[jberchtold-nvidia]

Description

Default is no logging. If the user sets NVTE_LOG_COLLECTIVE_GEMM_STREAM_IDS=1, then stream IDs will be logged

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Log stream IDs behind env var

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[jberchtold-nvidia]

/te-ci arm64 L1

[greptile-apps]

Confidence Score: 4/5

[Low impact] The change affects logging during collective GEMM context creation.

Add the missing cuBLASMp skip before merging so optional builds do not fail the distributed tests.

Findings

1.  Builds without cuBLASMp fail ▶

Summary

Adds opt-in logging of collective GEMM stream IDs and ranks.

• Collective GEMM contexts can print their CUDA stream IDs by rank.

_{Reviews (4) · Last reviewed commit: "Merge branch 'main' into jberchtold/cgem..." · Reviewed by Greptile}

[fheinecke]

/te-ci arm64 L1

[jberchtold-nvidia]

/te-ci L1 arm64

[greptile-apps]

Builds without cuBLASMp fail

The new test calls nvte_comm_gemm_ctx_create without checking nvte_built_with_cublasmp(). With NVTE_WITH_CUBLASMP disabled—the default—the call throws instead of skipping. The QA script launches this test whenever four GPUs are available, so builds without cuBLASMp now fail this suite.

Add the same early skip used by test_comm_gemm.cu, before creating the NCCL communicator.