[NPU]TP Communications compression For Qwen3 models for NPU (#20520)

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

test/registered/ascend/llm_models/test_npu_llama_2_7b_communications_compression.py

@@ -0,0 +1,37 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import LLAMA_2_7B_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="nightly-2-npu-a3")
+
+
+class TestLlama(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify that the inference accuracy of the LLM-Research/Llama-2-7B model on the GSM8K dataset with tp communications quantization is no less than 0.18.
+
+    [Test Category] Model
+    [Test Target] LLM-Research/Llama-2-7B
+    """
+
+    model = LLAMA_2_7B_WEIGHTS_PATH
+    accuracy = 0.18
+    other_args = [
+        "--trust-remote-code",
+        "--mem-fraction-static",
+        0.8,
+        "--max-running-requests",
+        32,
+        "--attention-backend",
+        "ascend",
+        "--cuda-graph-max-bs",
+        32,
+        "--tp-size",
+        2,
+        "--enable-quant-communications",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/ascend/llm_models/test_npu_qwen3_8b_communications_quantization.py

@@ -0,0 +1,37 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import QWEN3_8B_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="nightly-2-npu-a3")
+
+
+class TestQwen38BCommQuantization(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify that the inference accuracy of the Qwen/Qwen3-8B model with TP communications quantization on the GSM8K dataset is no less than 0.85.
+
+    [Test Category] Model
+    [Test Target] Qwen/Qwen3-8B
+    """
+
+    model = QWEN3_8B_WEIGHTS_PATH
+    accuracy = 0.85
+    other_args = [
+        "--trust-remote-code",
+        "--mem-fraction-static",
+        0.8,
+        "--max-running-requests",
+        32,
+        "--attention-backend",
+        "ascend",
+        "--cuda-graph-max-bs",
+        32,
+        "--tp-size",
+        2,
+        "--enable-quant-communications",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()