[NVBUG-6448152][test] TEST ONLY; DO NOT REVIEW lower compatible C++ midpoint

[chienchunhung]

Purpose

TEST ONLY; DO NOT REVIEW. Do not merge.

Run one fixed-current-runtime source discriminator for the separate later-tree C++ PP forward/device-loop throughput regression. The production async-consensus change remains in the official C++ review.

The prior compatible midpoint diagnostic completed 512/512 requests at 804.06 output tokens/s, showing that the dominant regression was already present at that checkpoint. This draft moves to the next audited lower native checkpoint while preserving every other tested factor.

Exact factor

• Native source parent: 699e2277a4b8b159e67d6daed3085e20c050b62d
• Diagnostic head: 02ef2f06104494d5aac7859d29cd3343c046a726
• Product tree: 8e4a24b7ed69636f4f511be918495b96e5010e86
• Async-consensus stable patch ID: 549e0085a06212cde0c47ab47072ad6acdcd1f58

All eight modified-file preimages match the official factor parent, all three added paths are absent at the native checkpoint, and all eleven postimages are byte-identical to the official async-consensus implementation.

The current image tags, exact GB300 stage YAML, exact workload YAML, and performance harness remain byte-identical to the current official baseline. The checkpoint also contains the native DSpark interface expected by that runtime.

Requested validation

Run only:

GB300-12_GPUs-3_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU4-GEN1-NODE2-GPU8-Post-Merge-1

Exact selector:

disagg_upload-e2e-gb300_deepseek-r1-fp4_128k8k_con256_ctx1_pp4_gen1_dep8_eplb0_mtp1_ccb-NIXL

A valid result requires every submitted request to succeed, asynchronous coordinator activation on all four context pipeline ranks, clean shutdown, and an official output-token-throughput record. Any timeout, failed request, import failure, or build failure censors throughput.

Local verification

• Stable patch identity and all preimage/postimage blobs verified.
• Changed-line clang-format, CMake formatting, codespell, whitespace, test-list validation, DCO, and all applicable pre-commit hooks passed.
• Independent read-only review found no P0/P1/P2 or compatibility issue.

No production or merge-readiness claim is made by this test-only pull request.

[chienchunhung]

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

[tensorrt-cicd]

PR_Github #61170 [ run ] triggered by Bot. Commit: 02ef2f0 Link to invocation

[tensorrt-cicd]

PR_Github #61170 [ run ] completed with state FAILURE. Commit: 02ef2f0
/LLM/main/L0_MergeRequest_PR pipeline #49419 (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