Repository navigation
Conversation
Jinyao618
marked this pull request as draft
November 4, 2026 01:19
Jinyao618
force-pushed
the
feat/rdma-transport
branch
3 times, most recently
from
November 9, 2026 07:49
179539b to
91eb71e
Compare
Jinyao618
force-pushed
the
feat/rdma-transport
branch
2 times, most recently
from
November 16, 2026 08:15
4a8c729 to
1960ed5
Compare
Jinyao618
force-pushed
the
feat/rdma-transport
branch
from
November 17, 2026 03:02
1960ed5 to
9faa1ef
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Add an experimental
WITH_RDMAtransport attribute for the MPI backend. When enabled, a two-rank communicator initializes a native verbs RC QP over the existing MPI control channel. Device-memoryinfinicclSend/infinicclRecvuses direct RDMA when both peers can register the buffer; all unsupported cases automatically fall back to the existing MPI host-staging path.Changes
MPI integration
src/backends/mpi/ompi/comm_instance.h: attach anRdmaTransportinstance to the MPI communicator.src/backends/mpi/ompi/impl/comm_init_all.h: initialize and collectively negotiate RDMA availability during communicator setup.src/backends/mpi/ompi/impl/send.h,src/backends/mpi/ompi/impl/recv.h: attempt the RDMA path before the existing host-staging path and preserve the original fallback behavior.RDMA transport
src/backends/mpi/rdma/rdma_transport.{h,cc}: select an HCA/port, validate IB/RoCE port state, exchange peer metadata over MPI, create an RC QP, negotiate the per-message path, post send/receive work requests, validate completions, and tear down resources safely.mlx5dvwhen that capability is available; otherwise Mellanox QPs are rejected conservatively.Device-memory adapters
src/devices/nvidia/rdma_memory.{h,cc}: register current-device CUDA memory through NVIDIA peer-memory, validate allocation ranges and GPUDirect write-ordering requirements, and flush receives when required.src/devices/metax/rdma_memory.{h,cc}: export and register MACA device memory through DMA-BUF, guarded byINFINICCL_RDMA_METAX_EXPERIMENTAL=1.src/devices/cpu/rdma_memory.{h,cc}: provide the CPU adapter shape while intentionally selecting host staging.src/backends/mpi/rdma/memory_adaptor.{h,cc}: define the adapter interface, memory-region ownership, device dispatch, and an LRU MR cache capped at 256 MiB.Build and documentation
WITH_RDMAdetection and build wiring inCMakeLists.txtandsrc/CMakeLists.txt.README.mdPlatform and Backend Affected
Platform
Backend
Performance Impact
If applicable, provide benchmark results.
Known Issues & Future Work
INFINICCL_RDMA_METAX_EXPERIMENTAL=1until cache-visibility semantics are fully validated.Test Results
Test Involved Platform
Test Involved Backend
metax-rdma-send-recv.log
nvidia-metax-rdma-send-recv.log
nvidia-rdma-send-recv.log
Checklist
Title, Branch, and Commits
feat: …,fix(nccl): …).<type>/xxx-yyyy-zzzzwhere<type>matches the PR title's Conventional Commits type and words are joined with hyphens (seeCONTRIBUTING.md§Branches).CONTRIBUTING.md§Pull Requests).master— the branch is rebased cleanly on top of the currentmaster.fixup!/squash!/wipcommits remain.Scope and Design
CONTRIBUTING.md§Code/General).printf/std::cout/print(...)left behind, orTODOwithout an owner and issue link.General Code Hygiene
CONTRIBUTING.md§Code/General).CONTRIBUTING.md§Code/General).the `AllReduce` implementation) (CONTRIBUTING.md§Code/General).CONTRIBUTING.md§Code/General).CONTRIBUTING.md§Code/General; §Python).C++ Specific (if C++ files changed)
clang-format(version 16, per.github/workflows/clang-format.yml) has been run against all modified applicable files; the diff is clean.assertwith messages that include at least__FILE__,__LINE__, and__func__(CONTRIBUTING.md§C++).CONTRIBUTING.md§C++).CONTRIBUTING.md§C++).CONTRIBUTING.md§C++).CONTRIBUTING.md§C++).CONTRIBUTING.md§C++).Python Specific (if Python files changed)
ruff checkpasses cleanly on CI (see.github/workflows/ruff.yml).ruff format --checkpasses cleanly — if not, runruff formatand commit the result.CONTRIBUTING.md§Python).pytest.skipmessages without terminal period) are honored where applicable (CONTRIBUTING.md§Python).CONTRIBUTING.md§Python).if,for, and similar control-flow statements (CONTRIBUTING.md§Python).return, except when it directly follows a control-flow statement (CONTRIBUTING.md§Python).CONTRIBUTING.md§Python).Testing
Build, CI, and Tooling
clang-format.yml,ruff.yml) are green locally (or expected to be green on CI).Documentation
README.md,CONTRIBUTING.md, or inline docs updated when behavior, build flags, or developer workflow changed.!orBREAKING CHANGE:footer.Security and Safety