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

[chienchunhung]

TEST ONLY

Retry of the exact pre-admission historical-source/current-runtime discriminator after the prior attempt was censored in prerequisite compilation. This draft is diagnostic only and is not intended for merge.

• Native source checkpoint: 85665f5fd3, exact native tree 8ada9734721bcd127722b367152eeb5bb22fa4ca, parent eaf5693b3aee9acea0d322dc8be997c4c0975280.
• 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 are absent from the checkout tree. Current main appears only through no-tree CI-ancestry merges; it contributes no product content to the tested tree.
• Runtime overlay, Jenkins harness, no-home-mount behavior, asynchronous-consensus factor, protocol mode 2 with cancellation disabled, CMake normalization/restoration, and the compile-only Marlin stub repair are unchanged from the prior attempt and completed slow discriminators.
• The only new delta is the exact relevant three-file compile/toolchain portion of the later DLFW dependency compatibility change, with stable patch ID 2128f376c81a96a070037f6d139abaf6c9b37092. It pins DeepEP compilation to the vendored NVSHMEM headers, removes a Clang-only flag after the sub-build switches to GCC, and normalizes CUDA virtual-memory aggregate initialization. It changes no admission, scheduling, transceiver, workload, or benchmark behavior.
• The exact workload YAML, test list, selector, performance baseline, performance test, and Slurm workload files remain byte-identical to the native checkpoint and the completed slow discriminators. No DSpark, workload, selector, timeout, or product compatibility adaptation is included.
• The target remains the exact GB300 3-node/12-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"

[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 #62236 [ run ] triggered by Bot. Commit: fd6fb71 Link to invocation

[tensorrt-cicd]

PR_Github #62236 [ run ] completed with state FAILURE. Commit: fd6fb71
/LLM/main/L0_MergeRequest_PR pipeline #50405 (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 outcome: censored before pytest

The single pre-admission discriminator run reached the exact requested GB300 stage at head fd6fb71e35ea2dd2113294596c065ce0afee312d. The prerequisite SBSA build succeeded, the exact three-node/twelve-GPU topology and selector were established, and artifact installation completed.

Before pytest collection or workload startup, the disaggregated-server process failed while importing tensorrt_llm/_utils.py:

ModuleNotFoundError: No module named 'tensorrt'

This is a runtime/package compatibility censor, not a performance result. There is no JUnit result, request count, protocol-mode evidence, shutdown evidence, or output-token metric. Consequently:

• no fast/slow classification is valid;
• the source regression bound does not move;
• this run must not be compared with completed throughput checkpoints;
• no manual rerun was issued.

This TEST ONLY draft is being closed unmerged. Its diagnostic branch is preserved.

[chienchunhung]

The import blocker was traced to the historical TensorRT 10.15/PyTorch 26.02 dependency input being installed inside the frozen TensorRT 10.16.1 runtime. The dependency-compatible replacement discriminator ports the exact upstream DLFW 26.04 requirements update used by the completed comparator, without changing product source or workload behavior. Its single targeted GB300 launch is active; this closed branch remains preserved.