[NPU][CI] Fail fast and speed up long-running qwen3.6 accuracy cases (#39813)

.github/workflows/nightly-test-npu.yml

@@ -480,7 +480,7 @@ jobs:
 
   nightly-poc-multi-node-mix-tests-cann910:
     name: multi-node-mix-poc-cann910
-    if: ${{ !cancelled() }}
+    if: false
     needs: [set-image-config, nightly-poc-multi-node-tests-cann910]
     strategy:
       fail-fast: false

python/sglang/test/ascend/e2e/test_npu_accuracy_utils.py

@@ -314,6 +314,9 @@ class TestNpuAccuracyTestCaseBase(CustomTestCase):
     server_timeout = DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
     envs = None
     accuracy = 0.1
+    # Per-case evalscope accuracy retry count. None uses the dataset default
+    # (get_max_retries); set to 1 to fail fast on first measurement.
+    max_retries = None
     test_type = "accuracy"
 
     @classmethod
@@ -478,7 +481,11 @@ class TestNpuAccuracyTestCaseBase(CustomTestCase):
         port = parsed_url.port
         if self.benchmark_tool == EVALSCOPE:
             model_name = os.path.basename(self.model)
-            max_retries = get_max_retries(self.datasets)
+            max_retries = (
+                self.max_retries
+                if self.max_retries is not None
+                else get_max_retries(self.datasets)
+            )
             best_metrics = None
             for attempt in range(max_retries):
                 metrics = run_evalscope(

test/registered/npu/accuracy/qwen3_6_27b/test_npu_qwen3_6_27b_1p_gpqa.py

@@ -91,6 +91,7 @@ class TestNPUQwen3_6_27B_1P_In64k_Out1k_Prefix90_gpqa(TestNpuAccuracyTestCaseBas
     envs = QWEN3_6_27B_64K_PREFIX_ENVS
     other_args = QWEN3_6_27B_64K_PREFIX_OTHER_ARGS
     accuracy = 0.878
+    max_retries = 1
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     eval_batch_size = 64

test/registered/npu/accuracy/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_1p_in3k5_out1k5_50ms_gpqa.py

@@ -86,9 +86,10 @@ class TestNPUQwen3_6_27B_1P_In3k5_Out1k5_gpqa(TestNpuAccuracyTestCaseBase):
     envs = QWEN3_6_27B_3K5_1K5_ENVS
     other_args = QWEN3_6_27B_3K5_1K5_OTHER_ARGS
     accuracy = 0.878
+    max_retries = 1
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
-    eval_batch_size = 8
+    eval_batch_size = 16
     generation_config = {
         "max_tokens": 81920,
         "temperature": 1.0,