "Simple" };