[NVBUG-6448152][test] TEST ONLY: admission-control C++ throughput discriminator

[chienchunhung]

TEST ONLY — do not merge

This diagnostic pins the native C++ source at the admission and bounded-polling change itself. Its parent was measured fast by the pre-admission discriminator at 1515.84 output tok/s, while the later native 9ff00c86 discriminator measured 795.48 output tok/s.

Native checkpoint

• Native commit: b6eacd1f725d0fbeb888e09ff2585e5ab23b0856
• Native parent: 85665f5fd331d0154a78172954846d843085e83f
• Native tree: dbf8f5a0c754a318f48b138d5bba47720f0ae9f8
• Distance: one first-parent commit after the fast bound and 99 before the slow 9ff00c86 bound

Held constant

• Current NIXL/runtime image, Jenkins harness, Slurm launchers, TensorRT requirements, and compile-only toolchain compatibility are byte-identical to the successful pre-admission diagnostic.
• The protected workload YAML, GB300 test-list selector, and performance driver are unchanged.
• Intended stage: GB300-12_GPUs-3_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU4-GEN1-NODE2-GPU8-Post-Merge-1
• Intended selector: disagg_upload-e2e-gb300_deepseek-r1-fp4_128k8k_con256_ctx1_pp4_gen1_dep8_eplb0_mtp1_ccb-NIXL
• Topology remains three nodes, twelve tasks, four GPUs per node, with --no-container-mount-home.
• The reviewed admission-aware asynchronous-consensus factor uses protocol version 1 (encoded mode 2) with cancellation disabled.

Relative to the successful fast diagnostic, the source difference is exactly the 25-file native admission commit; no additional file differs. Consensus product-source blobs are byte-identical to the validated later discriminator. The two differing test blobs retain only their native admission-era context, while the factor's zero-context added/removed payload is identical.

Validation before CI

• Changed-file pre-commit suite passed under Python 3.13.
• Every first-parent diagnostic commit has a DCO sign-off.
• Latest-main ancestry is a no-tree merge: the final tree equals its first-parent tree.
• Independent review found no P0–P3 issue or experiment confound.

Interpretation

• A valid result near 800 output tok/s places the regression at this admission-control change.
• A valid result of at least 1402 output tok/s excludes this change and confines the regression to the following 99 first-parent commits.
• Build/import failure, timeout, protocol mismatch, cancellation, or any failed request censors the result.

CI has intentionally not been triggered yet.

[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 #62588 [ run ] triggered by Bot. Commit: dd4a7ac Link to invocation

[chienchunhung]

Terminal admission-change discriminator

The exact native b6eacd1f admission/bounded-polling checkpoint, the single first-parent commit after the known-fast 85665f5f checkpoint, completed the exact GB300 disaggregated DeepSeek selector under the verified three-node, twelve-GPU topology and current-runtime harness.

• 512/512 requests succeeded; 0 failed.
• 799.73 output tok/s, 0.10 req/s, 13,595.41 total tok/s, and 6.49 user tok/s.
• The official record is valid and regressed: 48.41% below its 1,550.25 output-tok/s baseline.
• Asynchronous consensus was active on CTX PP ranks 0-3 in both service starts: protocol v1, cancellation disabled (the reviewed local mode-2 encoding).
• CTX, GEN, and disaggregated applications shut down cleanly. The uploaded workload logs contain no fatal, protocol-mismatch, coordinator-failure, request-timeout, or failed-request evidence.
• The only pytest failure is the expected historical performance-regression gate.

This checkpoint is categorically slow: 47.24% below the immediate parent's valid 1,515.84 output tok/s result, while only 0.53% above the later slow 9ff00c86 result of 795.48 output tok/s.

Because the native checkpoints are adjacent in first-parent history and the protected workload, runtime, harness, and consensus factor are identical, the regression is introduced by the admission/bounded-polling change, native commit b6eacd1f. This bisect identifies the commit/change; it does not by itself distinguish the admission logic from the bounded-polling edits within that same change.

The TEST ONLY diagnostic is complete. Closing it unmerged while preserving the branch.

[tensorrt-cicd]

PR_Github #62588 [ run ] completed with state FAILURE. Commit: dd4a7ac
/LLM/main/L0_MergeRequest_PR pipeline #50730 (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