[NPU] Fix evalscope accuracy parsing and add glm5_1 aime26 request timeout (#36459)

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

@@ -227,17 +227,24 @@ def run_evalscope(
 
         if "accuracy" not in metrics:
             accuracy_patterns = [
+                # Add adaptation for evalscope 1.11+ table format
+                r"Accuracy\s*[↑↓]?\s*│\s*[^│]*│\s*\d+\s*│\s*([\d.]+)%?\s*│",
                 r"mean_acc\s*.*?│\s*\d+\s*│\s*([\d.]+)\s*│",
                 r"│\s+([\d.]+)\s+│\s+\S+\s+│\s*$",
                 r"accuracy\s*[:=]?\s*([\d.]+)",
+                # Keep compatibility with legacy evalscope 1.10 table format
                 r"Accuracy\s*[:=]?\s*([\d.]+)",
                 r"score\s*[:=]?\s*([\d.]+)",
             ]
 
             for pattern in accuracy_patterns:
-                matches = re.findall(pattern, full_output)
+                matches = list(re.finditer(pattern, full_output))
                 if matches:
-                    final_accuracy = float(matches[-1])
+                    final_accuracy = float(matches[-1].group(1))
+                    # evalscope 1.11+ reports accuracy as a percentage (e.g. 66.67%);
+                    # normalize it to a 0-1 fraction to compare against the baseline.
+                    if "%" in matches[-1].group(0):
+                        final_accuracy /= 100.0
                     metrics["accuracy"] = final_accuracy
                     logger.info(f"The Final Accuracy from output: {final_accuracy}")
                     break
@@ -307,7 +314,6 @@ class TestNpuAccuracyTestCaseBase(CustomTestCase):
     other_args = None
     server_timeout = DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
     envs = None
-    n_runs = 3
     accuracy = 0.1
     test_type = "accuracy"
 
@@ -504,67 +510,6 @@ class TestNpuAccuracyTestCaseBase(CustomTestCase):
                     )
             assert_metrics(self, best_metrics)
 
-    def run_accuracy_multiple(self, n_runs=None):
-        if n_runs is None:
-            n_runs = self.n_runs
-
-        parsed_url = urlparse(self.base_url)
-        host = parsed_url.hostname
-        port = parsed_url.port
-
-        if self.benchmark_tool != EVALSCOPE:
-            raise Exception(
-                "run_accuracy_multiple only supports evalscope benchmark tool"
-            )
-
-        model_name = os.path.basename(self.model)
-        all_metrics = []
-
-        for i in range(n_runs):
-            logger.info(f"=== Accuracy run {i + 1}/{n_runs} ===")
-            metrics = run_evalscope(
-                host=host,
-                port=port,
-                model=model_name,
-                datasets=self.datasets,
-                dataset_args=self.dataset_args,
-                eval_batch_size=self.eval_batch_size,
-                limit=self.limit,
-                generation_config=self.generation_config,
-                dataset_dir=self.dataset_dir,
-                stream=self.stream,
-                timeout=self.timeout,
-                eval_type=self.eval_type,
-            )
-            all_metrics.append(metrics)
-            if metrics and "accuracy" in metrics:
-                logger.info(f"Run {i + 1} accuracy: {metrics['accuracy']}")
-            else:
-                logger.warning(f"Run {i + 1} failed to get accuracy metric")
-
-        valid_metrics = [m for m in all_metrics if m and "accuracy" in m]
-        if not valid_metrics:
-            raise Exception("No valid accuracy metrics obtained from any run")
-
-        avg_accuracy = sum(float(m["accuracy"]) for m in valid_metrics) / len(
-            valid_metrics
-        )
-
-        logger.info("=" * 60)
-        logger.info("Multiple Run Accuracy Results:")
-        for i, m in enumerate(valid_metrics):
-            logger.info(f"  Run {i + 1}: {m['accuracy']}")
-        logger.info(f"  Average: {avg_accuracy}")
-        logger.info("=" * 60)
-
-        avg_metrics = {"accuracy": avg_accuracy}
-        dump_metric(
-            "accuracy_avg",
-            avg_accuracy,
-            labels={"test_case": self.__class__.__name__, "type": "accuracy"},
-        )
-        assert_metrics(self, avg_metrics)
-
 
 class TestNpuAccuracyMultiNodePdMixTestCaseBase(CustomTestCase):
     model_config = None

test/registered/npu/accuracy/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py

@@ -159,7 +159,7 @@ GLM_5_1_PD_SEP_MODEL_CONFIG = {
     "decode_args": GLM_5_1_PD_SEP_DECODE_ARGS,
     "prefill_envs": GLM_5_1_PD_SEP_PREFILL_ENVS,
     "decode_envs": GLM_5_1_PD_SEP_DECODE_ENVS,
-    "router_args": ["--policy", "round_robin"],
+    "router_args": ["--policy", "round_robin", "--request-timeout-secs", 7200],
     "router_envs": {},
 }
 

test/registered/npu/accuracy/qwen3_next_80b_a3b_instruct/test_npu_qwen3_next_80b_w8a8_2p_in6k_out1k5_bs16_aime25.py

@@ -112,7 +112,7 @@ class TestQwen3Next80BA3B_aime25(TestNpuAccuracyTestCaseBase):
     max_concurrency = 16
 
     def test_aime25(self):
-        self.run_accuracy_multiple(n_runs=3)
+        self.run_accuracy()
 
 
 if __name__ == "__main__":

test/registered/npu/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms.py

@@ -44,7 +44,7 @@ OTHER_ARGS = [
     "--nnodes",
     2,
     "--mem-fraction-static",
-    0.662,
+    0.68,
     "--max-running-requests",
     32,
     "--chunked-prefill-size",