[CI/Build] Warm the DP engines before measuring load balance in test_load (#60241)

Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
This commit is contained in:
stefankoncarevic
2026-10-06 11:23:37 -05:00
committed by GitHub
parent fcf53d240c
commit 4516896626
+20
View File
@@ -132,6 +132,26 @@ async def test_load(
)
after.callback(engine.shutdown)
# The first request an engine serves pays a few hundred milliseconds of
# one-off setup, and while it does the engine holds the requests already
# routed to it while reporting itself idle, which is enough to skew the
# split asserted on below. Pay it on every engine first, then measure.
await asyncio.gather(
*(
generate(
engine,
f"warmup-{rank}",
prompt,
output_kind,
max_tokens=1,
data_parallel_rank=rank,
)
for rank in range(DP_SIZE)
)
)
for sl in stats_loggers.values():
sl.finished_req_count = 0
NUM_REQUESTS = 100
NUM_EXPECTED_TOKENS = 10