Skip to content

Inference-perf BR v0.2 conversion omits input_token_rate #1744

Description

@danehans

Component

Analysis / Benchmark Report v0.2 conversion

Description

The inference-perf converter in native_to_br0_2.py writes aggregate output_token_rate, total_token_rate, and request_rate, but does not write input_token_rate.

The Benchmark Report v0.2 schema supports all three token-rate fields. Consumers such as Prism therefore cannot display input token throughput from the generated report, even though inference-perf summary_lifecycle_metrics.json contains enough information.

Current behavior

results.request_performance.aggregate.throughput.input_token_rate is absent from inference-perf BR v0.2 output.

Expected behavior

The converter should populate input_token_rate. It can use a native input-token throughput value when available, or derive it as total_token_rate - output_token_rate.

The converter should preserve None when the required source values are unavailable and include regression coverage for the mapping.

Reproduction

  1. Run an inference-perf benchmark with analysis enabled.
  2. Compare summary_lifecycle_metrics.json with the generated benchmark_report_v0.2,_*.yaml.
  3. Observe that output and total token rates are present in the report, while input token rate is missing.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions