[MoE Refactor] [NPU] Refactor Ascend MoE implementation to reduce code duplication and align with community design (#25663)

Co-authored-by: ronnie_zheng <zl19940307@163.com>

test/registered/ascend/llm_models/test_npu_qwen3_30b_fuseep.py

@@ -0,0 +1,45 @@
+import os
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import QWEN3_30B_A3B_W8A8_WEIGHTS_PATH
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_npu_ci(est_time=400, suite="per-commit-4-npu-a3")
+
+
+class TestQwen330Bw8a8FuseEP(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen3-30B-A3B-w8a8 model on the GSM8K dataset is no less than 0.90.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen3-30B-A3B-w8a8
+    """
+
+    model = QWEN3_30B_A3B_W8A8_WEIGHTS_PATH
+    accuracy = 0.90
+    other_args = [
+        "--trust-remote-code",
+        "--mem-fraction-static",
+        0.8,
+        "--attention-backend",
+        "ascend",
+        "--cuda-graph-max-bs-decode",
+        128,
+        "--tp-size",
+        4,
+        "--moe-a2a-backend",
+        "ascend_fuseep",
+        "--fuseep-mode",
+        2,
+    ]
+
+    env = {
+        **os.environ,
+        "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "100",
+        "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "188416",
+    }
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/llm_models/test_npu_qwen3_30b_w4a4.py

@@ -1,28 +1,28 @@
 import unittest
 
 from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
-from sglang.test.ascend.test_ascend_utils import QWEN3_30B_MODELSLIM_INT4_WEIGHTS_PATH
+from sglang.test.ascend.test_ascend_utils import (
+    ECO_TECH_QWEN3_30B_A3B_W4A4_LAOS_WEIGHTS_PATH,
+)
 from sglang.test.ci.ci_register import register_npu_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_npu_ci(est_time=400, suite="per-commit-2-npu-a2")
 
 
-class TestQwen317BGPTQInt8(GSM8KAscendMixin, CustomTestCase):
+class TestQwen330Bw4a4(GSM8KAscendMixin, CustomTestCase):
     """Testcase: Verify that the inference accuracy of the Eco-Tech/Qwen3-30B-A3B-w4a4-LAOS model on the GSM8K dataset is no less than 0.85.
 
     [Test Category] Model
-    [Test Target] Qwen/Qwen3-1.7B-GPTQ-Int8
+    [Test Target] Eco-Tech/Qwen3-30B-A3B-w4a4-LAOS
     """
 
-    model = QWEN3_30B_MODELSLIM_INT4_WEIGHTS_PATH
+    model = ECO_TECH_QWEN3_30B_A3B_W4A4_LAOS_WEIGHTS_PATH
     accuracy = 0.85
     other_args = [
         "--trust-remote-code",
         "--mem-fraction-static",
         0.8,
-        "--max-running-requests",
-        32,
         "--attention-backend",
         "ascend",
         "--disable-cuda-graph",

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

@@ -30,7 +30,6 @@ MINIMAX_M2_5_HIGH_THROUGHPUT_ENVS = {
     "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": "204800",
     "PYTHONPATH": f"{MINIMAX_M2_5_EAGLE3_MODEL_PATH}:{os.environ.get('PYTHONPATH', '')}",
     "SGLANG_EXTERNAL_MODEL_PACKAGE": "custom_eagle3",
@@ -68,8 +67,8 @@ MINIMAX_M2_5_HIGH_THROUGHPUT_OTHER_ARGS = [
     20,
     "--moe-a2a-backend",
     "ascend_fuseep",
-    "--deepep-mode",
-    "auto",
+    "--fuseep-mode",
+    2,
     "--quantization",
     "modelslim",
     "--speculative-algorithm",