Skip to content

Fix the IB DMABUF path to be used from CUDA 12.8 #830

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
chhwang wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix the IB DMABUF path to be used from CUDA 12.8 #830

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
00ae6d6
Fix the IB DMABUF path to be used from CUDA 12.8
chhwang Jul 9, 2026
f0c7c10
Merge branch 'main' into chhwang/fix-dmabuf-guard
chhwang Jul 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions src/core/ib.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,15 +90,15 @@ IbMr::IbMr(ibv_pd* pd, void* buff, std::size_t size, bool isDataDirect) : mr_(nu
// bridge that reorders posted writes (e.g., Grace/GB200 NVLink-C2C), the PCIe mapping flag
// routes DMA through the Data Direct engine for correct ordering and higher throughput.
// Fall back to the default (non-PCIe) mapping if the flag is unsupported.
#if (CUDA_VERSION >= 12030)
#if (CUDA_VERSION >= 12080)
CUresult cuRes = cuMemGetHandleForAddressRange(&fd, addr, rangeSize, CU_MEM_RANGE_HANDLE_TYPE_DMA_BUF_FD,
CU_MEM_RANGE_FLAG_DMA_BUF_MAPPING_TYPE_PCIE);
if (cuRes != CUDA_SUCCESS || fd < 0) {
if (fd >= 0) ::close(fd);
fd = -1;
}
bool usedPcieFlag = (fd >= 0);
#endif // CUDA_VERSION >= 12030
#endif // CUDA_VERSION >= 12080
if (fd < 0) {
MSCCLPP_CUTHROW(cuMemGetHandleForAddressRange(&fd, addr, rangeSize, CU_MEM_RANGE_HANDLE_TYPE_DMA_BUF_FD, 0));
}
Expand All @@ -119,13 +119,13 @@ IbMr::IbMr(ibv_pd* pd, void* buff, std::size_t size, bool isDataDirect) : mr_(nu
}

// If MR registration failed with a PCIe-mapped fd, retry with the default mapping.
#if (CUDA_VERSION >= 12030)
#if (CUDA_VERSION >= 12080)
if (mr_ == nullptr && usedPcieFlag) {
::close(fd);
MSCCLPP_CUTHROW(cuMemGetHandleForAddressRange(&fd, addr, rangeSize, CU_MEM_RANGE_HANDLE_TYPE_DMA_BUF_FD, 0));
mr_ = IBVerbs::ibv_reg_dmabuf_mr(pd, offsetInDmaBuf, size, buffIntPtr, fd, accessFlags);
}
#endif // CUDA_VERSION >= 12030
#endif // CUDA_VERSION >= 12080

::close(fd);
if (mr_ == nullptr) {
Expand Down
Loading