Skip to content

GitLab

Unverified Commit 74983b36 authored by Nicolas Castet's avatar Nicolas Castet Committed by GitHub
Browse files

Fix UB with MPI init (#1538) 


Signed-off-by: default avatarNicolas Castet <ncastet@nvidia.com>
parent bd278fff
Show whitespace changes
Inline Side-by-side
Showing with 4 additions and 2 deletions
transformer_engine/common/comm_gemm_overlap/userbuffers/userbuffers-host.cpp
View file @ 74983b36
...@@ -280,7 +280,7 @@ int create_communicator_grouped2(communicator **comm, int myrank, int numranks, ...@@ -280,7 +280,7 @@ int create_communicator_grouped2(communicator **comm, int myrank, int numranks,
int fd; int fd;
volatile uint32_t abortFlag = 0; volatile uint32_t abortFlag = 0;
IpcSocketHandle ipcSock = {0}; IpcSocketHandle ipcSock = {0};
uint64_t opId = 0xdeadcafeb000 + (*comm)->my_node + (*comm)->ar2_firstgpu; uint64_t opId = 0xdeadcafe0000 + (*comm)->my_node;
ipcSocketResult_t ret = ipcSocketSuccess; ipcSocketResult_t ret = ipcSocketSuccess;
IPCCHECK(ipcSocketInit(&ipcSock, (*comm)->ar2_nvrank, (uint64_t)opId, &abortFlag)); IPCCHECK(ipcSocketInit(&ipcSock, (*comm)->ar2_nvrank, (uint64_t)opId, &abortFlag));
(*comm)->_barrier((*comm)->comm_world); (*comm)->_barrier((*comm)->comm_world);
...@@ -416,6 +416,8 @@ int create_communicator_grouped2_mpi(communicator **comm, int pipegpus, int pipe ...@@ -416,6 +416,8 @@ int create_communicator_grouped2_mpi(communicator **comm, int pipegpus, int pipe
// find internode numbers and make internode communicator // find internode numbers and make internode communicator
NVTE_CHECK_CUDA(cudaFree(0)); NVTE_CHECK_CUDA(cudaFree(0));
int mynode, numnodes; int mynode, numnodes;
mynode = myrank / numlocal;
numnodes = numranks / numlocal;
// finally call the abstracted constructor with MPI info // finally call the abstracted constructor with MPI info
return create_communicator_grouped2(comm, myrank, numranks, mylocal, numlocal, mynode, numnodes, return create_communicator_grouped2(comm, myrank, numranks, mylocal, numlocal, mynode, numnodes,
...@@ -549,7 +551,7 @@ int register_user_buffer_collective(void **gpubuff, size_t bytes, communicator * ...@@ -549,7 +551,7 @@ int register_user_buffer_collective(void **gpubuff, size_t bytes, communicator *
volatile uint32_t abortFlag = 0; volatile uint32_t abortFlag = 0;
IpcSocketHandle ipcSock = {0}; IpcSocketHandle ipcSock = {0};
uint64_t opId = 0xdeadcafebeef + comm->my_node; uint64_t opId = 0xdeadcafe0000 + comm->my_node;
ipcSocketResult_t ret = ipcSocketSuccess; ipcSocketResult_t ret = ipcSocketSuccess;
// All-gather POSIX file descriptors across local ranks // All-gather POSIX file descriptors across local ranks
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment