[Scheduler] Cap prefill-delayer queue target by admission capacity (#35191)

This commit is contained in:
paulzhang-tm
2026-08-18 21:40:13 +08:00
committed by GitHub
parent 94eef833fe
commit 0065fbfae1
3 changed files with 40 additions and 4 deletions
@@ -67,6 +67,7 @@ class NegotiateTestCase:
# to exercise the legacy slot-only code paths.
queue_min_ratio: Optional[float] = None
max_delay_ms: Optional[float] = None
prefill_max_requests: Optional[int] = None
# Expected accumulated wait surfaced on the final (release) outcome. When
# set, asserts the wait histograms would observe this value instead of 0.
expected_wait_forward_passes: Optional[int] = None
@@ -90,6 +91,7 @@ def _run_negotiate_test(rank, test_cases):
disable_overlap_schedule=False,
prefill_delayer_queue_min_ratio=case.queue_min_ratio,
prefill_delayer_max_delay_ms=case.max_delay_ms,
prefill_max_requests=case.prefill_max_requests,
),
max_delay_passes=case.max_delay_passes,
token_usage_low_watermark=case.token_usage_low_watermark,
@@ -291,6 +293,34 @@ _NEGOTIATE_TEST_CASES = [
expected_allow=False,
expected_reason="delay",
),
NegotiateTestCase(
name="queue_trigger_uses_admission_capacity",
max_delay_passes=100,
token_usage_low_watermark=0.8,
queue_min_ratio=0.02,
max_delay_ms=5000,
prefill_max_requests=128,
calls=[
NegotiateCall(
prefillable=[True, True, True, True],
token_usage=[0.9, 0.9, 0.9, 0.9],
running_batch=[500, 500, 500, 500],
max_prefill_bs=[1, 1, 1, 1],
waiting_queue_len=[1, 1, 1, 1],
max_running_requests=1024,
),
NegotiateCall(
prefillable=[True, True, True, True],
token_usage=[0.9, 0.9, 0.9, 0.9],
running_batch=[500, 500, 500, 500],
max_prefill_bs=[1, 1, 1, 1],
waiting_queue_len=[1, 1, 1, 1],
max_running_requests=1024,
),
],
expected_allow=False,
expected_reason="delay",
),
# Waiting queue at or above queue_min: queue trigger must not fire.
NegotiateTestCase(
name="queue_trigger_above_threshold",