[alex-breslow-amd]

Motivation

Restore performance on single node for MI350 when WarpSpeed enabled.

Technical Details

New as of 7/22/2026
-Address @Kapil-Shyam-Pawar's feedback regarding changes getting picked up by the testrunner. Thanks for the pointers! :-)

New as of 7/17/2026

• Synced with team. We disable WarpSpeed on non-SPX modes (e.g., DPX, QPX, CPX). Thanks @mustafabar and @wenkaidu for the suggestion.
• Move WarpSpeed enablement logic into separate functions that we can unit test
• Add unit tests for enablement logic
• Fixes regression in single node performance from PR#8077

Old as of 7/16/2026
Drop in performance due to max channels being capped at 32 for large message sizes on single node. This is one possible fix. Need to verify that it plays nice with CPX and still prevents the issue mentioned in PR#8077.

We need to decide if we should just revert to not using WarpSpeed and go back to the original logic. Will consult with team tomorrow.

Another option is checking which mode we're in e.g., DPX or CPX and having custom max channel counts for that.

Issue Tracking

JIRA ID: AICOMRCCL-1553

Test Plan

Ran with this PR on develop and without it. Verified that performance returns and channel count is again 56 for 1GiB.

Test Result

See above

Submission Checklist

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

[therock-pr-bot]

❌ PR Check — Action Required

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	❌ Fail	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/enqueue.cc, projects/rccl/src/graph/connect.cc, projects/rccl/src/include/rccl_common.h, projects/rccl/src/init.cc, projects/rccl/src/rccl_wrap.cc (+1 more); no test file found
🔎 pre-commit	⏳ Pending	⏳ Still running…
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

⚠️ 1 policy check(s) failed. Please address the issues above before this PR can be Reviewed.

🚫 Please fix the failed policies

• ❌ Unit Test

The Not ready to Review label was added to this PR. Once all policies pass, the label is removed automatically.

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

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ Unit Test

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.

[Copilot]

Pull request overview

Restores single-node performance for MI350 when WarpSpeed is enabled by moving the single-node MAXCHANNELS/2 clamp to occur after the gfx950 8-rank single-node halving, avoiding the effective “cap-at-32” behavior introduced by the earlier cap placement.

Changes:

• Removes the early single-node MAXCHANNELS/2 cap when computing nc under WarpSpeed.
• Adds a final single-node clamp (nc = min(nc, MAXCHANNELS/2)) after the gfx950 single-node halving logic to preserve higher channel budgets while still protecting comm->channels[MAXCHANNELS] from overflow.

---

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

[wenkaidu]

LGTM. I think SPX should be used as gate for other resource intensive algorithm as well, like direct algorithms.

[alex-breslow-amd]

LGTM. I think SPX should be used as gate for other resource intensive algorithm as well, like direct algorithms.

I'll add a follow up task for this. I think that makes sense at least on gfx942 and gfx950.

[alex-breslow-amd]

Merging in develop because CI failed due to ROCK issue. Fix and bump have landed: ROCm/TheRock@2e7c190 and #9056