# gpt-oss-120b INTERACTIVE loadgen user config for 2x8 MI355X (2-node P/D).
# Driven by `harness.py --interactive`, which reads these rows under the
# `gpt-oss-120b-interactive` benchmark and the `Interactive` scenario string.
# Interactive is NOT a loadgen scenario (its enum is SingleStream/MultiStream/
# Server/Offline); the run is still driven as Server. mlperf.conf supplies the
# tighter constraints: TTFT 3000 -> 2000 ms, TPOT 80 -> 15 ms.
#
# WHICH PREFIX A KEY NEEDS is not uniform, and getting it wrong is silent.
# loadgen's FromConfig (test_settings_internal.cc) looks up the "keys that apply
# to all scenarios" with the scenario string it was PASSED, but the per-scenario
# ones with a HARDCODED literal:
#     lookupkv(model, "Server", "target_qps", nullptr, &server_target_qps);
# So target_qps / target_latency / coalesce_queries / max_async_queries must be
# written `.Server.` even here -- an `*.Interactive.target_qps` row is never read
# and the run silently uses loadgen's default of 1 qps.

*.Interactive.performance_sample_count = 6396
*.Interactive.performance_sample_count_override = 6396

# BOTH have to be satisfied, and 600000 is not optional: the submission checker reads
# `effective_min_duration_ms` out of the detail log and rejects anything under
# TEST_DURATION_MS = 600000 (performance_check.py min_duration_check), which is how a run
# that stopped on query count alone came back as NoResults. The baseline cannot supply it
# either -- mlperf.conf only carries *.Server.min_duration, and loadgen looks min_duration
# up under the scenario string it was passed, which for this file is Interactive.
# At ~43.5 qps the 600s floor lands around 26.1k queries, so duration is the binding rule
# and min_query_count is the floor under it.
*.Interactive.min_query_count = 6396
*.Interactive.min_duration = 600000

# .Server. on purpose -- see the note above. Sized for the 15 ms TPOT budget,
# which is 5.3x tighter than the 80 ms Server run, so it is far below that
# scenario's qps rather than a scaled version of it.
*.Server.target_qps = 21.8