feat(compliance): MLPerf TEST04 caching audit — extensible AuditTest framework

[wu6u3tw]

Summary

Adds an extensible MLPerf compliance-audit framework with TEST04 (caching detection) as the first test, driven by an audit: block in the benchmark YAML. This PR carries the full redesign: the approved design plan, the implementation, tests, and runnable WAN2.2 examples.

TEST04 issues one fixed sample for every query in an audit phase; if repeating an identical request makes the SUT meaningfully faster, it is serving from cache. Pass iff the audit run is at most 10% faster than the reference (matching upstream compliance/TEST04/verify_performance.py).

Design (the two axes)

• Axis A — run modification: expressed as a generic typed SampleOrderSpec (WITHOUT_REPLACEMENT | SINGLE(index)) carried on a RunSpec. No test-specific knowledge leaks into the load generator.
• Axis B — verification: a pure post-run check, AuditTest.verify(runs) -> AuditVerdict, registered per test.

A generic orchestrator (commands/audit.py::run_audit) runs each RunSpec phase back-to-back via the existing setup_benchmark / run_benchmark_async path, then verifies and writes the verdict. Adding TEST01/06/07/09 later is a new registry entry, not cross-cutting edits.

Config shape

audit:
  test: test04
  samples: 64         # reference phase query count
  audit_samples: 64   # audit (fixed-sample) phase count
  sample_index: 3     # MLCommons performance_issue_same_index
  threshold: 0.10     # audit qps must stay < ref qps * (1 + threshold)

AuditConfig is a discriminated-union-ready sub-model on BenchmarkConfig (parallel to AccuracyConfig) — no DatasetType.AUDIT, no audit fields polluting Dataset, no test04 boolean in RuntimeSettings.

What's included

• compliance/__init__.py — AuditTest protocol + RunSpec/RunStats/RunArtifacts + registry
• compliance/verdict.py — AuditVerdict + atomic write_verdict (tmp → fsync → rename → fsync)
• compliance/tests/test04.py — Test04Audit + verify_test04
• commands/audit.py — generic run_audit orchestrator
• config/schema.py — AuditTestId + Test04Config/AuditConfig + BenchmarkConfig.audit
• load_generator — SampleOrderSpec + SingleSampleOrder + factory dispatch
• Unit tests + e2e integration test (offline + single-stream) against the echo server
• docs/compliance_audit_plan.md — the design plan
• WAN2.2 submission examples: offline_wan22_submission.yaml, single_stream_wan22_submission.yaml

Exit codes

benchmark from-config with an audit: block exits 0 (PASS) / 1 (FAIL); errors propagate via the standard handler using the repo-wide scheme (InputValidationError → 2, SetupError → 3, ExecutionError → 4). The on-disk audit_verdict.json is the durable record.

Testing

Unit + integration green; pre-commit run --all-files clean. The e2e test exercises the full audit: → run_audit → AuditVerdict flow for both max_throughput (offline) and concurrency=1 (single-stream).

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[gemini-code-assist]

Code Review

This pull request implements the MLPerf TEST04 compliance audit to detect result caching by repeatedly issuing a single fixed sample and comparing the throughput against a reference run. It introduces configuration options, validation guards, a SingleSampleOrder generator, and a compliance verification module with a CLI tool and tests. The review feedback focuses on improving the robustness of the compliance verifier, specifically by handling potential OSError exceptions during file writes, catching AttributeError when parsing non-dictionary JSON configurations, and gracefully handling malformed snapshot files during parsing.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[wu6u3tw]

Update summary

All review feedback has been addressed. Here is what changed since the original submission:

Architecture (main concern)

• audit: test04 now runs both phases in a single command — reference run then audit run back-to-back against the same endpoint, with automatic comparison and verdict output. No more 3-step manual workflow.

Config shape

• type: audit dataset replaces the old settings.runtime.test04_sample_index and audit_n_samples runtime variables. Reference and audit sample counts are now independent and co-located with the dataset config — consistent with how type: accuracy datasets carry their own accuracy_config.

audit: "test04"

datasets:
  - name: wan22_prompts
    path: wan22_prompts.jsonl
    type: "performance"
    samples: 50          # reference phase query count (50–144)

  - name: wan22_audit
    path: wan22_prompts.jsonl
    type: "audit"
    samples: 25          # audit phase query count (25–50)
    audit_sample_index: 0

Robustness

• Warning logged when audit: test04 is set but no type: audit dataset is present (previously silent fallback to index 0).
• Phase failures (SetupError/ExecutionError) are caught and logged cleanly — no unhandled traceback, verdict not lost.
• Report.from_snapshot wrapped in try/except in _run_stats_from_dir — malformed snapshots exit with code 2 instead of crashing.
• Pre-flight audit_sample_index bounds check before dataset load.

Testing

• New e2e integration test (test_audit_test04_two_phase_flow) exercises the full run_benchmark → two-phase flow against the echo server and asserts both phase subdirs are created and the flow completes gracefully.

Example config

• Renamed offline_wan22_test04.yaml → wan22_audit_test04.yaml per review suggestion.

[wu6u3tw]

All review feedback has been addressed. Here's a summary of what changed:

Architecture
audit: test04 now runs reference and audit phases in a single command back-to-back against the same endpoint — no more 3-step workflow, no endpoint-change risk. A single type: audit dataset entry drives both phases (carrying ref_samples, audit_samples, audit_sample_index).

Sample counts & index
ref_samples: 50, audit_samples: 25 — sized for WAN2.2 throughput. audit_sample_index: 3 — fixed per MLCommons audit.config (performance_issue_same_index=3 for WAN2.2).

SingleStream
Added wan22_single_stream_test04.yaml (concurrency=1, ref/audit samples=20 matching MLCommons min_query_count).

Durations
Perf configs: min=10min, max=4hr. Audit configs: min=10min, max=2hr. The 10-min minimum documents MLCommons compliance intent; counts take priority in the current session stop logic, with AND-semantics available as a future improvement.

Robustness fixes (Gemini)

• write_verdict moved inside try-except in CLI
• _audit_marker uses isinstance guards — no AttributeError possible
• Report.from_snapshot wrapped in try/except (KeyError, TypeError) in _run_stats_from_dir

Cleanup

• Test renamed to test_audit_test04.py
• README.md removed from diff (rebased onto main)
• Orphaned type: audit datasets in non-TEST04 configs now emit a warning; multiple audit datasets raise InputValidationError

[nvzhihanj]

The naming of the file is not consistent. some have offline/singlestream, some have audit_test04. Can you fix it?

[wu6u3tw]

renamed and replace the test name into output-caching test instead of using the term 'test04'

[nvzhihanj]

Review Council — first-principles design review

Reviewed by: Claude (Codex review timed out on this 2046-line diff at xhigh reasoning) · Depth: thorough

Focus: design issues warranting re-design for a modular, extensible audit-test framework (TEST04 is the first of several). 11 findings; see the tiered summary comment. The ref_samples dead-write (#1) was independently verified against the source.

[nvzhihanj]

Review Council — Multi-AI Code Review (first-principles design review)

Reviewed by: Claude · Depth: thorough
Codex review timed out on this 2046-line diff at xhigh reasoning (the load-gen + compliance surface is large); this pass is Claude-led. The one HIGH bug below was independently verified against the source.

Framing: TEST04 is the first MLPerf compliance/audit test and is meant to become a modular, extensible framework. The findings below are design-led — what would adding the next audit (TEST01/05) cost, and where does TEST04-specific knowledge leak into general-purpose code. 11 findings, all posted inline.

🔴 Re-design / Must-fix

#	File	Line	Cat	Why it needs a re-design
1	commands/benchmark/execute.py	1151	bug	ref_samples is a dead write. Dataset.samples is consumed nowhere; ref_config never sets n_samples_to_issue, so the reference phase runs duration-driven and ignores ref_samples while the audit phase honors audit_samples → the compared phases run mismatched counts. Set n_samples_to_issue=ref_samples.
2	compliance/__init__.py	18	design	No AuditTest abstraction. run_benchmark hardcodes if audit==TEST04; package exports only test04_*. Adding TEST01/05 = cross-cutting edits everywhere. Introduce an AuditTest protocol (plan_runs+verify) registered by AuditMode.
3	config/schema.py	82	design	DatasetType.AUDIT is a fake dataset type the loader ignores, carrying test params on the shared Dataset model, then converted to PERFORMANCE. Move params to a structured audit: block; drop the fake type.
4	config/runtime_settings.py	90	design	test04 boolean leaks into core load-gen. RuntimeSettings.test04/test04_sample_index + create_sample_order's if settings.test04. Use a generic sample-order strategy selector, not a per-test flag.

🟡 Should-fix

#	File	Line	Cat	Summary
5	commands/benchmark/execute.py	113	design	_OVERRIDE_TEST04_SAMPLE_INDEX stringly-typed magic kwarg through **runtime_overrides; pass a typed run_spec instead.
6	commands/benchmark/execute.py	1146	design	Two-phase model_copy surgery is fragile (root cause of #1; ref phase also skips _validate_audit_test04). Use a declarative RunSpec + validate before any phase runs.
7	tests/integration/commands/test_benchmark_command.py	209	testing	_run_benchmark_test04 has no unit test; the one integration test asserts verdict OR error with min_duration_ms=0 — the regime that hides bug #1.
8	config/schema.py	666	design	audit bare top-level enum; params scattered, threshold hardcoded. Use a structured compliance sub-config (like accuracy_config).
9	compliance/test04.py	206	design	QPS compared across phases with different counts/contents (upstream TEST04 uses the same query set); completion guard only protects the FAIL direction. Extends the existing fairness thread; compounded by #1.

🔵 Consider

#	File	Line	Cat	Summary
10	compliance/test04.py	175	design	verify_test04_dirs vs verify_test04_from_reports duplication; dir-swap guard in one path only. Collapse to one core + thin adapters.
11	commands/benchmark/execute.py	446	bug	audit_sample_index bound-checked vs requested counts, not the loaded dataset size, until phase 2 — an out-of-range index wastes a full reference run.

Through-line: #1, #5, #6, #7 are all symptoms of the same root cause — TEST04 is bolted onto run_benchmark via per-phase config surgery and untyped overrides instead of a first-class audit-test abstraction (#2). Fixing #2/#3/#4 (an AuditTest that emits typed RunSpecs + a generic ordering strategy) would dissolve most of the others structurally.

Dedup: none overlap existing inline comments except #9, which extends the maintainer's existing fairness thread with upstream-parity / guard-direction substance.

[viraatc]

lgtm, looking forward to impl!

[nvzhihanj]

We shouldn't enable TEST04 for LLM.

[wu6u3tw]

Modifie.d

[nvzhihanj]

If possible, unify the output to be json so it's easier to parse.

And do we need 2 files? Ideally one json should be enough (containing all information). On traditional MLPerf side we can always use scripts to make it pass

[wu6u3tw]

MLCommons run_verification.py takes verify_.txt that req a Performance check pass: True We can remove it if MLComm run_verification changed.

[nvzhihanj]

Not sure verdict is the right word to use (I know AI loves this word). Maybe auditResults will be more explicit and lower the entropy of understanding

[wu6u3tw]

Done

[nvzhihanj]

Again, recommend using an explicit name like output_caching_audit

[wu6u3tw]

Done renaming

[wu6u3tw]

Review Council — Multi-AI Code Review

Reviewed by: Claude (Codex unavailable — bwrap sandbox blocked, no sudo to relax userns) · Depth: thorough

4 findings posted inline. Verdict-correctness finding (threshold) is the one to prioritize.