[NVBUG-6448152][test] TEST ONLY: bypass disagg transfer admission

[chienchunhung]

TEST ONLY — do not merge

This replacement diagnostic isolates the reachable PyExecutor transfer-admission gate within the adjacent native admission/bounded-polling change. The superseded conflicting draft was closed without any CI trigger; its branch remains frozen.

The immediate parent checkpoint completed at 1,515.84 output tok/s. The adjacent native admission-change checkpoint completed at 799.73 output tok/s. Both used the same current-runtime harness, protected workload, and asynchronous-consensus factor.

Exact construction

• Diagnostic commit: 235d69128ccd62b6e7cc98c67645a2ff1231a673.
• Head: a8b4c17046daa51ba6ba740f831cddee9697685a, a signed ancestry-only merge whose first parent is the diagnostic commit and whose second parent is current main at b8a4af190dad7429886e3c90fb11dc6e47ad85f2.
• The head and first-parent trees are byte-identical: e1426af69a95bf80ae77c0b2e03927b2794f0742; the merge has no first-parent file diff.
• The treatment stable patch ID remains 1f6f1e2677c83bef4f26c0e2f2a8c6c0aebbcc46.

Treatment

• Preserve the exact native b6eacd1f baseline construction, runtime overlay, image selection, harness, selector, topology, NIXL configuration, consensus mode, and cancellation setting.
• Enable TRTLLM_DIAGNOSTIC_DISAGG_ADMISSION_BYPASS=1 only in the exact GB300 workload YAML.
• On PyExecutor scheduling ranks, preserve the scheduler's original ordered disaggregated-generation candidates and return wait_for_progress=false instead of applying the transfer-admission selection.
• Preserve those allocations on non-root PP ranks as the immediate parent did; do not change the physical transfer-buffer size or compute-activation limit.
• Emit one ID-free DIAGNOSTIC OBSERVED marker per executor only when a pure comparison shows the default gate would have deferred requests. The marker reports candidate/admitted/deferred counts and transfer-block totals, never request IDs.

Verification

• Changed-Python compile: pass.
• Complete changed-file pre-commit, DCO, YAML, test-list, Ruff, formatting, and diff checks: pass.
• Focused isolated admission-path smoke: pass.
• Independent treatment and ancestry-merge P0-P3/no-confound reviews: clean.

Required experiment evidence

The sole target is GB300-12_GPUs-3_Nodes-PyTorch-Disagg-PerfSanity-CTX1-NODE1-GPU4-GEN1-NODE2-GPU8-Post-Merge-1, selector disagg_upload-e2e-gb300_deepseek-r1-fp4_128k8k_con256_ctx1_pp4_gen1_dep8_eplb0_mtp1_ccb-NIXL.

A valid result requires all 512 requests to succeed, the bypass marker to prove the default gate would bind on every relevant GEN scheduling rank, consensus active on CTX PP ranks 0-3 with cancellation disabled, clean CTX/GEN/application shutdown, and a valid official output-token metric.

• At least 1,402 output tok/s implicates the PyExecutor admission gate as the dominant regression source for this workload.
• Near 800 output tok/s with exercised markers rules the gate down as dominant and promotes the GEN caller-cadence discriminator.
• Missing markers, build/import failure, timeout, cancellation, protocol mismatch, failed requests, or an invalid metric censor the experiment.

[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 #62816 [ run ] triggered by Bot. Commit: a8b4c17 Link to invocation

[tensorrt-cicd]

PR_Github #62816 [ run ] completed with state FAILURE. Commit: a8b4c17
/LLM/main/L0_MergeRequest_PR pipeline #50944 (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 evidence: admission-bypass diagnostic

The exact frozen workload at head a8b4c17046daa51ba6ba740f831cddee9697685a completed successfully:

• 512/512 requests succeeded; 0 failed
• b_is_valid=true
• Output throughput: 800.49 tok/s
• Total throughput: 13,608.34 tok/s
• CTX, GEN, and router each recorded exactly 512 HTTP 200 responses and clean application shutdown
• Protocol mode 2 was active with cancellation disabled on CTX ranks 0–3 in both benchmark rounds

This is only +0.095% versus the matched exact-admission result of 799.73 output tok/s. The reported CI failure is the expected performance-threshold result, not a request, protocol, or shutdown failure.

However, the required DIAGNOSTIC OBSERVED marker appeared zero times across all retained logs. That marker was the reachability proof that the default admission gate would have deferred a candidate. Its absence means the bypass treatment was not proven to exercise the decision path in this workload.

Therefore the throughput record is valid, but the causal discriminator is censored: this run does not prove that admission was bypassed in a consequential case and does not rule admission control in or out. The unchanged head should not be rerun. Closing this TEST ONLY draft unmerged and preserving its branch.