[NVBUG-6448152][test] TEST ONLY; DO NOT REVIEW: benchmark Python generation-first flag-off control

[chienchunhung]

Purpose

TEST ONLY control for the generation-first Python-transceiver consensus performance experiment. The branch is exactly one signed YAML-only child of corrected harness base da6e140bda4c831bc813b3b192d7cb98a37025b2. That base is one harness-only child of the official asynchronous Python transceiver draft at product head 0f9acdd4dddbde7ccbf31973ce420bf4da6cb28a: it only passes the allowlisted server_config_extra.schedule_style value into the generated runtime config and adds focused tests. It does not change product behavior.

The previous run is not generation-first evidence: although its YAML requested generation_first, the old harness omitted that key from the generated server config and the runtime used context_first. This corrected head requires both the generated config and runtime log to confirm generation_first.

Experiment

Setting	Value
Schedule	generation_first
Async terminal agreement	off
Async readiness agreement	off
Context transceiver	Python
Generation transceiver	Python
Transport	NIXL

The control explicitly sets both agreement flags to zero on the CTX workers, preventing ambient opt-in. The GEN workers use the Python transceiver but receive no CTX-only agreement flags.

The run must use the exact GB300 disaggregated perf-sanity selector, complete every submitted request, and publish an official output-token-throughput metric. Any failed request censors the throughput result.

This diagnostic is not intended to merge.

[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 #60498 [ run ] triggered by Bot. Commit: 93d1f7e Link to invocation

[tensorrt-cicd]

PR_Github #60498 [ run ] completed with state FAILURE. Commit: 93d1f7e
/LLM/main/L0_MergeRequest_PR pipeline #48822 (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]

/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 #60513 [ run ] triggered by Bot. Commit: d2b0c45 Link to invocation

[tensorrt-cicd]

PR_Github #60513 [ run ] completed with state FAILURE. Commit: d2b0c45
/LLM/main/L0_MergeRequest_PR pipeline #48836 (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]

/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 #60820 [ run ] triggered by Bot. Commit: 11eb420 Link to invocation

[tensorrt-cicd]

PR_Github #60820 [ run ] completed with state FAILURE. Commit: 11eb420
/LLM/main/L0_MergeRequest_PR pipeline #49095 (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]

/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 #60874 [ run ] triggered by Bot. Commit: 156ca8e Link to invocation

[tensorrt-cicd]

PR_Github #60874 [ run ] completed with state FAILURE. Commit: 156ca8e
/LLM/main/L0_MergeRequest_PR pipeline #49146 (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