[NVBUG-6448152][test] TEST ONLY earlier native-source discriminator

[chienchunhung]

TEST ONLY — do not merge.

This is one fixed-current-runtime source discriminator for the separate later C++ CTX PP forward/device-loop throughput regression. The asynchronous-consensus product change is already merged and is not under review here.

One-factor construction

• Native source: 58d8964d131a163dc91c871afbf85785857594f3
• Native source tree: 3455e1542c4f5a0ee47e7439f811db059f3c5c1f
• Native source parent: 602718366c40e7078cec1d520c531a4fd119ccc2
• This source is 53 first-parent commits earlier than the valid slow 274043a7 checkpoint from the preceding discriminator.
• Current-runtime overlay is limited to the NIXL installer/development pins and all five image tags using build 202607151440-16194; stable patch ID: 7e73140158499de670b40490317c6f77601d06fc.
• The official asynchronous-consensus factor is byte-identical by stable patch ID: 549e0085a06212cde0c47ab47072ad6acdcd1f58.
• One CMake list-order normalization was restored after applying the factor and has no net tree effect beyond adding the factor source.
• The native five-way stage mapping, workload YAML, selector, and timeout are unchanged.
• Current main is present only as a no-tree second parent for CI mergeability.

Interpretation

• At least 1402 output tokens/s places the regression in the following 53 first-parent commits through 274043a7.
• Near 800 output tokens/s places the regression at or before this checkpoint and moves the source bound earlier.
• Any partial recovery will be quantified.
• A build/import failure, timeout, or failed request is compatibility/censored evidence and is not a throughput result.

Requested evidence

Run only GB300-12_GPUs-3_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU4-GEN1-NODE2-GPU8-Post-Merge-1, requiring the exact selector, three nodes, twelve tasks, four GPUs per node, --no-container-mount-home, 512/512 successful requests, coordinator activation on all four CTX PP ranks, clean shutdown, and a valid official throughput metric.

The draft will be closed after terminal evidence is captured; the branch will be preserved.

[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 #61606 [ run ] triggered by Bot. Commit: c1c1671 Link to invocation

[tensorrt-cicd]

PR_Github #61606 [ run ] completed with state FAILURE. Commit: c1c1671
/LLM/main/L0_MergeRequest_PR pipeline #49816 (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 discriminator evidence for exact head c1c1671a4512781fa4aedfa4c0dc7e47e965db3f / native source 58d8964d131a163dc91c871afbf85785857594f3:

• Exact SBSA child / Slurm 2670797 ran the requested DeepSeek GB300 selector only, on three nodes / twelve tasks / four GPUs per node with --no-container-mount-home.
• Result artifact records 512/512 successful requests, 0 failed requests, 0.10 req/s, 801.91 output tok/s, and 13632.41 total tok/s.
• Logs show asynchronous context-transfer consensus enabled on CTX PP ranks 0, 1, 2, and 3. CTX, GEN, and disaggregated servers each reached complete application shutdown; no fatal, protocol, segmentation, or double-free errors were found.
• Jenkins is red only because the intentionally historical source misses the expected performance gate; the benchmark itself is valid.

801.91 output tok/s is +0.92% versus the later-tree 794.61 result and 51.77% of the known-fast historical 1548.84 result. There is no meaningful recovery: the slowdown is present at or before native source 58d8964d, excluding the following 53 first-parent commits through 274043a7 from being its introduction interval. The next source discriminator must move earlier.

[chienchunhung]

Closing this terminal TEST ONLY discriminator after capturing the valid benchmark evidence above. The branch is preserved for audit.