[NVBUG-6448152][test] TEST ONLY bisect pre-admission native 85665f5f

[chienchunhung]

TEST ONLY

Historical-source/current-runtime discriminator for the separate C++ token-throughput regression. This draft is diagnostic only and is not intended for merge.

• Native source checkpoint: 85665f5fd3, with exact native tree 8ada9734721bcd127722b367152eeb5bb22fa4ca and parent eaf5693b3aee9acea0d322dc8be997c4c0975280.
• This checkpoint is the exact first-parent immediately before the admission and bounded-polling change, and 100 first-parent commits before the known-slow native checkpoint 9ff00c869a36055fcf2ebf4de4950a0471a3039f.
• Admission-specific controller, requester timestamping, bounded polling, scheduler/configuration, and serialization changes are absent from the native tree. Current main is joined only as a no-tree second parent for CI ancestry; the exact checkout tree remains the audited first-parent tree.
• Runtime-only overlay is held byte-identical to the completed slow discriminators: NIXL 1.3.1, aiperf 0.8.0, and the same five CI image tags from build 202607151440-16194.
• CI-only compatibility holds the Jenkins harness and no-home-mount launch behavior byte-identical to those discriminators. A compile-only upstream Marlin empty-stub repair is also carried; the selected GB300 CUTEDSL workload does not execute Marlin.
• Because the native checkpoint predates cancellation APIs, the reviewed asynchronous-consensus factor is a semantic pre-cancellation port. It runs protocol mode 2 with cancellation fixed off, retains the native blocking/finite-wait behavior, and is valid for categorical fast-versus-near-800 localization rather than small-percent attribution.
• CMake context normalization and restoration have zero net ordering effect.
• The exact workload YAML, test list, selector, performance baseline, performance test, and Slurm workload files are byte-identical to both the native checkpoint and completed slow discriminators. No DSpark, workload, selector, timeout, scheduler, or product compatibility adaptation is included.
• The target remains the exact GB300 3-node/12-GPU disaggregated performance selector. A result at or above 1402 output tokens/s places the regression at the admission change or within the following 99 commits; a result near 800 proves the slowdown predates admission and moves the source bound earlier. Any build/import failure, timeout, protocol mismatch, cancellation, or failed request is non-interpretable.

The merged asynchronous-consensus change is not under re-review here; this draft only localizes the separate source throughput regression.

[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 #62089 [ run ] triggered by Bot. Commit: b6ba1cb Link to invocation

[tensorrt-cicd]

PR_Github #62089 [ run ] completed with state FAILURE. Commit: b6ba1cb
/LLM/main/L0_MergeRequest_PR pipeline #50274 (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

[chienchunhung]

Terminal result: censored before the benchmark

The exact historical source and sole requested GB300 stage were selected, but both prerequisite build families failed before any test child was created:

• the standard builds exposed a DeepEP NVSHMEM header/static-library compatibility mismatch;
• the LLVM builds treated an existing virtualMemory aggregate-initialization warning as an error.

There was no SBSA test child, Slurm allocation, pytest collection, serving request, or performance record. This run therefore provides no fast/slow evidence and does not move the source bound.

Both build failures are addressed by the later DLFW dependency compatibility change. Its relevant three-file compile/toolchain delta applies cleanly to this exact native checkpoint and introduces no admission-control or benchmark-path behavior. The next diagnostic will keep native source 85665f5fd331d0154a78172954846d843085e83f, the frozen runtime/harness, and the asynchronous-consensus factor unchanged, adding only that separately audited compatibility port.

Closing this TEST ONLY draft unmerged and preserving its branch.