[NVBUG-6448152][test] TEST ONLY pre-admission dependency-compatible retry

[chienchunhung]

TEST ONLY

Replacement for the closed pre-admission discriminator whose result was censored before pytest collection. This draft is diagnostic only and is not intended for merge.

• Native source checkpoint: 85665f5fd3, exact native tree 8ada9734721bcd127722b367152eeb5bb22fa4ca, parent eaf5693b3aee9acea0d322dc8be997c4c0975280.
• Exact diagnostic head: 17a1ffe8a3e50c34fc83560cf9db38fe78e2afe1, tree 6991f7968a1bc322d51881e7f3a146ef0fc207ad.
• This 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 remain absent. The asynchronous-consensus factor, protocol mode 2 with cancellation disabled, runtime overlay, Jenkins harness, no-home-mount behavior, CMake normalization/restoration, compile-only Marlin repair, and protected workload files are unchanged from the censored attempt.
• The prior attempt exposed a historical/current-runtime dependency mismatch: its development installation recursively consumed the native TensorRT 10.15/PyTorch 26.02 dependency bounds inside a frozen TensorRT 10.16.1 runtime, after which pytest collection could not import the tensorrt Python module.
• The only tested-tree delta from that attempt is the exact requirements.txt portion of the upstream DLFW 26.04 dependency-stack update, stable patch ID d00164dfea32bcf659400f20064852268a3e90d4. Its resulting blob, a243f9b26ab561018960ce9ffe2d546c7678f830, is byte-identical to the dependency input used by the completed native 9ff00c86 discriminator.
• This compatibility port changes no C++ product source, admission, scheduling, transceiver, selector, workload, timeout, or performance-test behavior. The latest upstream main is included only as a no-tree second parent for CI mergeability.
• The target remains the exact GB300 three-node/twelve-GPU disaggregated performance selector. A valid result at or above 1402 output tokens/s places the regression at the admission change or within the following 99 commits; a valid 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 #62277 [ run ] triggered by Bot. Commit: 17a1ffe Link to invocation

[tensorrt-cicd]

PR_Github #62277 [ run ] completed with state FAILURE. Commit: 17a1ffe
/LLM/main/L0_MergeRequest_PR pipeline #50446 (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]

Build-only follow-up on the prior dependency-compatible head:

• The first exact launch passed the public launch gates, fetched the full exact head, and started the sole requested GB300 stage.
• The primary x86 and SBSA builds succeeded, including SBSA artifact publication. A parallel LLVM build then stopped before any test child or pytest because FlashMLA invoked nvcc -ccbin=/usr/bin/g++ while still forwarding PyTorch’s Clang-only -Xcompiler=-fclang-abi-compat=17; GCC rejected the option. Therefore this run produced no request result and no throughput metric.
• The replacement compatibility commit applies only the exact nine-line FlashMLA hunk from the upstream dependency-stack fix. It removes that Clang-only flag only inside the existing ARM64 Clang-to-GCC fallback. Stable patch ID: fd7ac62fe956431597af3dab68b0c7a2d70fd290.
• The final FlashMLA CMake blob is byte-identical to current upstream. Product source, consensus factor, mode 2 with cancellation disabled, image, dependency set, workload, test list, topology, and harness remain unchanged. Two independent audits found no P0-P3 issue or experiment confound.

The exact replacement head is 378d5b2264a2b82147ff71635df1b60f4ee95529; its latest-main merge is ancestry-only and tree-identical to its first parent. GitHub reports it mergeable. The next run will request the same sole GB300 stage once.

[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 #62329 [ run ] triggered by Bot. Commit: 378d5b2 Link to invocation

[chienchunhung]

Terminal pre-admission discriminator

The exact native 85665f5f checkpoint, immediately before the admission/bounded-polling change, completed the exact GB300 disaggregated DeepSeek selector under the verified 3-node/12-GPU topology and current-runtime harness.

• 512/512 requests succeeded; 0 failed.
• 1515.84 output tok/s, 0.19 req/s, and 25769.26 total tok/s.
• The official performance record is valid and non-regressed: 2.22% below its 1550.25 output-tok/s baseline, within the 10% gate.
• Asynchronous context-transfer consensus was active on all four CTX PP ranks in both service starts. The reviewed mode-2 encoding was present as protocol version 1 with in-flight cancellation disabled.
• CTX, GEN, and disaggregated applications shut down cleanly. No fatal, protocol-mismatch, coordinator-shutdown, timeout, or failed-request evidence appeared.

This is a categorical fast result: +90.56% versus the valid slow 9ff00c86 checkpoint at 795.48 output tok/s, and 97.87% of the historical repaired 1548.84 result. The regression is therefore after native 85665f5f: it lies in the admission/bounded-polling change or one of the following 99 first-parent commits through 9ff00c86. This run does not yet attribute the regression specifically to admission control; it establishes the new fast lower bound.

This TEST ONLY diagnostic is complete. I am closing it unmerged while preserving the branch.

[tensorrt-cicd]

PR_Github #62329 [ run ] completed with state SUCCESS. Commit: 378d5b2
/LLM/main/L0_MergeRequest_PR pipeline #50501 (Partly Tested) completed with status: 'SUCCESS'

CI Report

Link to invocation