biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Benchmark][2/2] Use spline interpolation to tune SLA variables (#32095)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-11 12:27:27 +08:00
committed by GitHub
parent 2a4dbe24ea
commit ef96fa3f1f
3 changed files with 232 additions and 264 deletions
Download Patch File Download Diff File Expand all files Collapse all files

272
tests/benchmarks/sweep/test_serve_sla.py
View File

@@ -5,7 +5,7 @@ from pathlib import Path
from unittest.mock import patch
from vllm.benchmarks.sweep.param_sweep import ParameterSweepItem
from vllm.benchmarks.sweep.serve_sla import _estimate_sla_bounds, _find_sla_value
from vllm.benchmarks.sweep.serve_sla import solve_sla
from vllm.benchmarks.sweep.server import ServerProcess
from vllm.benchmarks.sweep.sla_sweep import (
SLACriterionBase,
@@ -39,18 +39,70 @@ def _set_return_value(
return patch("vllm.benchmarks.sweep.serve_sla.run_sla", side_effect=mock_run_sla)
def _var2metric_identity(bench_comb):
return [{"request_throughput": float(bench_comb["request_rate"])}]
def _var2metric_linear():
def wrapped(bench_comb):
x = float(bench_comb["request_rate"])
y = x
return [{"request_throughput": y}]
return wrapped
def _run_estimate_sla_bounds(
def _var2metric_concave(elbow_point: float):
def wrapped(bench_comb):
x = float(bench_comb["request_rate"])
if x < elbow_point:
y = 0.5 * (x - elbow_point) + elbow_point
else:
y = 1.5 * (x - elbow_point) + elbow_point
return [{"request_throughput": y}]
return wrapped
def _var2metric_convex(elbow_point: float):
def wrapped(bench_comb):
x = float(bench_comb["request_rate"])
if x < elbow_point:
y = 1.5 * (x - elbow_point) + elbow_point
else:
y = 0.5 * (x - elbow_point) + elbow_point
return [{"request_throughput": y}]
return wrapped
def _var2metric_quadratic(y_intercept: float):
def wrapped(bench_comb):
x = float(bench_comb["request_rate"])
y = y_intercept + 0.1 * x**2
return [{"request_throughput": y}]
return wrapped
def _var2metric_sqrt(y_intercept: float):
def wrapped(bench_comb):
x = float(bench_comb["request_rate"])
y = y_intercept + 10 * x**0.5
return [{"request_throughput": y}]
return wrapped
def _run_solve_sla(
var2metric: Callable[[ParameterSweepItem], list[dict[str, float]]],
criterion: SLACriterionBase,
init_value: int,
max_value: int,
min_value: int = 1,
max_value: int = 100,
):
with _set_return_value(var2metric):
return _estimate_sla_bounds(
result = solve_sla(
server=None,
bench_cmd=[],
serve_comb=ParameterSweepItem(),
@@ -60,143 +112,129 @@ def _run_estimate_sla_bounds(
num_runs=1,
dry_run=False,
sla_variable="request_rate",
init_value=init_value,
max_value=max_value,
sla_min_value=min_value,
sla_max_value=max_value,
)
assert result is not None
return result
def test_estimate_sla_bounds_le():
sla_data, (max_passing, min_failing), history = _run_estimate_sla_bounds(
_var2metric_identity,
def test_solve_linear_sla_le():
sla_data, history = _run_solve_sla(
_var2metric_linear(),
SLALessThanOrEqualTo(target=32),
init_value=1,
max_value=100,
)
assert max_passing == 32
assert min_failing == 64
assert history.get_max_passing() == 32
assert {val: margin <= 0 for val, margin in history.items()} == {
100: False,
1: True,
2: True,
4: True,
8: True,
16: True,
32: True,
64: False,
33: False,
}
def test_estimate_sla_bounds_lt():
sla_data, (max_passing, min_failing), history = _run_estimate_sla_bounds(
_var2metric_identity,
def test_solve_linear_sla_lt():
sla_data, history = _run_solve_sla(
_var2metric_linear(),
SLALessThan(target=32),
init_value=1,
max_value=100,
)
assert max_passing == 16
assert min_failing == 32
assert history.get_max_passing() == 31
assert {val: margin <= 0 for val, margin in history.items()} == {
100: False,
1: True,
2: True,
4: True,
8: True,
16: True,
31: True,
32: False,
}
def test_estimate_sla_bounds_oob():
sla_data, (max_passing, min_failing), history = _run_estimate_sla_bounds(
_var2metric_identity,
def test_solve_linear_sla_oob():
sla_data, history = _run_solve_sla(
_var2metric_linear(),
SLALessThanOrEqualTo(target=32),
init_value=64,
max_value=128,
)
assert max_passing == 0
assert min_failing == 64
assert {val: margin <= 0 for val, margin in history.items()} == {
64: False,
}
def _run_test_find_sla_value_le(
var2metric: Callable[[ParameterSweepItem], list[dict[str, float]]],
criterion: SLACriterionBase,
min_value: int,
max_value: int,
):
with _set_return_value(var2metric):
return _find_sla_value(
server=None,
bench_cmd=[],
serve_comb=ParameterSweepItem(),
bench_comb=ParameterSweepItem(),
sla_comb=SLASweepItem({"request_throughput": criterion}),
base_path=Path(""),
num_runs=1,
dry_run=False,
sla_variable="request_rate",
min_value=min_value,
max_value=max_value,
)
def test_find_sla_value_le():
sla_data, sla_value, history = _run_test_find_sla_value_le(
_var2metric_identity,
SLALessThanOrEqualTo(target=50.0),
min_value=32,
max_value=64,
)
assert sla_value == 50
assert {val: margin <= 0 for val, margin in history.items()} == {
48: True,
56: False,
52: False,
50: True,
51: False,
}
def test_find_sla_value_lt():
sla_data, sla_value, history = _run_test_find_sla_value_le(
_var2metric_identity,
SLALessThan(target=50.0),
min_value=32,
max_value=64,
)
assert sla_value == 49
assert {val: margin <= 0 for val, margin in history.items()} == {
48: True,
56: False,
52: False,
50: False,
49: True,
}
def test_find_sla_value_oob():
sla_data, sla_value, history = _run_test_find_sla_value_le(
_var2metric_identity,
SLALessThanOrEqualTo(target=50.0),
min_value=64,
max_value=128,
)
assert sla_value == 64
assert history.get_max_passing() == 64
assert history.get_min_failing() == 64
assert {val: margin <= 0 for val, margin in history.items()} == {
96: False,
80: False,
72: False,
68: False,
66: False,
65: False,
100: False,
64: False,
}
def test_solve_concave_sla_le():
sla_data, history = _run_solve_sla(
_var2metric_concave(elbow_point=32),
SLALessThanOrEqualTo(target=24),
)
assert history.get_max_passing() == 16
assert {val: margin <= 0 for val, margin in history.items()} == {
100: False,
1: True,
7: True,
13: True,
15: True,
16: True,
17: False,
}
def test_solve_convex_sla_le():
sla_data, history = _run_solve_sla(
_var2metric_convex(elbow_point=32),
SLALessThanOrEqualTo(target=24),
)
assert history.get_max_passing() == 26
assert {val: margin <= 0 for val, margin in history.items()} == {
100: False,
1: True,
48: False,
30: False,
24: True,
26: True,
27: False,
}
def test_solve_quadratic_sla_le():
sla_data, history = _run_solve_sla(
_var2metric_quadratic(y_intercept=10),
SLALessThanOrEqualTo(target=50),
)
assert history.get_max_passing() == 20
assert {val: margin <= 0 for val, margin in history.items()} == {
100: False,
1: True,
4: True,
20: True,
21: False,
}
def test_solve_sqrt_sla_le():
sla_data, history = _run_solve_sla(
_var2metric_sqrt(y_intercept=10),
SLALessThanOrEqualTo(target=100),
)
assert history.get_max_passing() == 81
assert {val: margin <= 0 for val, margin in history.items()} == {
100: False,
1: True,
89: False,
81: True,
82: False,
}