feat: sync npu nightly test improvements from Ascend testcases (#29403)

test/registered/ascend/accuracy/deepseek_v3_2/test_npu_deepseek_v3_2_8p_aime25.py

@@ -25,6 +25,10 @@ OTHER_ARGS = [
     "--quantization",
     "modelslim",
     "--disable-radix-cache",
+    "--reasoning-parser",
+    "deepseek-v3",
+    "--tool-call-parser",
+    "deepseekv32",
 ]
 
 

test/registered/ascend/accuracy/glm4_6v_flash/test_npu_glm4_6v_flash_1p_mmmu.py

@@ -50,6 +50,10 @@ OTHER_ARGS = [
     32,
     "--watchdog-timeout",
     9000,
+    "--reasoning-parser",
+    "glm45",
+    "--tool-call-parser",
+    "glm45",
 ]
 
 

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

@@ -50,6 +50,10 @@ OTHER_ARGS = [
     32,
     "--watchdog-timeout",
     9000,
+    "--reasoning-parser",
+    "glm45",
+    "--tool-call-parser",
+    "glm47",
 ]
 
 

test/registered/ascend/accuracy/qwen3_32b/test_npu_qwen3_32b_bf16_8p_gpqa.py

@@ -65,6 +65,10 @@ QWEN3_32B_OTHER_ARGS = [
     64,
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 
@@ -72,7 +76,7 @@ class TestQwen32B_GPQA(TestNpuAccuracyTestCaseBase):
     model = QWEN3_32B_MODEL_PATH
     envs = QWEN3_32B_ENVS
     other_args = QWEN3_32B_OTHER_ARGS
-    accuracy = 0.516
+    accuracy = 0.4949
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     eval_batch_size = 64

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

@@ -80,7 +80,7 @@ QWEN3_6_27B_64K_PREFIX_OTHER_ARGS = [
 ]
 
 
-class TestNPUQwen3_6_27B_2P_In64k_Out1k_Prefix90_gpqa(TestNpuAccuracyTestCaseBase):
+class TestNPUQwen3_6_27B_1P_In64k_Out1k_Prefix90_gpqa(TestNpuAccuracyTestCaseBase):
     model = QWEN3_6_27B_MODEL_PATH
     envs = QWEN3_6_27B_64K_PREFIX_ENVS
     other_args = QWEN3_6_27B_64K_PREFIX_OTHER_ARGS
@@ -88,7 +88,13 @@ class TestNPUQwen3_6_27B_2P_In64k_Out1k_Prefix90_gpqa(TestNpuAccuracyTestCaseBas
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     eval_batch_size = 64
-    generation_config = {"max_tokens": 81920, "temperature": 1.0}
+    generation_config = {
+        "max_tokens": 81920,
+        "temperature": 1.0,
+        "extra_body": {
+            "chat_template_kwargs": {"enable_thinking": True},
+        },
+    }
 
     def test_gpqa(self):
         self.run_accuracy()

test/registered/ascend/accuracy/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_aime26.py

@@ -71,6 +71,10 @@ QWEN3_6_35B_A3B_3K5_1K5_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -84,7 +88,11 @@ class TestNPUQwen3_6_35BA3B_1P_In3k5_Out1k5_aime26(TestNpuAccuracyTestCaseBase):
     eval_batch_size = 4
     generation_config = {
         "max_tokens": 131072,
-        "temperature": 0.2,
+        "temperature": 0.6,
+        "top_p": 0.95,
+        "top_k": 20,
+        "min_p": 0.0,
+        "presence_penalty": 0.0,
         "repetition_penalty": 1.08,
     }
 

test/registered/ascend/accuracy/qwen3_vl_30b_a3b_thinking/test_npu_qwen3_vl_30b_a3b_thinking_1p_mmmu.py

@@ -16,71 +16,42 @@ register_npu_ci(
 )
 
 ENVS = {
-    "ASCEND_LAUNCH_BLOCKING": "0",
     "SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
     "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
-    "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
-    "SGLANG_ENABLE_SPEC_V2": "1",
-    "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
-    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200",
-    "HCCL_BUFFSIZE": "400",
+    "HCCL_BUFFSIZE": "2000",
 }
 
 OTHER_ARGS = [
     "--trust-remote-code",
-    "--nnodes",
-    "1",
-    "--node-rank",
-    "0",
     "--attention-backend",
     "ascend",
     "--device",
     "npu",
     "--max-running-requests",
-    162,
+    128,
     "--disable-radix-cache",
-    # "--speculative-draft-model-quantization",
-    # "unquant",
     "--chunked-prefill-size",
     -1,
     "--max-prefill-tokens",
-    35000,
-    # "--speculative-algorithm",
-    # "EAGLE3",
-    # "--speculative-draft-model-path",
-    # QWEN3_A3B_EAGLE_MODEL_PATH,
-    # "--speculative-num-steps",
-    # 3,
-    # "--speculative-eagle-topk",
-    # 1,
-    # "--speculative-num-draft-tokens",
-    # 4,
+    12800,
+    "--prefill-max-requests",
+    10,
     "--tp-size",
     2,
     "--mem-fraction-static",
-    0.87,
-    "--cuda-graph-bs",
-    1,
-    5,
-    15,
-    40,
-    70,
-    100,
-    120,
-    130,
-    140,
-    146,
-    150,
-    154,
-    156,
-    158,
-    160,
-    162,
+    0.8,
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3-thinking",
+    "--tool-call-parser",
+    "qwen",
+    "--enable-multimodal",
+    "--mm-attention-backend",
+    "ascend_attn",
 ]
 
 
@@ -91,7 +62,7 @@ class TestQwen3(TestNpuAccuracyTestCaseBase):
     accuracy = 0.76
     datasets = ["mmmu"]
     few_shot_num = 0
-    generation_config = {"max_tokens": 65536, "temperature": 1.0}
+    generation_config = {"max_tokens": 65536}
     eval_batch_size = 64
 
     def test_mmmu(self):

test/registered/ascend/accuracy/qwen3_vl_8b_thinking/test_npu_qwen3_vl_8b_thinking_1p_mmmu.py

@@ -21,22 +21,17 @@ ENVS = {
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
-    "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
-    "SGLANG_ENABLE_SPEC_V2": "1",
+    "HCCL_BUFFSIZE": "2000",
 }
 
 OTHER_ARGS = [
     "--trust-remote-code",
-    "--nnodes",
-    "1",
-    "--node-rank",
-    "0",
     "--attention-backend",
     "ascend",
     "--device",
     "npu",
     "--max-running-requests",
-    16,
+    64,
     "--max-prefill-tokens",
     16384,
     "--disable-radix-cache",
@@ -45,26 +40,16 @@ OTHER_ARGS = [
     "--tp-size",
     2,
     "--mem-fraction-static",
-    0.894,
-    "--cuda-graph-bs",
-    1,
-    5,
-    15,
-    16,
+    0.8,
     "--dtype",
     "bfloat16",
-    # "--speculative-draft-model-quantization",
-    # "unquant",
-    # "--speculative-algorithm",
-    # "EAGLE3",
-    # "--speculative-draft-model-path",
-    # QWEN3_8B_EAGLE_MODEL_PATH,
-    # "--speculative-num-steps",
-    # 4,
-    # "--speculative-eagle-topk",
-    # 1,
-    # "--speculative-num-draft-tokens",
-    # 5,
+    "--reasoning-parser",
+    "qwen3-thinking",
+    "--tool-call-parser",
+    "qwen",
+    "--enable-multimodal",
+    "--mm-attention-backend",
+    "ascend_attn",
 ]
 
 
@@ -75,8 +60,8 @@ class TestQwen3(TestNpuAccuracyTestCaseBase):
     accuracy = 0.741
     datasets = ["mmmu"]
     few_shot_num = 0
-    generation_config = {"max_tokens": 65536, "temperature": 1.0}
-    eval_batch_size = 16
+    generation_config = {"max_tokens": 65536}
+    eval_batch_size = 64
 
     def test_mmmu(self):
         self.run_accuracy()

test/registered/ascend/performance/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py

@@ -97,6 +97,10 @@ GLM_5_1_PD_SEP_PREFILL_ARGS = [
     1,
     "--pp-size",
     8,
+    "--reasoning-parser",
+    "glm45",
+    "--tool-call-parser",
+    "glm47",
 ]
 
 GLM_5_1_PD_SEP_DECODE_ARGS = [
@@ -147,6 +151,10 @@ GLM_5_1_PD_SEP_DECODE_ARGS = [
     "round_robin",
     "--speculative-draft-model-quantization",
     "unquant",
+    "--reasoning-parser",
+    "glm45",
+    "--tool-call-parser",
+    "glm47",
 ]
 
 GLM_5_1_PD_SEP_MODEL_CONFIG = {
@@ -187,7 +195,7 @@ class TestNPUGLM5_1_W4A8_PD_SEP_In3k5_Out1k5(TestNpuPerfMultiNodePdSepTestCaseBa
     output_len = 1024
     random_range_ratio = 1
     tpot = 50
-    output_token_throughput = 16
+    output_token_throughput = 160
 
     def test_npu_glm5_1_w4a8_pd_sep_in3k5_out1k5(self):
         """Run NPU performance test for GLM-5.1-w4a8 PD separation"""

test/registered/ascend/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms_aime25.py

@@ -48,7 +48,7 @@ OTHER_ARGS = [
     "--nnodes",
     2,
     "--mem-fraction-static",
-    0.55,
+    0.662,
     "--max-running-requests",
     32,
     "--chunked-prefill-size",
@@ -82,6 +82,10 @@ OTHER_ARGS = [
     4,
     "--speculative-draft-model-quantization",
     "unquant",
+    "--reasoning-parser",
+    "kimi_k2",
+    "--tool-call-parser",
+    "kimi_k2",
 ]
 
 MODEL_CONFIG = {

test/registered/ascend/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms_gpqa.py

@@ -86,6 +86,10 @@ MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS = [
     "--dtype",
     "bfloat16",
     "--trust-remote-code",
+    "--reasoning-parser",
+    "minimax-append-think",
+    "--tool-call-parser",
+    "minimax-m2",
 ]
 
 
@@ -106,7 +110,7 @@ class TestNPUMiniMaxM2_5W8A8_4P_In64k_Out1k_Prefix90_50ms(
     random_range_ratio = 1
     repeat_rate = 0.9
     tpot = 50
-    output_token_throughput = 390.5839
+    output_token_throughput = 390.5859
     request_rate = float("inf")
 
     def test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms(self):
@@ -122,7 +126,7 @@ class TestNPUMiniMaxM2_5_W8A8_4P_In3k5_Out1k5_GPQA(TestNpuAccuracyTestCaseBase):
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     generation_config = {"max_tokens": 65536, "temperature": 1.0}
-    max_concurrency = 64
+    eval_batch_size = 64
 
     def test_accuracy(self):
         self.run_accuracy()

test/registered/ascend/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_8p_in3k5_out1k5_50ms_gpqa.py

@@ -87,6 +87,10 @@ MINIMAX_M2_5_HIGH_THROUGHPUT_OTHER_ARGS = [
     "unquant",
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "minimax-append-think",
+    "--tool-call-parser",
+    "minimax-m2",
 ]
 
 
@@ -122,7 +126,7 @@ class TestNPUMiniMaxM2_5_W8A8_8P_In3k5_Out1k5_GPQA(TestNpuAccuracyTestCaseBase):
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     generation_config = {"max_tokens": 65536, "temperature": 1.0}
-    max_concurrency = 64
+    eval_batch_size = 64
 
     def test_accuracy(self):
         self.run_accuracy()

test/registered/ascend/performance/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in3k5_out1k5_50ms_gpqa.py

@@ -82,6 +82,10 @@ QWEN3_8B_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 

test/registered/ascend/performance/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in6k_out1k5_bs16_gpqa.py

@@ -69,6 +69,10 @@ QWEN3_8B_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     5,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 
@@ -84,7 +88,7 @@ class TestQwen8B(TestNpuPerformanceTestCaseBase):
     output_len = 1500
     random_range_ratio = 1
     tpot = 11.79
-    output_token_throughput = 930
+    output_token_throughput = 1040.96
 
     def test_qwen3_8b(self):
         self.run_throughput()

test/registered/ascend/performance/qwen3_30b_a3b/test_npu_qwen3_30b_w8a8_1p_in3k5_out1k5_50ms_aime25.py

@@ -87,6 +87,10 @@ QWEN3_30B_A3B_OTHER_ARGS = [
     162,
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 

test/registered/ascend/performance/qwen3_32b/test_npu_qwen3_32b_bf16_8p_in18k_out4k_6ms.py

@@ -65,12 +65,16 @@ QWEN3_32B_OTHER_ARGS = [
     1,
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 
 class TestQwen32B(TestNpuPerformanceTestCaseBase):
     benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    aisbench_dataset_type = AISBENCHMARK_DATASET_DEFAULT
+    dataset_type = AISBENCHMARK_DATASET_DEFAULT
     model = QWEN3_32B_MODEL_PATH
     other_args = QWEN3_32B_OTHER_ARGS
     envs = QWEN3_32B_ENVS

test/registered/ascend/performance/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_a2.py

@@ -84,6 +84,10 @@ QWEN3_32B_OTHER_ARGS = [
     101,
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 
@@ -93,16 +97,12 @@ class TestQwen32B_GPQA(TestNpuAccuracyTestCaseBase):
     model = QWEN3_32B_W8A8_MODEL_PATH
     other_args = QWEN3_32B_OTHER_ARGS
     envs = QWEN3_32B_ENVS
-    accuracy = 0.516
+    accuracy = 0.4949
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     eval_batch_size = 64
     generation_config = {"max_tokens": 40000, "temperature": 1.0}
 
-    @classmethod
-    def tearDownClass(cls):
-        pass
-
     def test_qwen3_32b_qpqa(self):
         """Run NPU accuracy test for Qwen3-32B-W8A8 on qpqa"""
         self.run_accuracy()
@@ -110,7 +110,7 @@ class TestQwen32B_GPQA(TestNpuAccuracyTestCaseBase):
 
 class TestQwen32B(TestNpuPerformanceTestCaseBase):
     base_url = DEFAULT_URL_FOR_TEST
-    aisbench_dataset_type = AISBENCHMARK_DATASET_DEFAULT
+    dataset_type = AISBENCHMARK_DATASET_DEFAULT
     model = QWEN3_32B_W8A8_MODEL_PATH
     other_args = QWEN3_32B_OTHER_ARGS
     envs = QWEN3_32B_ENVS
@@ -123,10 +123,6 @@ class TestQwen32B(TestNpuPerformanceTestCaseBase):
     tpot = 50
     output_token_throughput = 1600
 
-    @classmethod
-    def setUpClass(cls):
-        pass
-
     def test_qwen3_32b(self):
         self.run_throughput()
 

test/registered/ascend/performance/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_gpqa.py

@@ -84,6 +84,10 @@ QWEN3_32B_OTHER_ARGS = [
     101,
     "--dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
 ]
 
 

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_1p_in1024x1024_30_out1024_50ms.py

@@ -80,6 +80,10 @@ QWEN3_6_27B_1024_OTHER_ARGS = [
     "--speculative-num-draft-tokens",
     4,
     "--mm-enable-dp-encoder",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -87,11 +91,12 @@ class TestNPUQwen3_6_27B_1P_In1024x1024_30_Out1024_50ms(TestNpuPerformanceTestCa
     """Test NPU performance for Qwen3.6-27B 1p in1024x1024 30 out1024 50ms"""
 
     benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    aisbench_dataset_type = AISBENCHMARK_DATASET_MM_CUSTOM_GEN
+    dataset_type = AISBENCHMARK_DATASET_MM_CUSTOM_GEN
     model = QWEN3_6_27B_MODEL_PATH
     other_args = QWEN3_6_27B_1024_OTHER_ARGS
     envs = QWEN3_6_27B_1024_ENVS
-    dataset_name = "random"
+    backend = "sglang-oai-chat"
+    dataset_name = "image"
     max_concurrency = 48
     num_prompts = 48
     input_len = 30

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_1p_in1080p_30_out256_50ms.py

@@ -77,6 +77,10 @@ QWEN3_6_27B_1080P_OTHER_ARGS = [
     "--speculative-num-draft-tokens",
     4,
     "--mm-enable-dp-encoder",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -84,11 +88,12 @@ class TestNPUQwen3_6_27B_1P_In1080p_30_Out256_50ms(TestNpuPerformanceTestCaseBas
     """Test NPU performance for Qwen3.6-27B 1p in1080p 30 out256 50ms"""
 
     benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    aisbench_dataset_type = AISBENCHMARK_DATASET_MM_CUSTOM_GEN
+    dataset_type = AISBENCHMARK_DATASET_MM_CUSTOM_GEN
     model = QWEN3_6_27B_MODEL_PATH
     other_args = QWEN3_6_27B_1080P_OTHER_ARGS
     envs = QWEN3_6_27B_1080P_ENVS
-    dataset_name = "random"
+    backend = "sglang-oai-chat"
+    dataset_name = "image"
     max_concurrency = 30
     num_prompts = 120
     input_len = 30

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_2p_in64k_out1k_prefix90_50ms.py

@@ -46,24 +46,23 @@ QWEN3_6_27B_64K_PREFIX_OTHER_ARGS = [
     "--max-running-requests",
     20,
     "--max-mamba-cache-size",
-    120,
+    160,
     "--mem-fraction-static",
-    0.8,
+    0.82,
     "--cuda-graph-bs",
     1,
     2,
-    4,
-    8,
+    5,
     10,
-    12,
-    16,
-    18,
+    15,
+    17,
+    19,
     20,
     "--enable-prefill-delayer",
     "--prefill-delayer-queue-min-ratio",
-    0.5,
+    0.7,
     "--prefill-delayer-max-delay-ms",
-    30000,
+    20000,
     "--dtype",
     "bfloat16",
     "--mamba-ssm-dtype",
@@ -76,11 +75,15 @@ QWEN3_6_27B_64K_PREFIX_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
-class TestNPUQwen3_6_27B_2P_In64k_Out1k_Prefix90_50ms(TestNpuPerformanceTestCaseBase):
-    """Test NPU performance for Qwen3.6-27B-w8a8 2p in64k out1k prefix90 50ms"""
+class TestNPUQwen3_6_27B_1P_In64k_Out1k_Prefix90_50ms(TestNpuPerformanceTestCaseBase):
+    """Test NPU performance for Qwen3.6-27B 1p in64k out1k prefix90 50ms"""
 
     benchmark_tool = BENCHMARK_TOOL_DEFAULT
     model = QWEN3_6_27B_MODEL_PATH
@@ -94,11 +97,12 @@ class TestNPUQwen3_6_27B_2P_In64k_Out1k_Prefix90_50ms(TestNpuPerformanceTestCase
     random_range_ratio = 1
     repeat_rate = 0.9
     request_rate = float("inf")
+    warmup_requests = 1
     tpot = 50
     output_token_throughput = 225
 
     def test_npu_qwen3_6_27b_2p_in64k_out1k_prefix90_50ms(self):
-        """Run NPU performance test for Qwen3.6-27B-w8a8 in64k out1k prefix90 50ms"""
+        """Run NPU performance test for Qwen3.6-27B in64k out1k prefix90 50ms"""
         self.run_throughput()
 
 

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_1p_in3k5_out1k5_50ms_gpqa.py

@@ -58,8 +58,10 @@ QWEN3_6_27B_3K5_1K5_OTHER_ARGS = [
     8,
     16,
     32,
-    48,
-    64,
+    40,
+    45,
+    50,
+    54,
     "--enable-multimodal",
     "--quantization",
     "modelslim",
@@ -77,6 +79,10 @@ QWEN3_6_27B_3K5_1K5_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -106,11 +112,17 @@ class TestNPUQwen3_6_27B_1P_In3k5_Out1k5_gpqa(TestNpuAccuracyTestCaseBase):
     model = QWEN3_6_27B_W8A8_MODEL_PATH
     envs = QWEN3_6_27B_3K5_1K5_ENVS
     other_args = QWEN3_6_27B_3K5_1K5_OTHER_ARGS
-    accuracy = 0.855
+    accuracy = 0.878
     datasets = ["gpqa_diamond"]
     few_shot_num = 0
     eval_batch_size = 8
-    generation_config = {"max_tokens": 81920, "temperature": 1.0}
+    generation_config = {
+        "max_tokens": 81920,
+        "temperature": 1.0,
+        "extra_body": {
+            "chat_template_kwargs": {"enable_thinking": True},
+        },
+    }
 
     def test_accuracy(self):
         self.run_accuracy()

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_1p_in64k_out1k_50ms.py

@@ -0,0 +1,99 @@
+import unittest
+
+from sglang.test.ascend.e2e.test_npu_performance_utils import (
+    AISBENCHMARK_DATASET_DEFAULT,
+    BENCHMARK_TOOL_DEFAULT,
+    QWEN3_6_27B_W8A8_MODEL_PATH,
+    TestNpuPerformanceTestCaseBase,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(
+    est_time=3600,
+    suite="",
+    nightly=True,
+    disabled="performance testcase",
+)
+
+QWEN3_6_27B_64K_1K_ENVS = {
+    "STREAMS_PER_DEVICE": "32",
+    "HCCL_SOCKET_IFNAME": "lo",
+    "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",
+}
+
+QWEN3_6_27B_64K_1K_OTHER_ARGS = [
+    "--tp-size",
+    2,
+    "--nnodes",
+    1,
+    "--attention-backend",
+    "ascend",
+    "--device",
+    "npu",
+    "--chunked-prefill-size",
+    -1,
+    "--max-prefill-tokens",
+    48000,
+    "--disable-radix-cache",
+    "--trust-remote-code",
+    "--max-running-requests",
+    6,
+    "--max-mamba-cache-size",
+    16,
+    "--mem-fraction-static",
+    0.6,
+    "--cuda-graph-bs",
+    1,
+    2,
+    4,
+    5,
+    6,
+    "--quantization",
+    "modelslim",
+    "--dtype",
+    "bfloat16",
+    "--mamba-ssm-dtype",
+    "bfloat16",
+    "--speculative-algorithm",
+    "NEXTN",
+    "--speculative-num-steps",
+    3,
+    "--speculative-eagle-topk",
+    1,
+    "--speculative-num-draft-tokens",
+    4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
+]
+
+
+class TestNPUQwen3_6_27B_2P_In64k_Out1k_50ms(TestNpuPerformanceTestCaseBase):
+    """Test NPU performance for Qwen3.6-27B-w8a8 2p in64k out1k 50ms"""
+
+    benchmark_tool = BENCHMARK_TOOL_DEFAULT
+    dataset_type = AISBENCHMARK_DATASET_DEFAULT
+    model = QWEN3_6_27B_W8A8_MODEL_PATH
+    other_args = QWEN3_6_27B_64K_1K_OTHER_ARGS
+    envs = QWEN3_6_27B_64K_1K_ENVS
+    dataset_name = "random"
+    max_concurrency = 6
+    num_prompts = 12
+    input_len = 64000
+    output_len = 1000
+    random_range_ratio = 1
+    tpot = 50
+    output_token_throughput = 57.85
+
+    def test_npu_qwen3_6_27b_2p_in64k_out1k_50ms(self):
+        """Run NPU performance test for Qwen3.6-27B-w8a8 in64k out1k 50ms"""
+        self.run_throughput()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_2p_in128k_out1k_50ms.py

@@ -65,6 +65,10 @@ QWEN3_6_27B_128K_OTHER_ARGS = [
     "bfloat16",
     "--mamba-ssm-dtype",
     "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -72,7 +76,7 @@ class TestNPUQwen3_6_27B_2P_In128k_Out1k_50ms(TestNpuPerformanceTestCaseBase):
     """Test NPU performance for Qwen3.6-27B-w8a8 2p in128k out1k 50ms"""
 
     benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    aisbench_dataset_type = AISBENCHMARK_DATASET_DEFAULT
+    dataset_type = AISBENCHMARK_DATASET_DEFAULT
     model = QWEN3_6_27B_W8A8_MODEL_PATH
     other_args = QWEN3_6_27B_128K_OTHER_ARGS
     envs = QWEN3_6_27B_128K_ENVS

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

@@ -16,7 +16,6 @@ register_npu_ci(
 )
 
 QWEN3_6_27B_16K_1k_ENVS = {
-    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
     "STREAMS_PER_DEVICE": "32",
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
@@ -25,7 +24,8 @@ QWEN3_6_27B_16K_1k_ENVS = {
     "SGLANG_ENABLE_SPEC_V2": "1",
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
     "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
-    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "30",
+    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
+    "GDN_ATTN_BACKEND_TRITON": "1",
     "ASCEND_USE_FIA": "1",
 }
 
@@ -41,28 +41,28 @@ QWEN3_6_27B_16K_1k_OTHER_ARGS = [
     "--chunked-prefill-size",
     -1,
     "--max-prefill-tokens",
-    50000,
+    58000,
     "--disable-radix-cache",
     "--trust-remote-code",
     "--max-running-requests",
-    28,
+    29,
     "--max-mamba-cache-size",
-    50,
+    58,
     "--mem-fraction-static",
-    0.7,
+    0.68,
     "--cuda-graph-bs",
+    1,
     2,
     8,
     12,
     16,
     20,
     24,
+    26,
     28,
-    "--enable-multimodal",
+    29,
     "--quantization",
     "modelslim",
-    "--mm-attention-backend",
-    "ascend_attn",
     "--dtype",
     "bfloat16",
     "--mamba-ssm-dtype",
@@ -75,6 +75,10 @@ QWEN3_6_27B_16K_1k_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -82,13 +86,13 @@ class TestNPUQwen3_6_27B_2P_In16k_Out1k_50ms(TestNpuPerformanceTestCaseBase):
     """Test NPU performance for Qwen3.6-27B-w8a8 2p in16k out1k 50ms"""
 
     benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    aisbench_dataset_type = AISBENCHMARK_DATASET_DEFAULT
+    dataset_type = AISBENCHMARK_DATASET_DEFAULT
     model = QWEN3_6_27B_W8A8_MODEL_PATH
     other_args = QWEN3_6_27B_16K_1k_OTHER_ARGS
     envs = QWEN3_6_27B_16K_1k_ENVS
     dataset_name = "random"
-    max_concurrency = 28
-    num_prompts = 112
+    max_concurrency = 29
+    num_prompts = 116
     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_in128k_out1k_50ms.py

@@ -18,6 +18,7 @@ register_npu_ci(
 QWEN3_6_35B_A3B_128K_1K_ENVS = {
     "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
     "STREAMS_PER_DEVICE": "32",
+    "HCCL_BUFFSIZE": "1600",
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
@@ -26,6 +27,7 @@ QWEN3_6_35B_A3B_128K_1K_ENVS = {
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
     "ASCEND_USE_FIA": "1",
     "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "20",
+    "GDN_ATTN_BACKEND_TRITON": "1",
 }
 
 QWEN3_6_35B_A3B_128K_1K_OTHER_ARGS = [
@@ -40,21 +42,21 @@ QWEN3_6_35B_A3B_128K_1K_OTHER_ARGS = [
     "--chunked-prefill-size",
     -1,
     "--max-total-tokens",
-    600000,
+    420000,
     "--max-prefill-tokens",
-    65536,
+    128000,
     "--disable-radix-cache",
     "--trust-remote-code",
-    "--enable-prefill-delayer",
     "--max-running-requests",
-    4,
+    3,
     "--max-mamba-cache-size",
-    12,
+    3,
     "--mem-fraction-static",
-    0.6,
-    "--max-mamba-cache-size",
-    20,
-    "--disable-cuda-graph",
+    0.9,
+    "--cuda-graph-bs",
+    1,
+    2,
+    3,
     "--enable-multimodal",
     "--mm-attention-backend",
     "ascend_attn",
@@ -70,6 +72,10 @@ QWEN3_6_35B_A3B_128K_1K_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -82,11 +88,12 @@ class TestNPUQwen3_6_35BA3B_1P_In128k_Out1k_50ms(TestNpuPerformanceTestCaseBase)
     other_args = QWEN3_6_35B_A3B_128K_1K_OTHER_ARGS
     envs = QWEN3_6_35B_A3B_128K_1K_ENVS
     dataset_name = "random"
-    max_concurrency = 4
-    num_prompts = 16
+    max_concurrency = 3
+    num_prompts = 3
     input_len = 128000
     output_len = 1000
     random_range_ratio = 1
+    seed = 1
     tpot = 50
     output_token_throughput = 60.57
 

test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in128k_out1k_prefix90_50ms.py

@@ -76,6 +76,10 @@ QWEN3_6_35B_A3B_128K_PREFIX_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -91,10 +95,11 @@ class TestNPUQwen3_6_35BA3B_1P_In128k_Out1k_Prefix90_50ms(
     dataset_name = "generated-shared-prefix"
     max_concurrency = 103
     num_prompts = 412
-    input_len = 64000
+    input_len = 128000
     output_len = 1000
     random_range_ratio = 1
     repeat_rate = 0.9
+    seed = 1
     tpot = 50
     request_rate = float("inf")
     output_token_throughput = 308.2

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": "800",
+    "HCCL_BUFFSIZE": "1",
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
@@ -26,7 +26,6 @@ QWEN3_6_35B_A3B_3K5_1K5_ENVS = {
     "SGLANG_ENABLE_SPEC_V2": "1",
     "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
     "ASCEND_USE_FIA": "1",
-    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "50",
 }
 
 QWEN3_6_35B_A3B_3K5_1K5_OTHER_ARGS = [
@@ -40,25 +39,29 @@ QWEN3_6_35B_A3B_3K5_1K5_OTHER_ARGS = [
     "npu",
     "--chunked-prefill-size",
     -1,
+    "--max-total-tokens",
+    659840,
     "--max-prefill-tokens",
-    35000,
+    43400,
     "--disable-radix-cache",
     "--trust-remote-code",
-    "--enable-prefill-delayer",
+    "--prefill-max-requests",
+    "12",
     "--max-running-requests",
-    110,
+    122,
     "--max-mamba-cache-size",
-    115,
+    122,
     "--mem-fraction-static",
-    0.78,
+    0.9,
     "--cuda-graph-bs",
     4,
     16,
     32,
     64,
-    84,
-    105,
-    110,
+    96,
+    116,
+    120,
+    122,
     "--enable-multimodal",
     "--mm-attention-backend",
     "ascend_attn",
@@ -74,6 +77,10 @@ QWEN3_6_35B_A3B_3K5_1K5_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -86,11 +93,12 @@ class TestNPUQwen3_6_35BA3B_1P_In3k5_Out1k5_50ms(TestNpuPerformanceTestCaseBase)
     other_args = QWEN3_6_35B_A3B_3K5_1K5_OTHER_ARGS
     envs = QWEN3_6_35B_A3B_3K5_1K5_ENVS
     dataset_name = "random"
-    max_concurrency = 110
-    num_prompts = 440
+    max_concurrency = 122
+    num_prompts = 122
     input_len = 3500
     output_len = 1500
     random_range_ratio = 1
+    seed = 1
     tpot = 50
     output_token_throughput = 2031.71
 

test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_50ms.py

@@ -74,6 +74,10 @@ QWEN3_6_35B_A3B_64K_1K_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -91,6 +95,7 @@ class TestNPUQwen3_6_35BA3B_1P_In64k_Out1k_50ms(TestNpuPerformanceTestCaseBase):
     input_len = 64000
     output_len = 1000
     random_range_ratio = 1
+    seed = 1
     tpot = 50
     output_token_throughput = 141.72
 

test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_prefix90_50ms_aime26.py

@@ -43,18 +43,19 @@ QWEN3_6_35B_A3B_64K_PREFIX_OTHER_ARGS = [
     "npu",
     "--chunked-prefill-size",
     -1,
+    "--max-total-tokens",
+    470784,
     "--max-prefill-tokens",
     65536,
     "--trust-remote-code",
-    "--enable-prefill-delayer",
     "--mamba-scheduler-strategy",
     "extra_buffer",
     "--max-running-requests",
-    42,
+    40,
     "--max-mamba-cache-size",
-    210,
+    200,
     "--mem-fraction-static",
-    0.71,
+    0.9,
     "--cuda-graph-bs",
     2,
     8,
@@ -63,7 +64,6 @@ QWEN3_6_35B_A3B_64K_PREFIX_OTHER_ARGS = [
     32,
     36,
     40,
-    42,
     "--enable-multimodal",
     "--mm-attention-backend",
     "ascend_attn",
@@ -79,6 +79,10 @@ QWEN3_6_35B_A3B_64K_PREFIX_OTHER_ARGS = [
     1,
     "--speculative-num-draft-tokens",
     4,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -118,15 +122,17 @@ class TestNPUQwen3_6_35BA3B_1P_In64k_Out1k_Prefix90_50ms(
     other_args = QWEN3_6_35B_A3B_64K_PREFIX_OTHER_ARGS
     envs = QWEN3_6_35B_A3B_64K_PREFIX_ENVS
     dataset_name = "generated-shared-prefix"
-    max_concurrency = 42
-    num_prompts = 42
+    max_concurrency = 40
+    num_prompts = 40
     input_len = 65536
     output_len = 1024
     random_range_ratio = 1
     repeat_rate = 0.9
+    seed = 1
     tpot = 50
     request_rate = float("inf")
     output_token_throughput = 660
+    pop_sglang_is_in_ci_for_gsp = True
 
     @classmethod
     def setUpClass(cls):

test/registered/ascend/performance/qwen3_next_80b_a3b_instruct/test_npu_qwen3_next_80b_w8a8_2p_in6k_out1k5_bs16_aime25.py

@@ -91,6 +91,10 @@ QWEN3_NEXT_80B_A3B_OTHER_ARGS = [
     "bfloat16",
     "--speculative-draft-model-path",
     QWEN3_NEXT_80B_A3B_MODEL_PATH,
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen3_coder",
 ]
 
 
@@ -123,12 +127,12 @@ class TestQwen3Next80BA3B_aime25(TestNpuAccuracyTestCaseBase):
         "temperature": 0.7,
         "top_p": 0.8,
         "top_k": 20,
-        "extra_body": {"chat_template_kwargs": {"enable_thinking": True}},
+        "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
     }
     max_concurrency = 16
 
     def test_aime25(self):
-        self.run_accuracy()
+        self.run_accuracy_multiple(n_runs=3)
 
 
 if __name__ == "__main__":