benchmark_methods
=================

.. py:module:: benchmark_methods

.. autoapi-nested-parse::

   Benchmark inference latency for all production API methods.

   For each method, this command can run up to three benchmark modes:
   1) uncached CPU baseline (existing path)
   2) uncached GPU precompute path (requires actual GPU execution)
   3) cached repeat, for methods with a persistent cache

   Each method is benchmarked for exactly one target: kcat if supported, otherwise
   Km. With --kcat_km_ratio, only kcat/Km-capable methods are benchmarked and each
   run uses the kcat/Km target. GPU mode is submitted only for method/target pairs
   with GPU offload support.
   Methods with ephemeral embedding use (currently OmniESI and EITLEM) skip the
   cached repeat because there is no persistent sequence cache to benchmark.

   Example summary:
     DLKcat [kcat] - 100.00s compute (not cached cpu), n/a compute (not cached gpu), 10.00s compute (cached)

   The command submits jobs to the public REST API (/api/v1/submit/) and polls
   /api/v1/status/<job_id>/, reporting only server-side compute time
   (`computeSeconds`) and not queue time.