[Test] add a new Qwen3-32b-int8 test case with feature_stack3 (#3676)

### What this PR does / why we need it?
This PR add a new Qwen3-32b-int8 test case for nightly test. This test
case mainly test the performance and accuracy of Qwen3-32b-int8 with a
new feature.

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
By running the test.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: root <root@hostname-2pbfv.foreman.pxe>
Co-authored-by: root <root@hostname-2pbfv.foreman.pxe>

.github/workflows/vllm_ascend_test_nightly.yaml

@@ -98,6 +98,17 @@ jobs:
       runner: ${{ matrix.os }}
       image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.2.rc1-a3-ubuntu22.04-py3.11
       tests: tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py
+  qwen3-32b-int8-a3-feature-stack3:
+    if: contains(github.event.pull_request.labels.*.name, 'run-nightly')
+    strategy:
+      matrix:
+        os: [ linux-aarch64-a3-4 ]
+    uses: ./.github/workflows/_e2e_nightly.yaml
+    with:
+      vllm: v0.11.0
+      runner: ${{ matrix.os }}
+      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.2.rc1-a3-ubuntu22.04-py3.11
+      tests: tests/e2e/nightly/features/test_qwen3_32b_int8_a3_feature_stack3.py
   qwen2-5-vl-7b:
     if: contains(github.event.pull_request.labels.*.name, 'run-nightly')
     strategy:

tests/e2e/nightly/features/test_qwen3_32b_int8_a3_feature_stack3.py

@@ -0,0 +1,106 @@
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+from typing import Any
+
+import openai
+import pytest
+from vllm.utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/Qwen3-32B-W8A8",
+]
+
+TENSOR_PARALLELS = [4]
+
+prompts = [
+    "San Francisco is a",
+]
+
+api_keyword_args = {
+    "max_tokens": 10,
+}
+
+aisbench_cases = [{
+    "case_type": "accuracy",
+    "dataset_path": "vllm-ascend/gsm8k-lite",
+    "request_conf": "vllm_api_general_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_noncot_chat_prompt",
+    "max_out_len": 10240,
+    "batch_size": 32,
+    "baseline": 96,
+    "threshold": 4
+}, {
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/GSM8K-in3500-bs400",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 240,
+    "max_out_len": 1500,
+    "batch_size": 60,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
+async def test_models(model: str, tp_size: int) -> None:
+    port = get_open_port()
+    env_dict = {
+        "VLLM_USE": "1",
+        "TASK_QUEUE_ENABLE": "1",
+        "HCCL_OP_EXPANSION_MODE": "AIV",
+        "OMP_PROC_BIND": "false",
+        "VLLM_ASCEND_ENABLE_TOPK_OPTIMIZE": "1",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM": "1",
+        "VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE": "1",
+        "VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"
+    }
+    server_args = [
+        "--quantization", "ascend", "--tensor-parallel-size",
+        str(tp_size), "--port",
+        str(port), "--trust-remote-code", "--reasoning-parser", "qwen3",
+        "--distributed_executor_backend", "mp", "--gpu-memory-utilization",
+        "0.9", "--block-size", "128", "--max-num-seqs", "256",
+        "--enforce-eager", "--max-model-len", "35840",
+        "--max-num-batched-tokens", "35840", "--additional-config",
+        '{"ascend_scheduler_config":{"enabled":true},"enable_weight_nz_layout":true}',
+        "--compilation-config",
+        '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[1,8,24,48,60]}'
+    ]
+    request_keyword_args: dict[str, Any] = {
+        **api_keyword_args,
+    }
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False) as server:
+        client = server.get_async_client()
+        batch = await client.completions.create(
+            model=model,
+            prompt=prompts,
+            **request_keyword_args,
+        )
+        choices: list[openai.types.CompletionChoice] = batch.choices
+        assert choices[0].text, "empty response"
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)