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 14, 05:30 UTC
[CI] Tweaks to GPT-OSS Eval (Blackwell) for stability (#26030)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Oct 1, 2025 at 15:02 UTC
ee04c0cd04cfdf331168b5696a0411d1f8fc5fdf
2 files changed
+3
-4
.buildkite/test-pipeline.yaml
+1
-1
@@ -845,7 +845,7 @@ steps:
845
- vllm/v1/attention/backends/flashinfer.py
846
commands:
847
- uv pip install --system 'gpt-oss[eval]==0.0.5'
848
- - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58 --server-args '--tensor-parallel-size 2'
848
+ - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
849
850
- label: Blackwell Quantized MoE Test
851
timeout_in_minutes: 60
tests/evals/gpt_oss/test_gpqa_correctness.py
+2
-3
@@ -26,7 +26,8 @@ def run_gpqa_eval(model_name: str, base_url: str) -> float:
26
# Build the command to run the evaluation
27
cmd = [
28
sys.executable, "-m", "gpt_oss.evals", "--eval", "gpqa", "--model",
29
- model_name, "--reasoning-effort", "low", "--base-url", base_url
29
+ model_name, "--reasoning-effort", "low", "--base-url", base_url,
30
+ "--n-threads", "200"
31
]
32
33
try:
@@ -72,8 +73,6 @@ def test_gpqa_correctness(request):
73
74
# Add standard server arguments
75
server_args.extend([
75
- "--max-model-len",
76
- "32768",
76
"--trust-remote-code",
77
])
78