[CI] Tweaks to GPT-OSS Eval (Blackwell) for stability (vllm-project#26030)

Signed-off-by: mgoin <[email protected]>

Author: mgoin

.buildkite/test-pipeline.yaml

@@ -845,7 +845,7 @@ steps:
   - vllm/v1/attention/backends/flashinfer.py
   commands:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - 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'
+    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
 
 - label: Blackwell Quantized MoE Test
   timeout_in_minutes: 60

tests/evals/gpt_oss/test_gpqa_correctness.py

@@ -26,7 +26,8 @@ def run_gpqa_eval(model_name: str, base_url: str) -> float:
     # Build the command to run the evaluation
     cmd = [
         sys.executable, "-m", "gpt_oss.evals", "--eval", "gpqa", "--model",
-        model_name, "--reasoning-effort", "low", "--base-url", base_url
+        model_name, "--reasoning-effort", "low", "--base-url", base_url,
+        "--n-threads", "200"
     ]
 
     try:
@@ -72,8 +73,6 @@ def test_gpqa_correctness(request):
 
     # Add standard server arguments
     server_args.extend([
-        "--max-model-len",
-        "32768",
         "--trust-remote-code",
     ])