Fix GLM4-7B-Flash accuracy test configuration, tune Qwen3.6-27B/35B performance test parameters, and harden Ascend NPU multi-node E2E test utilities against pod name format errors. (#32371)

python/sglang/test/ascend/e2e/run_npu_e2e_test.py

@@ -310,7 +310,9 @@ def create_or_update_configmap(cm_name: str, data: dict, namespace: str):
 
 def prepare_cm_data(namespace, pod_string):
     """Prepare a configmap data: {pod_name: pod_ip} by the running pod's information."""
-    pods = core_api.list_namespaced_pod(namespace=namespace)
+    pods = core_api.list_namespaced_pod(
+        namespace=namespace, label_selector="app=sgl-ascend"
+    )
     data = {}
     for pod in pods.items:
         pod_name = pod.metadata.name

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

@@ -426,7 +426,12 @@ def check_role(allowed_roles: Union[str, Iterable[str]]):
 def launch_pd_mix_node(model_config):
     logger.info(f"Launch pd mix node start ......")
     host_name = get_host_name()
-    pod_index = int(host_name.rsplit("-", 1)[-1])
+    last_part = host_name.rsplit("-", 1)[-1]
+    if not last_part.isdigit():
+        raise RuntimeError(
+            f"Unexpected hostname format, expected numeric suffix: {host_name}"
+        )
+    pod_index = int(last_part)
 
     # Monitor ConfigMap to generate dist-init-addr and node-rank
     is_ready = False
@@ -497,7 +502,12 @@ def launch_pd_mix_node(model_config):
 def launch_pd_separation_node(model_config):
     logger.info(f"Launch pd separation node start ......")
     host_name = get_host_name()
-    pod_index = int(host_name.rsplit("-", 1)[-1])
+    last_part = host_name.rsplit("-", 1)[-1]
+    if not last_part.isdigit():
+        raise RuntimeError(
+            f"Unexpected hostname format, expected numeric suffix: {host_name}"
+        )
+    pod_index = int(last_part)
     role = "prefill" if "prefill" in host_name else "decode"
 
     bootstrap_init_port = BOOTSTRAP_INIT_PORT
@@ -717,7 +727,14 @@ def launch_router(model_config):
         node_ip_list.clear()
 
         for pod_name, pod_ip in configmap.data.items():
-            pod_index = int(pod_name.rsplit("-", 1)[-1])
+            # Skip unexpected entries that don't end with a numeric index
+            last_part = pod_name.rsplit("-", 1)[-1]
+            if not last_part.isdigit():
+                logger.info(
+                    "Skipping ConfigMap entry with non-numeric suffix: %s", pod_name
+                )
+                continue
+            pod_index = int(last_part)
 
             if "prefill" in pod_name:
                 if is_multi_node_prefill_instance:

test/registered/ascend/accuracy/glm4_7_flash/test_npu_glm4_7_flash_1p_aime25.py

@@ -16,7 +16,7 @@ register_npu_ci(
 ENVS = {
     "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
     "STREAMS_PER_DEVICE": "32",
-    "HCCL_BUFFSIZE": "1000",
+    "DEEPEP_HCCL_BUFFSIZE": "1000",
     "HCCL_OP_EXPANSION_MODE": "AIV",
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
@@ -57,12 +57,12 @@ OTHER_ARGS = [
 ]
 
 
-class TestNPUDeepSeek_V3_2_8P_AIME2025(TestNpuAccuracyTestCaseBase):
+class TestNPUGLM_4_7_FLASH_1P_AIME25(TestNpuAccuracyTestCaseBase):
 
     model = GLM_4_7_FLASH_MODEL_PATH
     envs = ENVS
     other_args = OTHER_ARGS
-    accuracy = 0.916
+    accuracy = 0.7667
     datasets = ["aime25"]
     few_shot_num = 0
     generation_config = {"max_tokens": 65536, "temperature": 1.0}

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

@@ -76,6 +76,10 @@ QWEN3_6_27B_64K_PREFIX_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_2p_in16k_out1k_50ms.py

@@ -21,9 +21,10 @@ QWEN3_6_27B_16K_1k_ENVS = {
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
     "SGLANG_SET_CPU_AFFINITY": "1",
+    "SGLANG_ENABLE_SPEC_V2": "1",
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
     "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
-    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
+    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "50",
     "GDN_ATTN_BACKEND_TRITON": "1",
     "ASCEND_USE_FIA": "1",
 }
@@ -44,22 +45,37 @@ QWEN3_6_27B_16K_1k_OTHER_ARGS = [
     "--disable-radix-cache",
     "--trust-remote-code",
     "--max-running-requests",
-    29,
+    37,
     "--max-mamba-cache-size",
-    58,
+    74,
     "--mem-fraction-static",
-    0.68,
+    0.70,
     "--cuda-graph-bs",
     1,
     2,
+    3,
+    4,
+    6,
     8,
+    10,
     12,
+    14,
     16,
+    18,
     20,
+    21,
+    23,
     24,
+    25,
     26,
+    27,
     28,
     29,
+    30,
+    31,
+    33,
+    35,
+    37,
     "--quantization",
     "modelslim",
     "--dtype",
@@ -69,11 +85,11 @@ QWEN3_6_27B_16K_1k_OTHER_ARGS = [
     "--speculative-algorithm",
     "NEXTN",
     "--speculative-num-steps",
-    3,
+    4,
     "--speculative-eagle-topk",
     1,
     "--speculative-num-draft-tokens",
-    4,
+    5,
     "--reasoning-parser",
     "qwen3",
     "--tool-call-parser",
@@ -90,8 +106,9 @@ class TestNPUQwen3_6_27B_2P_In16k_Out1k_50ms(TestNpuPerformanceTestCaseBase):
     other_args = QWEN3_6_27B_16K_1k_OTHER_ARGS
     envs = QWEN3_6_27B_16K_1k_ENVS
     dataset_name = "random"
-    max_concurrency = 29
-    num_prompts = 116
+    max_concurrency = 37
+    warmup_requests = 4
+    num_prompts = 37
     input_len = 16000
     output_len = 1000
     random_range_ratio = 1

test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in3k5_out1k5_50ms.py

@@ -18,7 +18,7 @@ register_npu_ci(
 QWEN3_6_35B_A3B_3K5_1K5_ENVS = {
     "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
     "STREAMS_PER_DEVICE": "32",
-    "HCCL_BUFFSIZE": "1",
+    "HCCL_BUFFSIZE": "100",
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",