TransferBench v1.70.02

[AtlantaPepsi]

Motivation

• v1.70.01 Skipped due to incorrect minor tagging on v1.70.00
• Update minor version tags.
• Modify rings to use all available executors by default
• Disable smoketest tests that don't scale well in multi-node, i.e. all-to-all, gather, broadcast

[Copilot]

🟡 Changes recommended

Fix the rings summary to accurately report variable per-executor counts or validate/aggregate them.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Updates TransferBench to v1.70.01 with broader rings executor usage and safer multi-node smoketest defaults.

Changes:

• Defaults rings to all available subexecutors.
• Disables non-scalable collective smoketests by default.
• Updates versioning, documentation, changelog, and CodeQL setup.

File summaries

File	Summary
src/client/Presets/SmokeTest.hpp	Adjusts multi-rank smoketest selection.
src/client/Presets/Rings.hpp	Uses all available subexecutors; summary reporting has an unresolved issue for heterogeneous counts.
docs/reference/presets.rst	Documents the new rings default.
CMakeLists.txt	Updates the fallback patch version.
CHANGELOG.md	Adds v1.70.01 release notes.
.github/workflows/codeql.yml	Simplifies Git installation.

Review details

• Files reviewed: 6/6 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

Rings may resolve inconsistent subexecutor counts across ranks, causing validation failure.

Review details

Suppressed comments (1)

src/client/Presets/Rings.hpp:146

• With NUM_SUB_EXEC=0, this now resolves numSubExecs independently from each transfer's rank/GPU. However, RunTransfers requires every rank's corresponding Transfer.numSubExecs to match rank 0, while this preset only validates equal GPU counts. A multi-rank run with different CU counts will therefore fail before execution with a Num SubExecutors consistency error; validate matching counts up front or choose one common count for all ranks.

      t.numSubExecs = (numSubExecs > 0) ? numSubExecs : TransferBench::GetNumSubExecutors(t.exeDevice);

• Files reviewed: 7/7 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

Rings auto mode can fail when ranks have different subexecutor counts.

Review details

Suppressed comments (1)

src/client/Presets/Rings.hpp:146

• When NUM_SUB_EXEC is 0, this resolves each hop from the local GPU's executor. However, RunTransfers requires each rank's corresponding Transfer, including numSubExecs, to match rank 0; this preset only validates that GPU counts match. A multi-rank run with the same GPU count but different CU counts will therefore fail in CheckMultiNodeTransferConsistency rather than run. Please either reject heterogeneous subexecutor counts in auto mode with a preset-level error or resolve one common count before building the transfers.

      t.numSubExecs = (numSubExecs > 0) ? numSubExecs : TransferBench::GetNumSubExecutors(t.exeDevice);

• Files reviewed: 7/7 changed files
• Comments generated: 0 new
• Review effort level: Lite