perf(workflow): qualify integrated encoder ensemble

[furionw]

Based on #13315 and #11848.

Why

The integrated-encoder workflow retains orchestration, classifier fan-out, and response joining around stock Dynamo vLLM while keeping the custom encoder in the vLLM process. The benchmark selects request rate and response-stage placement explicitly, so the final join can run inline with the orchestrator or through a discovered remote worker without changing the authored name-only workflow.

The qualification uses an uncapped 50 req/s open-loop source. This shows whether the workflow sustains offered load without a client concurrency cap obscuring saturation. New benchmark runs enable vLLM prefix caching and ZMQ KV-event publication by default and fail the audit unless Dynamo starts the event reader.

Effect

flowchart LR
    Workload[Audited Qwen workload]
    Frontend[Dynamo frontend]
    Orchestrator[WorkflowOrchestrator]
    Generator[Stock Dynamo vLLM<br/>integrated encoder]
    Classifier[Remote classifier]
    Inline[Inline response stage]
    Remote[Remote response worker]
    Audit[Offered-load and correctness audit]

    Workload --> Frontend --> Orchestrator
    Orchestrator --> Generator
    Orchestrator --> Classifier
    Generator --> Inline
    Classifier --> Inline
    Generator --> Remote
    Classifier --> Remote
    Inline --> Frontend
    Remote --> Frontend
    Frontend --> Audit

Loading

H100 qualification

These results are historical and remain tied to exact source commit 7143606b1710bc3e39ab6cf9b745c238acc7ae3f, an empty working diff, and the then-explicit --no-enable-prefix-caching setting. They are not relabeled as results for the rebased source or the new cache-enabled default.

• One NVIDIA H100 80GB HBM3 at 700 W and 1980 MHz; exactly one GPU visible to PyTorch.
• Container: nvcr.io/nvstaging/ai-dynamo/vllm-runtime:qiwa-localdev-vllm-x86-07-31-e092448f.
• Versions: Dynamo 1.4.0, AIPerf 0.8.0, PyTorch 2.11.0+cu130, Transformers 5.15.0, vLLM 0.26.0.
• Workload SHA-256: 743e859f895ee0e22df2476f74e5d3fa4d48db059273f5fe517634f31d9ef7cc.
• 1,000 measured requests plus 20 excluded warmups in each of three fresh-server repetitions per placement.
• Constant 50 req/s source, no measured AIPerf concurrency limit, non-streaming OSL 7, and average decoder ISL 874.5.

Response placement	Request-window runs (req/s)	Mean	Achieved/offered	Full-process runs (req/s)	From mean wall
Inline	49.871, 49.854, 49.857	49.861	99.72%	31.330, 31.335, 32.350	31.665
Remote	49.776, 49.754, 49.771	49.767	99.53%	31.533, 29.179, 32.009	30.856

All six measured cells completed 1,000 requests with zero errors or cancellations and exceeded the ≥49 req/s gate. Each produced exact OSL 7, average ISL 874.5, 907,800 patches across both audited grids, 14 CUDA graphs, and normalized classifier scores. Moving the response join to a discovered remote worker changed mean request-window throughput by about 0.19%.

Artifacts:

• /dynamo-tmp/logs/08-17/integrated-qwen-inline-response-r50-unlimited-edgeplan-7143606
• /dynamo-tmp/logs/08-17/integrated-qwen-remote-response-r50-unlimited-edgeplan-7143606

What Change

• Add audited mixed-image Qwen workflow benchmarking and provenance capture.
• Select request rate and response placement independently.
• Enable vLLM prefix caching and ZMQ KV-event publication for new runs and audit both from worker startup logs.
• Preserve benchmark behavior while rebasing authoring onto name-only contracts.

Test Plan

• The cumulative workflow suite passes with two environment-specific skips on the rebased branch.
• Relevant pre-commit hooks pass.
• Historical exact-SHA H100 qualification remains recorded above; it is not attributed to the rebased SHA.

[datadog-official]

🔄 Datadog auto-retried 10 jobs - 10 passed on retry

🎯 Code Coverage (details)
• Patch Coverage: 21.14%
• Overall Coverage: 47.72%

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 7143606 | Docs | Datadog PR Page | Give us feedback!}

[furionw]

Parking this qualification branch because its base, #13315, is no longer in the retained workflow stack. Historical benchmark results remain here; the active path is #13293 followed by #12924.