[CRL] Add intra-node P2P transport and migrate to window-based API

[MC952-arch]

PR Category

CRL

PR Types

New Features

PR Description

This PR introduces a window-based, stream-integrated one-sided RMA API (intended to be NCCL-aligned) and adds an intra-node D2D “bypass” path in the hetero RMA proxy, while updating the host_api perf benchmarks and CI to exercise the new API.

[Copilot]

Pull request overview

This PR introduces a window-based, stream-integrated one-sided RMA API (intended to be NCCL-aligned) and adds an intra-node D2D “bypass” path in the hetero RMA proxy, while updating the host_api perf benchmarks and CI to exercise the new API.

Changes:

• Replaces the public flagcxPutSignal/flagcxSignal/flagcxWaitSignal API surface with new window/stream-based signatures and updates the host_api perf tests accordingly.
• Extends the hetero RMA proxy with GPU-visible completion sequence counters for stream-based flush/wait, plus initial scaffolding for intra-node IPC/D2D bypass.
• Adds a new RMA test script and a GitHub Actions workflow to build and run the perf-based RMA tests.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 13 comments.

Show a summary per file

File	Description
test/script/rma_test.sh	New script to run RMA perf binaries under MPI.
test/perf/host_api/test_put.cpp	Migrates put benchmark to window-based RMA API.
test/perf/host_api/test_one_side_register.cpp	Migrates multi-comm one-sided registration test to window-based API.
test/perf/host_api/test_get.cpp	Migrates get benchmark signaling/waiting to the new API.
flagcx/include/flagcx.h	Updates public one-sided API declarations to window/stream-based forms.
flagcx/flagcx.cc	Implements the new high-level one-sided APIs and WaitSignal logic.
flagcx/core/include/flagcx_hetero.h	Adds hetero proxy state for stream-based synchronization and IPC bypass.
flagcx/core/flagcx_hetero.cc	Implements stream-based flush and intra-node D2D bypass helpers.
.github/workflows/unittest-rma.yml	Adds CI job to build and run the RMA perf tests.

Comments suppressed due to low confidence (1)

test/perf/host_api/test_one_side_register.cpp:69

• Receiver waits with desc = {1, senderRank} but the current flagcxWaitSignal implementation flushes local outgoing opSeqs for that peer, so it won't reliably wait for the incoming PutSignal completion in this test.

    flagcxWaitSignalDesc_t desc = {1, senderRank};
    res = flagcxWaitSignal(1, &desc, comm, waitStream);
    if (res != flagcxSuccess) {
      fprintf(stderr, "[rank %d] flagcxWaitSignal failed (err=%d)\n", proc,
              int(res));
      return false;
    }
    devHandle->streamSynchronize(waitStream);

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 15 comments.

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 8 comments.

[Copilot]

Pull request overview

Copilot reviewed 15 out of 15 changed files in this pull request and generated 8 comments.

[Copilot]

Pull request overview

Copilot reviewed 15 out of 15 changed files in this pull request and generated 7 comments.

[Copilot]

Pull request overview

Copilot reviewed 22 out of 22 changed files in this pull request and generated 7 comments.