A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 02:00 UTC
[CI] Fix async scheduling + spec decoding test flake (#28902)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Nov 17, 2025 at 21:26 UTC
5bdd15527770ef39cc4c3cdca008fb4f9cf8a15f
1 file changed
+5
-3
tests/v1/e2e/test_async_scheduling.py
+5
-3
@@ -84,6 +84,7 @@ def test_with_spec_decoding(monkeypatch: pytest.MonkeyPatch):
84
"num_speculative_tokens": 2,
85
"model": "nm-testing/Llama3_2_1B_speculator.eagle3",
86
}
87
+ # Set small draft model len to force doesn't-fit-in-drafter case.
88
spec_config_short = spec_config | {"max_model_len": 50}
89
90
# test_preemption, executor, async_scheduling,
@@ -174,13 +175,14 @@ def run_tests(
175
):
176
if "spec_mml=None" in test_config:
177
assert (
177
- pytest.approx(test_acceptance_rate, rel=5e-2)
178
- == base_acceptance_rate
178
+ test_acceptance_rate > base_acceptance_rate
179
+ or test_acceptance_rate
180
+ == pytest.approx(base_acceptance_rate, rel=5e-2)
181
)
182
else:
183
# Currently the reported acceptance rate is expected to be
184
# lower when we sometimes skip drafting altogether.
183
- assert test_acceptance_rate > 0.05
185
+ assert test_acceptance_rate > 0.1
186
print(
187
f"PASSED: config=[{test_config}], params={params}"
188
f" accept_rate={test_acceptance_rate}"