Decision record · ADR
ADR-0049: Restore the AgentX profile to the release-pinned AIPerf replay surface
Status: superseded | Date: 2026-09-26 Superseded by: ADR-0050
References: RFC-0004, ADR-0044, ADR-0035, ADR-0031
Context
Section titled “Context”ADR-0044 made the release-qualified AgentX profile declare a 300-second per-trace idle-gap cap and a cache-pressure warmup of 10 requests per lane, on the premise that the release-pinned AIPerf offered both as native options. That premise was false for every official AIPerf distribution. The per-lane warmup budget (warmupRequestsPerLane) and a trace idle-gap cap that the inferencex-agentx-mvp scenario admits exist only in the SemiAnalysisAI/aiperf fork that InferenceX installs from its own checkout. The official AIPerf 0.12.0 and 0.13.0 releases have no per-lane budget, and their inferencex-agentx-mvp scenario sets forbid_trace_idle_gap_cap. A downstream workspace on InferLab 0.15.0 therefore could not run any AgentX case: the pinned AIPerf rejected the generated configuration before the first transport request.
The deterministic suite did not catch the drift because it asserts the lowered configuration, not its acceptance by the pinned tool, and the recorded real qualification ran outside the release-owned toolchain. The catalog still binds InferenceX revision 900b3d8, whose script used the 600-second duration-mode warmup and no per-trace cap; the ADR-0044 values came from a later upstream script that runs against the fork.
Upstream is converging: ai-dynamo/aiperf#1413 ports the fork’s per-lane warmup budget and a runtime per-trajectory-tree idle cap, but it is unmerged and in no release. The operator considered maintaining an InferLab-owned AIPerf fork or shipping the SemiAnalysis fork in a second Bench environment, and chose to stay on official AIPerf releases.
Decision
Section titled “Decision”The release-qualified AgentX profile declares only replay parameters that the release-pinned official AIPerf distribution accepts under its inferencex-agentx-mvp scenario. The embedded Bench toolchain moves to official AIPerf 0.13.0. The profile returns to duration-mode cache-pressure warmup of 600 seconds and declares no per-trace idle-gap cap; the 10-second global idle cap and every other profile value stay unchanged. The removal of the extra zero-warmup-error acceptance gate from ADR-0044 is retained: AIPerf’s native submission validity, failure threshold, and context-overflow decisions remain the sole acceptance authority, and cache-pressure warmup failures remain recorded warmup evidence.
The result is parameter-identical with the cataloged InferenceX revision, but not with later InferenceX scripts that use the fork’s per-lane budget and per-trace cap. When an official AIPerf release carries both capabilities, a new decision may realign the profile with the then-current InferenceX script and re-qualify it with a real run. Until then, InferLab does not reimplement fork scheduling in its release-owned shim, because ADR-0035 assigns AgentX replay semantics to AIPerf.
Consequences
Section titled “Consequences”Positive: AgentX cases run again on the shipped toolchain, and the profile once more matches the configuration shape that the original real qualification exercised. It also matches the InferenceX revision the catalog names. The toolchain remains a single official AIPerf distribution, so the exact-version shim, runner handshake, and record identity keep one code base to qualify against.
Negative: InferLab AgentX numbers are not parameter-identical to InferenceX submissions produced by scripts that use the fork’s per-lane warmup budget and 300-second per-trace idle cap. Comparisons across that boundary need to call out the difference; records carry the effective replay policy, so the difference is auditable. AgentX records from releases 0.13.2 through 0.15.0 declare the count-mode policy and per-trace cap in their replay-policy fields; those runs failed before measurement, and the restored policy fields differ from theirs.
Neutral: RFC-0004 already admits either warmup form and makes the per-trace cap optional, so only its informative rationale, which claimed parity with the per-trace cap and count-bounded warmup, needs correction. Realigning with InferenceX later is a profile-data and lowering change, not a new operator surface.
Alternatives Considered
Section titled “Alternatives Considered”Restore the profile to parameters the pinned official AIPerf accepts (600-second duration warmup, no per-trace cap) and upgrade to official AIPerf 0.13.0, keeping the gate removal (accepted)
Section titled “Restore the profile to parameters the pinned official AIPerf accepts (600-second duration warmup, no per-trace cap) and upgrade to official AIPerf 0.13.0, keeping the gate removal (accepted)”- Pros: Unblocks AgentX on the shipped toolchain with only official, PyPI-distributed AIPerf, Returns to the configuration shape of the original real qualification and the cataloged InferenceX revision, Keeps one AIPerf code base behind the exact-version shim and runner identity
- Cons: Not parameter-identical with later InferenceX scripts that use the fork’s per-lane budget and per-trace cap
Ship the SemiAnalysisAI/aiperf fork as a separate AgentX Bench environment pinned by git revision (rejected)
Section titled “Ship the SemiAnalysisAI/aiperf fork as a separate AgentX Bench environment pinned by git revision (rejected)”- Pros: Immediate full parameter parity with current InferenceX submissions
- Cons: The release toolchain depends on a git source whose self-reported version collides with official releases, so identity must move to a revision, The entrypoint and phase-barrier shim must be qualified against two diverging AIPerf code bases
Maintain an InferLab-owned AIPerf fork of an official release plus the upstream AgentX port (rejected)
Section titled “Maintain an InferLab-owned AIPerf fork of an official release plus the upstream AgentX port (rejected)”- Pros: One AIPerf for every Bench path with an InferLab-controlled release cadence
- Cons: The upstream AgentX port conflicts with the official release in core timing, credit, and record modules and must be rebased on every upstream release, InferLab would own the correctness of a third AIPerf lineage and still track the SemiAnalysis fork’s semantics for comparability
Reimplement per-lane warmup budgets and per-trace idle caps inside the release-owned AIPerf shim (rejected)
Section titled “Reimplement per-lane warmup budgets and per-trace idle caps inside the release-owned AIPerf shim (rejected)”- Pros: Keeps official AIPerf while reproducing the fork’s parameters
- Cons: Moves native replay scheduling into InferLab, contradicting ADR-0035; the per-tree idle cap is a runtime watchdog rewrite, not a parameter