[NVBUG-6448152][test] TEST ONLY bisect native 9ff00c86

[chienchunhung]

TEST ONLY

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

• Native source checkpoint: 9ff00c869a, with exact native tree dd93f3cecf225b261951d7d315560dfa01a4081c and parent b155a516154198d77c65309680dfa1a761465a23.
• This is the balanced first-parent split: 100 commits after pre-admission anchor 85665f5fd331d0154a78172954846d843085e83f, 99 commits after the admission and bounded-polling change, and 100 commits before the known-slow native checkpoint.
• Runtime-only overlay: NIXL 1.3.1 and the same five CI image tags from build 202607151440-16194.
• CI-only compatibility holds the Jenkins harness byte-identical to the known-slow diagnostic. A compile-only upstream Marlin empty-stub repair is also carried; the selected CUTEDSL workload does not execute Marlin.
• The reviewed pre-cancellation asynchronous-consensus factor runs protocol mode 2 with cancellation fixed off.
• CMake context normalization and restoration have zero net ordering effect.
• No DSpark, workload YAML, test-list, selector, timeout, scheduler, or product 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 in the following 100 commits; a result near 800 moves the slow bound to this checkpoint. 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 later 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 #61958 [ run ] triggered by Bot. Commit: 2de8762 Link to invocation

[tensorrt-cicd]

PR_Github #61958 [ run ] completed with state FAILURE. Commit: 2de8762
/LLM/main/L0_MergeRequest_PR pipeline #50151 (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 TEST ONLY result

The exact native checkpoint 9ff00c869a36055fcf2ebf4de4950a0471a3039f completed the sole requested GB300 discriminator:

• 512/512 requests succeeded; 0 failed
• 6,005.89 s selector duration
• 795.48 output tok/s
• 0.10 requests/s
• 13,523.19 total tok/s
• protocol mode 2 with cancellation disabled on all four CTX PP ranks
• clean CTX, GEN, and disaggregated-service shutdown

The red result is solely the expected performance gate. This is +0.11% versus the later-tree 794.61 result and 51.36% of the historical repaired 1,548.84 result, so there is no recovery. The slowdown is present at or before native 9ff00c86; the following 100 first-parent commits through f4742f68 are excluded.

No rerun is needed. Closing this TEST ONLY draft unmerged while preserving the branch for reproducibility.