profile_export_aiperf.json Schema

After every aiperf profile run, AIPerf writes a summary JSON file (default name profile_export_aiperf.json) under the artifact directory. Each top-level metric entry holds a stats block; this page documents which fields appear in that block, when they appear, and how the schema is versioned.

The on-disk shape is produced by JsonMetricResult in src/aiperf/common/models/export_models.py. Fields that are unset are omitted from the JSON output (exclude_none=True), so the field set per metric varies by metric type — this page is the source of truth for which fields to expect where.

Per-metric stats fields

The metric type (record / aggregate / derived) is documented per-metric in Metrics Reference. At a glance: latencies and per-request lengths are record; counts and timestamps are aggregate; throughputs and run-level totals are derived.

Example

A run with 20 requests against a streaming chat endpoint produces entries shaped like this:

1
{
2
  "schema_version": "1.3",
3
  "request_latency": {
4
    "unit": "ms",
5
    "avg": 2620.71,
6
    "min": 2145.06,
7
    "max": 3411.10,
8
    "p50": 2568.73,
9
    "p99": 3371.24,
10
    "std": 297.93,
11
    "count": 20,
12
    "sum": 52414.29
13
  },
14
  "request_throughput": {
15
    "unit": "requests/sec",
16
    "avg": 1.45
17
  },
18
  "request_count": {
19
    "unit": "requests",
20
    "avg": 20.0
21
  }
22
}

Note that request_throughput (derived) and request_count (aggregate) carry only unit + avg — no count, no sum, no percentiles. request_latency (record) carries the full set.

Top-level fields

In addition to the per-metric stats blocks, profile_export_aiperf.json includes top-level provenance:

run_info

Schema 1.2 introduced run_info to surface the seed and sweep coordinates needed to reproduce a run from the JSON file alone, without consulting the internal run_config.json handoff file. Schema 1.3 extends it with identifiers and the redacted CLI command.

Example for variation 2 of a concurrency grid sweep with --random-seed 42:

1
"run_info": {
2
  "benchmark_id": "abc123def456",
3
  "sweep_id": "8c4f9a2e-1234-4567-89ab-0123456789ab",
4
  "random_seed": 44,
5
  "trial": 0,
6
  "run_label": "run_0001",
7
  "variation_label": "concurrency_40",
8
  "variation_index": 2,
9
  "variation_values": {"phases.profiling.concurrency": 40},
10
  "cli_command": "aiperf profile --model meta-llama/Llama-3.1-8B-Instruct --url http://localhost:8000 --request-count 500"
11
}

Schema versions

The current schema version is exported as the top-level schema_version field on the JSON document. Bump on additive changes; coordinate a major bump for any field rename or removal.

Other JSON exports use independent schema versions

aiperf writes additional JSON files when --num-profile-runs >= 2:

• profile_export_aiperf_aggregate.json — confidence aggregation across runs. Per-metric blocks have a different shape (mean, std, cv, se, ci_low, ci_high, t_critical, unit) and own their own schema_version (AggregateConfidenceJsonExporter.SCHEMA_VERSION, currently "1.0").
• profile_export_aiperf_collated.json — pools per-request values from all runs into a single population, then emits combined percentiles (mean, std, p50, p90, p95, p99, count) under a combined key plus a per_run list of run-level summaries. Uses its own schema_version ("1.0.0").

The schema_version documented on this page applies only to profile_export_aiperf.json. The other files evolve on their own cadence.

For downstream parsers

• Treat absent fields as “not applicable to this metric type,” not “data missing.” A derived-metric block with no count is normal; a record-metric block with no count indicates a bug.
• Do not assume the field set is closed. Future minor schema bumps may add fields. Use schema_version to detect compat; ignore unknown fields.
• unit is authoritative for the value’s interpretation. Do not infer units from the metric tag.