microsoft/mscclpp
1
0
Fork 0
mirror of https://github.com/microsoft/mscclpp.git synced 2026-04-20 06:49:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

PosixFx path falls through to RuntimeIpc when unix socket is unreachable (cross-node)

Browse Source
This commit is contained in:
Qinghua Zhou
2026-03-05 14:34:42 +00:00
parent cba13f83fc
commit 237302258d
1 changed files with 14 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/core/gpu_ipc_mem.cc
View File

@@ -274,13 +274,21 @@ GpuIpcMem::GpuIpcMem(const GpuIpcMemHandle& handle)
}
}
if ((type_ == GpuIpcMemHandle::Type::None) && (handle_.typeFlags & GpuIpcMemHandle::Type::PosixFd)) {
int fileDesc = UnixSocketClient::instance().requestFd(UnixSocketServer::generateSocketPath(handle_.posixFd.pid),
static_cast<uint32_t>(handle_.posixFd.fd));
if (cuMemImportFromShareableHandle(&allocHandle_, reinterpret_cast<void*>(fileDesc),
CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR) == CUDA_SUCCESS) {
type_ = GpuIpcMemHandle::Type::PosixFd;
// PosixFd uses a host-local unix domain socket to transfer the file descriptor.
// On GB200 NVL (cross-node NVLink domain), the peer process may be on a different
// host where the unix socket does not exist. Catch the failure and fall through to
// RuntimeIpc which works cross-node via cudaIpcMemHandle.
try {
int fileDesc = UnixSocketClient::instance().requestFd(UnixSocketServer::generateSocketPath(handle_.posixFd.pid),
static_cast<uint32_t>(handle_.posixFd.fd));
if (cuMemImportFromShareableHandle(&allocHandle_, reinterpret_cast<void*>(fileDesc),
CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR) == CUDA_SUCCESS) {
type_ = GpuIpcMemHandle::Type::PosixFd;
}
::close(fileDesc);
} catch (const SysError&) {
// Unix socket unreachable (cross-node) — fall through to RuntimeIpc
}
::close(fileDesc);
}
if ((type_ == GpuIpcMemHandle::Type::None) && (handle_.typeFlags & GpuIpcMemHandle::Type::RuntimeIpc)) {
type_ = GpuIpcMemHandle::Type::RuntimeIpc;