Edgar
0336ffdf70
This is a single commit of the source code changes required to introduce support for multiple ranks per device. A new interface (ncclCommRankInitMulti) has to be used to make use of this new feature.
14 lignes
789 B
Bash
Fichiers exécutables
14 lignes
789 B
Bash
Fichiers exécutables
#!/bin/bash
|
|
export MPI_INSTALL_PATH=
|
|
export RCCL_INSTALL_PATH=
|
|
export ROCM_INSTALL_PATH=
|
|
|
|
$MPI_INSTALL_PATH/bin/mpiCC -o rccl-allreduce-mpi-multirank rccl-allreduce-mpi-multirank.cc -I$ROCM_INSTALL_PATH/include -I$RCCL_INSTALL_PATH/include -D__HIP_PLATFORM_AMD__ -L$ROCM_INSTALL_PATH/lib -lamdhip64 -L$RCCL_INSTALL_PATH/lib -lrccl
|
|
|
|
$MPI_INSTALL_PATH/bin/mpiCC -o rccl-reducescatter-mpi-multirank rccl-reducescatter-mpi-multirank.cc -I$ROCM_INSTALL_PATH/include -I$RCCL_INSTALL_PATH/include -D__HIP_PLATFORM_AMD__ -L$ROCM_INSTALL_PATH/lib -lamdhip64 -L$RCCL_INSTALL_PATH/lib -lrccl
|
|
|
|
|
|
export GPU_MAX_HW_QUEUES=16
|
|
$MPI_INSTALL_PATH/bin/mpirun --mca pml ucx -np 4 ./rccl-allreduce-mpi-multirank 0 0 2
|
|
$MPI_INSTALL_PATH/bin/mpirun --mca pml ucx -np 4 ./rccl-reducescatter-mpi-multirank 0 0 2
|