Add pr tests (#31952)

Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
Co-authored-by: Cherry_ming <136634645@qq.com>

test/registered/README.md

@@ -18,6 +18,6 @@ Tests under this directory are auto-discovered by `run_suite.py` via CI registra
 | Model inference correctness | `models/`, `4-gpu-models/`, `8-gpu-models/` | GPU |
 | Feature-specific (OpenAI API, LoRA, speculative, distributed, VLM, etc.) | `openai_server/`, `lora/`, `spec/`, `distributed/`, ... | GPU |
 | Benchmarks (performance, accuracy, stress) | `benchmark/` | GPU |
-| Platform-specific | `amd/`, `ascend/` | Vendor GPU |
+| Platform-specific | `amd/`, `npu/` | Vendor GPU |
 
 See [`unit/README.md`](unit/README.md) for unit test conventions.

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

@@ -5,10 +5,7 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
 )
 from sglang.test.ascend.e2e.test_npu_multi_node_utils import NIC_NAME
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
-    AISBENCHMARK_DATASET_DEFAULT,
-    BENCHMARK_TOOL_DEFAULT,
     GLM_5_1_W4A8_MODEL_PATH,
-    TestNpuPerfMultiNodePdSepTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -16,7 +13,7 @@ register_npu_ci(
     est_time=3600,
     suite="",
     nightly=True,
-    disabled="performance testcase",
+    disabled="accuracy testcase",
 )
 
 GLM_5_1_PD_SEP_PREFILL_ENVS = {
@@ -181,26 +178,5 @@ class TestNPUGLM5_1_W4A8_PD_SEP_AIME2026(TestNpuAccuracyMultiNodePdSepTestCaseBa
         self.run_accuracy()
 
 
-class TestNPUGLM5_1_W4A8_PD_SEP_In3k5_Out1k5(TestNpuPerfMultiNodePdSepTestCaseBase):
-    """Test NPU performance for GLM-5.1-w4a8 PD separation 4 nodes in3k5 out1k5"""
-
-    model_config = GLM_5_1_PD_SEP_MODEL_CONFIG
-    benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    dataset_type = AISBENCHMARK_DATASET_DEFAULT
-    dataset_name = "random"
-    max_concurrency = 1
-    num_prompts = 1
-    input_len = 65536
-    output_len = 1024
-    random_range_ratio = 1
-    seed = 1
-    tpot = 50
-    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"""
-        self.run_throughput()
-
-
 if __name__ == "__main__":
     unittest.main()

test/registered/npu/accuracy/kimi_k2_6/test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms_aime25.py

@@ -5,11 +5,8 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
 )
 from sglang.test.ascend.e2e.test_npu_multi_node_utils import NIC_NAME
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
-    AISBENCHMARK_DATASET_DEFAULT,
-    BENCHMARK_TOOL_DEFAULT,
     KIMI_K2_6_EAGLE3_MODEL_PATH,
     KIMI_K2_6_W4A8_MODEL_PATH,
-    TestNpuPerfMultiNodePdMixTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -17,7 +14,7 @@ register_npu_ci(
     est_time=1800,
     suite="nightly-8-npu-a3",
     nightly=True,
-    disabled="Currently it is executed by the npu performance workflow.",
+    disabled="accuracy testcase",
 )
 
 ENVS = {
@@ -107,23 +104,5 @@ class TestNPUKimiK2_6_W4A8_16P_AIME2025(TestNpuAccuracyMultiNodePdMixTestCaseBas
         self.run_accuracy()
 
 
-class TestNPUKimiK2_6_W4A8_16P_In64k_Out1k_100ms(TestNpuPerfMultiNodePdMixTestCaseBase):
-    benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    dataset_type = AISBENCHMARK_DATASET_DEFAULT
-    model_config = MODEL_CONFIG
-    dataset_name = "random"
-    max_concurrency = 32
-    num_prompts = 32
-    input_len = 64000
-    output_len = 1000
-    random_range_ratio = 1
-    seed = 1
-    tpot = 100
-    output_token_throughput = 160
-
-    def test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms(self):
-        self.run_throughput()
-
-
 if __name__ == "__main__":
     unittest.main()

test/registered/npu/accuracy/minimax_m2_5/test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms_gpqa.py

@@ -0,0 +1,109 @@
+import os
+import unittest
+
+from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
+    TestNpuAccuracyTestCaseBase,
+)
+from sglang.test.ascend.e2e.test_npu_performance_utils import (
+    MINIMAX_M2_5_EAGLE3_MODEL_PATH,
+    MINIMAX_M2_5_W8A8_MODEL_PATH,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(
+    est_time=3600,
+    suite="",
+    nightly=True,
+    disabled="accuracy testcase",
+)
+
+MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS = {
+    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
+    "STREAMS_PER_DEVICE": "32",
+    "HCCL_SOCKET_IFNAME": "lo",
+    "GLOO_SOCKET_IFNAME": "lo",
+    "TASK_QUEUE_ENABLE": "1",
+    "ASCEND_USE_FIA": "1",
+    "SGLANG_SET_CPU_AFFINITY": "1",
+    "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
+    "SGLANG_NPU_FUSED_MOE_MODE": "2",
+    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "140000",
+    "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
+    "DEEPEP_HCCL_BUFFSIZE": "1024",
+    "SGLANG_EXTERNAL_MODEL_PACKAGE": "custom_eagle3",
+    "PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
+}
+
+MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS = [
+    "--tp-size",
+    8,
+    "--mem-fraction-static",
+    0.63,
+    "--max-running-requests",
+    26,
+    "--reasoning-parser",
+    "minimax-append-think",
+    "--tool-call-parser",
+    "minimax-m2",
+    "--enable-prefill-delayer",
+    "--prefill-max-requests",
+    10,
+    "--chunked-prefill-size",
+    67072,
+    "--max-prefill-token",
+    67000,
+    "--cuda-graph-bs",
+    2,
+    4,
+    8,
+    12,
+    16,
+    18,
+    20,
+    22,
+    24,
+    26,
+    "--moe-a2a-backend",
+    "ascend_fuseep",
+    "--deepep-mode",
+    "auto",
+    "--quantization",
+    "modelslim",
+    "--speculative-algorithm",
+    "EAGLE3",
+    "--speculative-draft-model-path",
+    MINIMAX_M2_5_EAGLE3_MODEL_PATH,
+    "--speculative-num-steps",
+    3,
+    "--speculative-eagle-topk",
+    1,
+    "--speculative-num-draft-tokens",
+    4,
+    "--speculative-draft-model-quantization",
+    "unquant",
+    "--dtype",
+    "bfloat16",
+    "--trust-remote-code",
+    "--reasoning-parser",
+    "minimax-append-think",
+    "--tool-call-parser",
+    "minimax-m2",
+]
+
+
+class TestNPUMiniMaxM2_5_W8A8_4P_Gpqa(TestNpuAccuracyTestCaseBase):
+    model = MINIMAX_M2_5_W8A8_MODEL_PATH
+    other_args = MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_OTHER_ARGS
+    envs = MINIMAX_M2_5_W8A8_4P_IN64K_OUT1K_PREFIX90_ENVS
+    accuracy = 0.852
+    datasets = ["gpqa_diamond"]
+    few_shot_num = 0
+    generation_config = {"max_tokens": 65536, "temperature": 1.0}
+    eval_batch_size = 64
+
+    def test_accuracy(self):
+        self.run_accuracy()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/npu/accuracy/minimax_m2_5/test_npu_minimax_m2_5_w8a8_8p_in3k5_out1k5_50ms_gpqa.py

@@ -0,0 +1,107 @@
+import os
+import unittest
+
+from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
+    TestNpuAccuracyTestCaseBase,
+)
+from sglang.test.ascend.e2e.test_npu_performance_utils import (
+    MINIMAX_M2_5_EAGLE3_MODEL_PATH,
+    MINIMAX_M2_5_W8A8_MODEL_PATH,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(
+    est_time=3600,
+    suite="full-16-npu-a3",
+    nightly=True,
+    disabled="accuracy testcase",
+)
+
+MINIMAX_M2_5_HIGH_THROUGHPUT_ENVS = {
+    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
+    "STREAMS_PER_DEVICE": "32",
+    "HCCL_SOCKET_IFNAME": "lo",
+    "GLOO_SOCKET_IFNAME": "lo",
+    "TASK_QUEUE_ENABLE": "1",
+    "DEEPEP_HCCL_BUFFSIZE": "1024",
+    "ASCEND_USE_FIA": "1",
+    "SGLANG_SET_CPU_AFFINITY": "1",
+    "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
+    "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "204800",
+    "PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
+    "SGLANG_EXTERNAL_MODEL_PACKAGE": "custom_eagle3",
+}
+
+MINIMAX_M2_5_HIGH_THROUGHPUT_OTHER_ARGS = [
+    "--tp-size",
+    16,
+    "--enable-dp-attention",
+    "--dp-size",
+    16,
+    "--mem-fraction-static",
+    0.75,
+    "--max-running-requests",
+    320,
+    "--disable-radix-cache",
+    "--reasoning-parser",
+    "minimax-append-think",
+    "--tool-call-parser",
+    "minimax-m2",
+    "--prefill-delayer-max-delay-passes",
+    500,
+    "--enable-prefill-delayer",
+    "--chunked-prefill-size",
+    196608,
+    "--max-prefill-token",
+    8192,
+    "--cuda-graph-bs",
+    1,
+    2,
+    4,
+    8,
+    12,
+    16,
+    20,
+    "--moe-a2a-backend",
+    "ascend_fuseep",
+    "--fuseep-mode",
+    2,
+    "--quantization",
+    "modelslim",
+    "--speculative-algorithm",
+    "EAGLE3",
+    "--speculative-draft-model-path",
+    MINIMAX_M2_5_EAGLE3_MODEL_PATH,
+    "--speculative-num-steps",
+    3,
+    "--speculative-eagle-topk",
+    1,
+    "--speculative-num-draft-tokens",
+    4,
+    "--speculative-draft-model-quantization",
+    "unquant",
+    "--dtype",
+    "bfloat16",
+    "--reasoning-parser",
+    "minimax-append-think",
+    "--tool-call-parser",
+    "minimax-m2",
+]
+
+
+class TestNPUMiniMaxM2_5_W8A8_8P_In3k5_Out1k5_GPQA(TestNpuAccuracyTestCaseBase):
+    model = MINIMAX_M2_5_W8A8_MODEL_PATH
+    envs = MINIMAX_M2_5_HIGH_THROUGHPUT_ENVS
+    other_args = MINIMAX_M2_5_HIGH_THROUGHPUT_OTHER_ARGS
+    accuracy = 0.852
+    datasets = ["gpqa_diamond"]
+    few_shot_num = 0
+    generation_config = {"max_tokens": 65536, "temperature": 1.0}
+    eval_batch_size = 64
+
+    def test_accuracy(self):
+        self.run_accuracy()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/npu/accuracy/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in3k5_out1k5_50ms_gpqa.py

@@ -4,11 +4,8 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
     TestNpuAccuracyTestCaseBase,
 )
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
-    AISBENCHMARK_DATASET_DEFAULT,
-    BENCHMARK_TOOL_DEFAULT,
     QWEN3_8B_EAGLE_MODEL_PATH,
     QWEN3_8B_W8A8_MODEL_PATH,
-    TestNpuPerformanceTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -16,7 +13,7 @@ register_npu_ci(
     est_time=3600,
     suite="",
     nightly=True,
-    disabled="performance testcase",
+    disabled="accuracy testcase",
 )
 
 QWEN3_8B_ENVS = {
@@ -88,26 +85,6 @@ QWEN3_8B_OTHER_ARGS = [
 ]
 
 
-class TestQwen8B(TestNpuPerformanceTestCaseBase):
-    benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    dataset_type = AISBENCHMARK_DATASET_DEFAULT
-    model = QWEN3_8B_W8A8_MODEL_PATH
-    other_args = QWEN3_8B_OTHER_ARGS
-    envs = QWEN3_8B_ENVS
-    dataset_name = "random"
-    max_concurrency = 64
-    num_prompts = 256
-    input_len = 3500
-    output_len = 1500
-    random_range_ratio = 1
-    seed = 1
-    tpot = 37
-    output_token_throughput = 1586
-
-    def test_qwen3_8b(self):
-        self.run_throughput()
-
-
 class TestQwen8B_gpqa(TestNpuAccuracyTestCaseBase):
     model = QWEN3_8B_W8A8_MODEL_PATH
     envs = QWEN3_8B_ENVS

test/registered/npu/accuracy/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in6k_out1k5_bs16_gpqa.py

@@ -6,7 +6,6 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
     QWEN3_8B_EAGLE_MODEL_PATH,
     QWEN3_8B_W8A8_MODEL_PATH,
-    TestNpuPerformanceTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -14,7 +13,7 @@ register_npu_ci(
     est_time=3600,
     suite="",
     nightly=True,
-    disabled="performance testcase",
+    disabled="accuracy testcase",
 )
 
 QWEN3_8B_ENVS = {
@@ -75,24 +74,6 @@ QWEN3_8B_OTHER_ARGS = [
 ]
 
 
-class TestQwen8B(TestNpuPerformanceTestCaseBase):
-    max_attempts = 5
-    model = QWEN3_8B_W8A8_MODEL_PATH
-    other_args = QWEN3_8B_OTHER_ARGS
-    envs = QWEN3_8B_ENVS
-    dataset_name = "random"
-    max_concurrency = 16
-    num_prompts = 16
-    input_len = 6144
-    output_len = 1500
-    random_range_ratio = 1
-    tpot = 11.79
-    output_token_throughput = 1040.96
-
-    def test_qwen3_8b(self):
-        self.run_throughput()
-
-
 class TestQwen8B_gpqa(TestNpuAccuracyTestCaseBase):
     model = QWEN3_8B_W8A8_MODEL_PATH
     envs = QWEN3_8B_ENVS

test/registered/npu/accuracy/qwen3_30b_a3b/test_npu_qwen3_30b_w8a8_1p_in3k5_out1k5_50ms_aime25.py

@@ -4,11 +4,8 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
     TestNpuAccuracyTestCaseBase,
 )
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
-    AISBENCHMARK_DATASET_DEFAULT,
-    BENCHMARK_TOOL_DEFAULT,
     QWEN3_30B_A3B_W8A8_VLLM_MODEL_PATH,
     QWEN3_A3B_EAGLE_MODEL_PATH,
-    TestNpuPerformanceTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -16,7 +13,7 @@ register_npu_ci(
     est_time=3600,
     suite="",
     nightly=True,
-    disabled="performance testcase",
+    disabled="accuracy testcase",
 )
 
 QWEN3_30B_A3B_ENVS = {
@@ -93,26 +90,6 @@ QWEN3_30B_A3B_OTHER_ARGS = [
 ]
 
 
-class TestQwen30B(TestNpuPerformanceTestCaseBase):
-    benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    dataset_type = AISBENCHMARK_DATASET_DEFAULT
-    model = QWEN3_30B_A3B_W8A8_VLLM_MODEL_PATH
-    other_args = QWEN3_30B_A3B_OTHER_ARGS
-    envs = QWEN3_30B_A3B_ENVS
-    dataset_name = "random"
-    max_concurrency = 160
-    num_prompts = int(max_concurrency) * 4
-    input_len = 3500
-    output_len = 1500
-    random_range_ratio = 1
-    seed = 1
-    tpot = 50
-    output_token_throughput = 3200
-
-    def test_qwen3_30b(self):
-        self.run_throughput()
-
-
 class TestQwen30B_A3B_aime25(TestNpuAccuracyTestCaseBase):
     model = QWEN3_30B_A3B_W8A8_VLLM_MODEL_PATH
     envs = QWEN3_30B_A3B_ENVS

test/registered/npu/accuracy/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_gpqa.py

@@ -4,11 +4,8 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
     TestNpuAccuracyTestCaseBase,
 )
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
-    AISBENCHMARK_DATASET_DEFAULT,
-    BENCHMARK_TOOL_DEFAULT,
     QWEN3_32B_EAGLE_MODEL_PATH,
     QWEN3_32B_W8A8_MODEL_PATH,
-    TestNpuPerformanceTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -16,7 +13,7 @@ register_npu_ci(
     est_time=3600,
     suite="",
     nightly=True,
-    disabled="performance testcase",
+    disabled="accuracy testcase",
 )
 
 QWEN3_32B_ENVS = {
@@ -90,27 +87,7 @@ QWEN3_32B_OTHER_ARGS = [
 ]
 
 
-class TestQwen32B(TestNpuPerformanceTestCaseBase):
-    benchmark_tool = BENCHMARK_TOOL_DEFAULT
-    dataset_type = AISBENCHMARK_DATASET_DEFAULT
-    model = QWEN3_32B_W8A8_MODEL_PATH
-    other_args = QWEN3_32B_OTHER_ARGS
-    envs = QWEN3_32B_ENVS
-    dataset_name = "random"
-    max_concurrency = 100
-    num_prompts = 400
-    input_len = 3584
-    output_len = 1536
-    random_range_ratio = 1
-    seed = 1
-    tpot = 50
-    output_token_throughput = 1600
-
-    def test_qwen3_32b(self):
-        self.run_throughput()
-
-
-class TestQwen32B_mmlupro(TestNpuAccuracyTestCaseBase):
+class TestQwen32B_gpqa(TestNpuAccuracyTestCaseBase):
     model = QWEN3_32B_W8A8_MODEL_PATH
     envs = QWEN3_32B_ENVS
     other_args = QWEN3_32B_OTHER_ARGS

test/registered/npu/accuracy/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_gpqa_a2.py

@@ -0,0 +1,108 @@
+import unittest
+
+from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
+    TestNpuAccuracyTestCaseBase,
+)
+from sglang.test.ascend.e2e.test_npu_performance_utils import (
+    QWEN3_32B_EAGLE_MODEL_PATH,
+    QWEN3_32B_W8A8_MODEL_PATH,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(
+    est_time=3600,
+    suite="",
+    nightly=True,
+    disabled="accuracy testcase",
+)
+
+QWEN3_32B_ENVS = {
+    "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_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
+    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
+    "SGLANG_NPU_USE_DEEPGEMM": "1",
+}
+
+QWEN3_32B_OTHER_ARGS = [
+    "--trust-remote-code",
+    "--nnodes",
+    "1",
+    "--node-rank",
+    "0",
+    "--attention-backend",
+    "ascend",
+    "--device",
+    "npu",
+    "--quantization",
+    "modelslim",
+    "--max-running-requests",
+    101,
+    "--disable-radix-cache",
+    "--speculative-draft-model-quantization",
+    "unquant",
+    "--chunked-prefill-size",
+    -1,
+    "--max-prefill-tokens",
+    35000,
+    "--speculative-algorithm",
+    "EAGLE3",
+    "--speculative-draft-model-path",
+    QWEN3_32B_EAGLE_MODEL_PATH,
+    "--speculative-num-steps",
+    3,
+    "--speculative-eagle-topk",
+    1,
+    "--speculative-num-draft-tokens",
+    4,
+    "--tp-size",
+    4,
+    "--mem-fraction-static",
+    0.845,
+    "--cuda-graph-bs",
+    16,
+    32,
+    64,
+    72,
+    88,
+    90,
+    92,
+    94,
+    96,
+    97,
+    98,
+    99,
+    100,
+    101,
+    "--dtype",
+    "bfloat16",
+    "--reasoning-parser",
+    "qwen3",
+    "--tool-call-parser",
+    "qwen",
+]
+
+
+class TestQwen32B_GPQA(TestNpuAccuracyTestCaseBase):
+    """Test NPU accuracy for Qwen3-32B-W8A8 on qpqa"""
+
+    model = QWEN3_32B_W8A8_MODEL_PATH
+    other_args = QWEN3_32B_OTHER_ARGS
+    envs = QWEN3_32B_ENVS
+    accuracy = 0.4949
+    datasets = ["gpqa_diamond"]
+    few_shot_num = 0
+    eval_batch_size = 64
+    generation_config = {"max_tokens": 40000, "temperature": 1.0}
+
+    def test_qwen3_32b_qpqa(self):
+        """Run NPU accuracy test for Qwen3-32B-W8A8 on qpqa"""
+        self.run_accuracy()
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -0,0 +1,106 @@
+import unittest
+
+from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
+    TestNpuAccuracyTestCaseBase,
+)
+from sglang.test.ascend.e2e.test_npu_performance_utils import (
+    QWEN3_6_27B_W8A8_MODEL_PATH,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(
+    est_time=3600,
+    suite="full-2-npu-a3",
+    nightly=True,
+    disabled="accuracy testcase",
+)
+
+QWEN3_6_27B_3K5_1K5_ENVS = {
+    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
+    "STREAMS_PER_DEVICE": "32",
+    "HCCL_SOCKET_IFNAME": "lo",
+    "GLOO_SOCKET_IFNAME": "lo",
+    "HCCL_OP_EXPANSION_MODE": "AIV",
+    "SGLANG_SET_CPU_AFFINITY": "1",
+    "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
+    "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1",
+    "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "130",
+    "ASCEND_USE_FIA": "1",
+}
+
+QWEN3_6_27B_3K5_1K5_OTHER_ARGS = [
+    "--tp-size",
+    2,
+    "--nnodes",
+    1,
+    "--attention-backend",
+    "ascend",
+    "--device",
+    "npu",
+    "--chunked-prefill-size",
+    -1,
+    "--max-prefill-tokens",
+    60000,
+    "--disable-radix-cache",
+    "--trust-remote-code",
+    "--max-running-requests",
+    64,
+    "--max-mamba-cache-size",
+    74,
+    "--mem-fraction-static",
+    0.7,
+    "--cuda-graph-bs",
+    2,
+    8,
+    16,
+    32,
+    40,
+    45,
+    50,
+    54,
+    "--enable-multimodal",
+    "--quantization",
+    "modelslim",
+    "--mm-attention-backend",
+    "ascend_attn",
+    "--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_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 = 0.878
+    datasets = ["gpqa_diamond"]
+    few_shot_num = 0
+    eval_batch_size = 8
+    generation_config = {
+        "max_tokens": 81920,
+        "temperature": 1.0,
+        "extra_body": {
+            "chat_template_kwargs": {"enable_thinking": True},
+        },
+    }
+
+    def test_accuracy(self):
+        self.run_accuracy()
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/npu/accuracy/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_prefix90_50ms_aime26.py

@@ -0,0 +1,107 @@
+import unittest
+
+from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
+    TestNpuAccuracyTestCaseBase,
+)
+from sglang.test.ascend.e2e.test_npu_performance_utils import (
+    QWEN3_6_35B_A3B_MODEL_PATH,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+
+register_npu_ci(
+    est_time=3600,
+    suite="",
+    nightly=True,
+    disabled="accuracy testcase",
+)
+
+QWEN3_6_35B_A3B_64K_PREFIX_ENVS = {
+    "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
+    "STREAMS_PER_DEVICE": "32",
+    "DEEPEP_HCCL_BUFFSIZE": "300",
+    "HCCL_SOCKET_IFNAME": "lo",
+    "GLOO_SOCKET_IFNAME": "lo",
+    "HCCL_OP_EXPANSION_MODE": "AIV",
+    "SGLANG_SET_CPU_AFFINITY": "1",
+    "SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "0",
+    "ASCEND_USE_FIA": "1",
+    "GDN_ATTN_BACKEND_TRITON": "1",
+}
+
+QWEN3_6_35B_A3B_64K_PREFIX_OTHER_ARGS = [
+    "--tp-size",
+    2,
+    "--nnodes",
+    1,
+    "--attention-backend",
+    "ascend",
+    "--device",
+    "npu",
+    "--chunked-prefill-size",
+    -1,
+    "--max-total-tokens",
+    470784,
+    "--max-prefill-tokens",
+    65536,
+    "--trust-remote-code",
+    "--mamba-scheduler-strategy",
+    "extra_buffer",
+    "--max-running-requests",
+    40,
+    "--max-mamba-cache-size",
+    200,
+    "--mem-fraction-static",
+    0.9,
+    "--cuda-graph-bs",
+    2,
+    8,
+    16,
+    24,
+    32,
+    36,
+    40,
+    "--enable-multimodal",
+    "--mm-attention-backend",
+    "ascend_attn",
+    "--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_35BA3B_1P_AIME2026(TestNpuAccuracyTestCaseBase):
+    """Test NPU accuracy for Qwen3.6-35B-A3B 1p on AIME2026"""
+
+    model = QWEN3_6_35B_A3B_MODEL_PATH
+    other_args = QWEN3_6_35B_A3B_64K_PREFIX_OTHER_ARGS
+    envs = QWEN3_6_35B_A3B_64K_PREFIX_ENVS
+    accuracy = 0.927
+    datasets = ["aime26"]
+    few_shot_num = 0
+    eval_batch_size = 64
+    generation_config = {
+        "max_tokens": 65536,
+        "temperature": 0.2,
+        "repetition_penalty": 1.08,
+    }
+
+    def test_npu_qwen3_6_35b_a3b_1p_aime2026(self):
+        """Run NPU accuracy test for Qwen3.6-35B-A3B on AIME2026"""
+        self.run_accuracy()
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -6,7 +6,6 @@ from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
 from sglang.test.ascend.e2e.test_npu_performance_utils import (
     QWEN3_NEXT_80B_A3B_MODEL_PATH,
     QWEN3_NEXT_80B_A3B_W8A8_MODEL_PATH,
-    TestNpuPerformanceTestCaseBase,
 )
 from sglang.test.ci.ci_register import register_npu_ci
 
@@ -14,7 +13,7 @@ register_npu_ci(
     est_time=3600,
     suite="",
     nightly=True,
-    disabled="performance testcase",
+    disabled="accuracy testcase",
 )
 
 QWEN3_NEXT_80B_A3B_ENVS = {
@@ -97,23 +96,6 @@ QWEN3_NEXT_80B_A3B_OTHER_ARGS = [
 ]
 
 
-class TestQwen3Next80BA3B(TestNpuPerformanceTestCaseBase):
-    max_attempts = 5
-    model = QWEN3_NEXT_80B_A3B_W8A8_MODEL_PATH
-    other_args = QWEN3_NEXT_80B_A3B_OTHER_ARGS
-    envs = QWEN3_NEXT_80B_A3B_ENVS
-    dataset_name = "random"
-    max_concurrency = 16
-    num_prompts = 16
-    input_len = 6144
-    output_len = 1500
-    random_range_ratio = 1
-    tpot = 15.62
-
-    def test_qwen3_next_80b_a3b(self):
-        self.run_throughput()
-
-
 class TestQwen3Next80BA3B_aime25(TestNpuAccuracyTestCaseBase):
     model = QWEN3_NEXT_80B_A3B_W8A8_MODEL_PATH
     envs = QWEN3_NEXT_80B_A3B_ENVS

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

@@ -1,3 +1,4 @@
+import os
 import unittest
 
 from sglang.test.ascend.e2e.test_npu_accuracy_utils import (
@@ -15,6 +16,8 @@ register_npu_ci(
     disabled="performance testcase",
 )
 
+_is_pr_pipeline = os.environ.get("GITHUB_EVENT_NAME") == "pull_request"
+
 ENVS = {
     "SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT": "600",
     "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
@@ -57,9 +60,10 @@ class TestQwen3(TestNpuAccuracyTestCaseBase):
     model = QWEN3_VL_8B_THINKING_MODEL_PATH
     envs = ENVS
     other_args = OTHER_ARGS
-    accuracy = 0.7011
+    accuracy = 0.64 if _is_pr_pipeline else 0.7011
     datasets = ["mmmu"]
     few_shot_num = 0
+    limit = 5 if _is_pr_pipeline else 100000
     generation_config = {"max_tokens": 65536}
     eval_batch_size = 64
 

test/registered/npu/basic_function/optimization_debug/test_npu_graph_tp1_bf16.py

@@ -12,8 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a2", nightly=False)
-register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-1-npu-a3", nightly=True)
 
 TEST_MODEL_MATRIX = {
     "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {

test/registered/npu/basic_function/optimization_debug/test_npu_graph_tp2_bf16.py

@@ -12,8 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a2", nightly=False)
-register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+register_npu_ci(est_time=400, suite="full-2-npu-a3", nightly=True)
 
 TEST_MODEL_MATRIX = {
     "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {

test/registered/npu/basic_function/quant/test_npu_autoround_moe.py

@@ -10,7 +10,6 @@ from sglang.test.ascend.test_ascend_utils import (
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
 from sglang.test.test_utils import (
-    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
     popen_launch_server,
@@ -53,7 +52,7 @@ class TestAscendAutoRoundMoE(CustomTestCase):
                 process = popen_launch_server(
                     model,
                     self.base_url,
-                    timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+                    timeout=3600,
                     other_args=[
                         *self.common_args,
                     ],