Merge remote-tracking branch 'rccl/develop' into 2.19.4

[ROCm/rccl commit: 1a134b283b]
此提交包含在:
Wenkai Du
2024-01-31 11:08:54 -06:00
當前提交 4f4e52aafe
共有 14 個檔案被更改,包括 31 行新增 和 30 行删除
+2 -3
查看文件
@@ -270,11 +270,10 @@ static ncclResult_t mscclInternalSchedulerSelectAlgo(struct mscclSchedulerParam*
mscclStatus& status = mscclGetStatus();
param->scheduled = false;
/*// Current MSCCL doesn't support pre/post op
// Current MSCCL doesn't support pre/post op
if (param->op >= ncclAvg) {
return ncclSuccess;
}*/
}
// Whether the algorithm is in-place
bool isInPlace = false;
+2 -4
查看文件
@@ -332,12 +332,10 @@ static ncclResult_t hostToDevRedOp(
#define MSCCL_KERNEL_ENTRY() \
MSCCL_KERNEL_ENTRY_DEVREDOP(Sum, false), \
MSCCL_KERNEL_ENTRY_DEVREDOP(Prod, false), \
MSCCL_KERNEL_ENTRY_DEVREDOP(MinMax, false), \
MSCCL_KERNEL_ENTRY_DEVREDOP(PreMulSum, false), \
MSCCL_KERNEL_ENTRY_DEVREDOP_NOFLOAT(SumPostDiv, false)
MSCCL_KERNEL_ENTRY_DEVREDOP(MinMax, false)
// Except for ncclDevPreMulSum and ncclDevSumPostDiv required by ncclAvg
void* mscclKernelEntries[ncclNumDevRedOps * ncclNumTypes * NCCL_NUM_PROTOCOLS] = {
void* mscclKernelEntries[(ncclNumDevRedOps-2) * ncclNumTypes * NCCL_NUM_PROTOCOLS] = {
#ifdef COMPILE_MSCCL_KERNEL
MSCCL_KERNEL_ENTRY()
#endif