[NVBUG-6448152][test] TEST ONLY; DO NOT REVIEW current source with historical runtime

[chienchunhung]

Purpose

TEST ONLY. DO NOT REVIEW OR MERGE.

This reverse crossover isolates the runtime/toolchain contribution after the historical-source/current-runtime diagnostic could not reach its requested performance stage because the prerequisite build was incompatible. It retains the exact reviewed current C++ source at commit a2fca780 and replaces only its five CI image assignments with the historical image set used by the known-fast validation tree.

Factor isolation

• Current source parent: commit a2fca780.
• The five current image assignments from build 202607151440-16194 are replaced with the exact five assignments from historical source commit 5e2498b, build 202606051544-14972.
• Product code, benchmark YAML, selector, workload, coordinator behavior, traffic, admission, and timeout settings are unchanged from the current reviewed source.

This is a one-factor reverse crossover: current source plus historical runtime/toolchain. It is not intended for review or merge.

Why the reverse crossover is required

The forward crossover used historical source with current runtime. Its exact-head helper entered pipeline 49303 but failed in the x86 prerequisite during DeepEP device linking against current NVSHMEM symbols. Fail-fast then aborted the SBSA build before any requested test child, Slurm allocation, selector execution, or performance metric. That result is build-incompatible and therefore inconclusive for throughput.

Validation

• The factor commit has exactly one parent, the reviewed current-source commit.
• The committed delta from that parent is exactly one file and five line substitutions.
• The resulting image-tag file is byte-identical to the historical comparison file.
• Shell parsing and all five required key/value checks passed.
• The image-renaming utility parsed all five entries in dry-run mode under Python 3.13.
• Repository pre-commit and DCO checks passed.

Targeted validation is limited to the exact GB300 PP4 disaggregated performance-sanity workload:

disagg_upload-e2e-gb300_deepseek-r1-fp4_128k8k_con256_ctx1_pp4_gen1_dep8_eplb0_mtp1_ccb-NIXL

Throughput from this diagnostic will be interpreted only if every submitted request succeeds, the generation-first configuration is confirmed, coordinator mode is active on all four context pipeline ranks, shutdown is clean, and the official output-token metric is valid.

[chienchunhung]

/bot run --stage-list "GB300-12_GPUs-3_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU4-GEN1-NODE2-GPU8-Post-Merge-1"

[tensorrt-cicd]

PR_Github #61061 [ run ] triggered by Bot. Commit: f44f23f Link to invocation

[tensorrt-cicd]

PR_Github #61061 [ run ] completed with state FAILURE. Commit: f44f23f
/LLM/main/L0_MergeRequest_PR pipeline #49317 (Partly Tested) completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation