[dlamd1dai]

Motivation

Follow-up to #11569 / AICOMRCCL-2339. ginAnvilRegisterLsaSignals allgathers each rank's LSA signal base through bootstrapAllGather, which writes one slot per world rank (bootstrapState::nranks). The buffer was ctx->nRanks long and the local VA was stored at ctx->rank. Under NCCL_GIN_CONNECTION_RAIL those are the GIN team, so the ring writes past the allocation and publishes this rank's VA in the wrong slot.

Technical Details

• Allocate comm->nRanks slots
• Publish and check the local VA at comm->rank, which matches the bootstrap rank index
• On a full (non-RAIL) connection, GIN rank and world rank are the same, so the check is unchanged

#11569 rewrites ginAnvilRegisterLsaSignals and still sizes this buffer from ctx->nRanks. If that PR merges first, rebase this change onto it.

JIRA ID: AICOMRCCL-2339

Test Plan

• Single-node Anvil (NCCL_GIN_TYPE=7, 1 node / 2 ranks) still binds LSA signals; the world size equals the GIN team, so the gathered slot is the same as before
• No multi-node / RAIL run planned for this change

Test Result

• Not run on a SUT in this session

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

Made with Cursor

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	⚠️ Warning	Error: Source/code files changed without an accompanying unit test. Expected: add at least one test file named like test_<name>.py / test_<name>.cpp (or <name>_test.*). Current: code file(s) changed: projects/rccl/src/gin/gin_plugin_anvil_sdma.cc; no test file found
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[amd-argus]

ARGUS

Findings

1 finding(s) that could not be posted as inline comments

• projects/rccl/src/gin/gin_plugin_rocshmem_gda.cc:303 (Medium) The same team-sized allgather buffer is still live in the sibling rocshmem-gda backend, where the gathered data does reach the device The same overflow is still live in the sibling backend. ginRocshmemGdaCreateContext at gin_plugin_rocshmem_gda.cc:303-306 sizes both allgather buffers from ctx->nRanks and publishes at ctx->rank, then bootstrapAllGather at 308-309 writes comm->nRanks slots, and unlike the block fixed here the result is copied to the device at 311-312. ginRocshmemGdaRegMrSym at gin_plugin_rocshmem_gda.cc:187-193 has the same shape. That file is outside this PR, so this looks like a separate change rather than something to fold in here.

Bottom line: Worth settling before approval: no test can observe this fix, since the GIN plugin suite's allgather stub discards non-int payloads and passes identically with the fix reverted. The sizing change itself checks out against the bootstrap allgather contract.

_{Comment @argus review to re-run, or @argus re-review after pushing.}

[amd-argus]

Medium

No test can observe this fix: the stub bootstrapAllGather at gin_anvil_plugin_test_stubs.cc:103 returns early for any element size other than sizeof(int), and this call passes sizeof(uintptr_t), so rccl-UnitTestsGinAnvilPlugin passes bit-identically with the fix reverted. ConnCheck_SkippedOnRailConnection at GinAnvilPlugin_test.cpp:526 already builds comm->rank=1 against ctx->rank=0, so the divergence is one fixture away. Could the stub record the buffer and size, the way bootstrapIntraNodeAllGather already records ranks at gin_anvil_plugin_test_stubs.cc:144, so the published slot becomes assertable?

[amd-argus]

Nit

Suggested change

	}
	uintptr_t* gatheredLocalBases = (uintptr_t*)calloc((size_t)comm->nRanks, sizeof(uintptr_t));
	if (gatheredLocalBases) {
	gatheredLocalBases[comm->rank] = (uintptr_t)lsaSelf;
	if (ginAnvilBootstrapAllgather(comm->bootstrap, gatheredLocalBases, sizeof(uintptr_t)) != 0) {
	WARN("GIN anvil-sdma: signal local-base allgather failed");
	} else if (gatheredLocalBases[comm->rank] != (uintptr_t)lsaSelf) {
	WARN("GIN anvil-sdma: signal local-base allgather mismatch rank=%d got=%#lx expect=%#lx", comm->rank,
	(unsigned long)gatheredLocalBases[comm->rank], (unsigned long)lsaSelf);
	}

[dlamd1dai]

Superseded by #11569 (AICOMRCCL-2339).

#11569 was rebased onto current develop and now carries this fix directly. It deletes ginAnvilRegisterLsaSignals and replaces it with ginAnvilRegisterSignalSpan, which registers one LSA signal stripe per logical context, so the standalone patch here no longer applies to a function that exists. The world-sizing change was folded into the new span probe: the buffer is sized with comm->nRanks and the local VA is published at comm->rank, matching what bootstrapAllGather actually writes.

Closing in favour of #11569.