[TEST] Add Qwen3-32b-w8a8 acc/perf A2/A3 test (#3541)

### What this PR does / why we need it?
This PR Qwen3-32b-w8a8 acc/perf 8 cases on A2 and A3, we need test them
daily.

### 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: jiangyunfan1 <jiangyunfan1@h-partners.com>
Signed-off-by: wangli <wangli858794774@gmail.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Signed-off-by: root <root@hostname-2pbfv.foreman.pxe>
Co-authored-by: wangli <wangli858794774@gmail.com>
Co-authored-by: Yikun Jiang <yikunkero@gmail.com>

.github/workflows/_e2e_nightly.yaml

@@ -109,6 +109,7 @@ jobs:
         env:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
           VLLM_USE_MODELSCOPE: True
+          VLLM_CI_RUNNER: ${{ inputs.runner }}
         run: |
           # TODO: enable more tests
           pytest -sv ${{ inputs.tests }}

.github/workflows/vllm_ascend_test_nightly.yaml

@@ -41,7 +41,7 @@ defaults:
 # and ignore the lint / 1 card / 4 cards test type
 concurrency:
   group: ascend-nightly-${{ github.ref }}
-  cancel-in-progress: true
+  #cancel-in-progress: true
 
 jobs:
   qwen3-32b:
@@ -56,3 +56,22 @@ jobs:
       vllm: v0.11.0
       runner: ${{ matrix.os }}
       tests: tests/e2e/nightly/models/test_qwen3_32b.py
+  qwen3-32b-in8-a3:
+    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/models/test_qwen3_32b_int8.py
+  qwen3-32b-in8-a2:
+    strategy:
+      matrix:
+        os: [linux-aarch64-a2-4]
+    uses: ./.github/workflows/_e2e_nightly.yaml
+    with:
+      vllm: v0.11.0
+      runner: ${{ matrix.os }}
+      tests: tests/e2e/nightly/models/test_qwen3_32b_int8.py

tests/e2e/nightly/models/test_qwen2_5_vl_7b.py

@@ -0,0 +1,110 @@
+# 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
+from tools.send_mm_request import send_image_request
+
+MODELS = [
+    "Qwen/Qwen2.5-VL-7B-Instruct",
+]
+
+TENSOR_PARALLELS = [4]
+
+prompts = [
+    "San Francisco is a",
+]
+
+api_keyword_args = {
+    "max_tokens": 10,
+}
+
+aisbench_cases = [{
+    "case_type": "accuracy",
+    "dataset_path": "vllm-ascend/textvqa-lite",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "textvqa/textvqa_gen_base64",
+    "max_out_len": 2048,
+    "batch_size": 128,
+    "baseline": 81,
+    "threshold": 5
+}, {
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/textvqa-perf-1080p",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "textvqa/textvqa_gen_base64",
+    "num_prompts": 512,
+    "max_out_len": 256,
+    "batch_size": 128,
+    "request_rate": 0,
+    "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 = {
+        "TASK_QUEUE_ENABLE": "1",
+        "VLLM_ASCEND_ENABLE_NZ": "0",
+        "HCCL_OP_EXPANSION_MODE": "AIV"
+    }
+    server_args = [
+        "--no-enable-prefix-caching",
+        "--disable-mm-preprocessor-cache",
+        "--tensor-parallel-size",
+        str(tp_size),
+        "--port",
+        str(port),
+        "--max-model-len",
+        "30000",
+        "--max-num-batched-tokens",
+        "40000",
+        "--max-num-seqs",
+        "400",
+        "--trust-remote-code",
+        "--gpu-memory-utilization",
+        "0.8",
+    ]
+    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"
+        print(choices)
+        send_image_request(model, server)
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)

tests/e2e/nightly/models/test_qwen3_32b_int8.py

@@ -0,0 +1,118 @@
+# 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.
+#
+import os
+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",
+]
+
+MODES = [
+    "aclgraph",
+    "single",
+]
+
+TENSOR_PARALLELS = [4]
+
+prompts = [
+    "San Francisco is a",
+]
+
+api_keyword_args = {
+    "max_tokens": 10,
+}
+
+batch_size_dict = {
+    "linux-aarch64-a2-4": 44,
+    "linux-aarch64-a3-4": 46,
+}
+VLLM_CI_RUNNER = os.getenv("VLLM_CI_RUNNER", "linux-aarch64-a2-4")
+performance_batch_size = batch_size_dict.get(VLLM_CI_RUNNER, 1)
+
+aisbench_cases = [{
+    "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": 4 * performance_batch_size,
+    "max_out_len": 1500,
+    "batch_size": performance_batch_size,
+    "baseline": 1,
+    "threshold": 0.97
+}, {
+    "case_type": "accuracy",
+    "dataset_path": "vllm-ascend/aime2024",
+    "request_conf": "vllm_api_general_chat",
+    "dataset_conf": "aime2024/aime2024_gen_0_shot_chat_prompt",
+    "max_out_len": 32768,
+    "batch_size": 32,
+    "baseline": 83.33,
+    "threshold": 17
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("mode", MODES)
+@pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
+async def test_models(model: str, mode: str, tp_size: int) -> None:
+    port = get_open_port()
+    env_dict = {
+        "TASK_QUEUE_ENABLE": "1",
+        "OMP_PROC_BIND": "false",
+        "HCCL_OP_EXPANSION_MODE": "AIV",
+        "PAGED_ATTENTION_MASK_LEN": "5500"
+    }
+    server_args = [
+        "--quantization", "ascend", "--no-enable-prefix-caching",
+        "--tensor-parallel-size",
+        str(tp_size), "--port",
+        str(port), "--max-model-len", "36864", "--max-num-batched-tokens",
+        "36864", "--block-size", "128", "--trust-remote-code",
+        "--gpu-memory-utilization", "0.9", "--additional-config",
+        '{"enable_weight_nz_layout":true}'
+    ]
+    if mode == "single":
+        server_args.append("--enforce-eager")
+    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"
+        print(choices)
+        if mode == "single":
+            return
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)

tools/aisbench.py

@@ -101,6 +101,9 @@ class AisbenchRunner:
         if self.task_type == "performance":
             conf_path = os.path.join(DATASET_CONF_DIR,
                                      f'{self.dataset_conf}.py')
+            if self.dataset_conf.startswith("textvqa"):
+                self.dataset_path = os.path.join(self.dataset_path,
+                                                 "textvqa_val.jsonl")
             with open(conf_path, 'r', encoding='utf-8') as f:
                 content = f.read()
             content = re.sub(r'path=.*', f'path="{self.dataset_path}",',
@@ -180,9 +183,13 @@ class AisbenchRunner:
     def _get_result_performance(self):
         result_dir = re.search(r'Performance Result files locate in (.*)',
                                self.result_line).group(1)[:-1]
-        result_csv_file = os.path.join(result_dir, "gsm8kdataset.csv")
+        dataset_type = self.dataset_conf.split('/')[0]
-        result_json_file = os.path.join(result_dir, "gsm8kdataset.json")
+        result_csv_file = os.path.join(result_dir,
+                                       f"{dataset_type}dataset.csv")
+        result_json_file = os.path.join(result_dir,
+                                        f"{dataset_type}dataset.json")
         self.result_csv = pd.read_csv(result_csv_file)
+        print("Getting performance results from file: ", result_json_file)
         with open(result_json_file, 'r', encoding='utf-8') as f:
             self.result_json = json.load(f)
 

tools/send_mm_request.py

@@ -0,0 +1,49 @@
+import base64
+import os
+
+import requests
+from modelscope import snapshot_download  # type: ignore
+
+mm_dir = snapshot_download("vllm-ascend/mm_request", repo_type='dataset')
+image_path = os.path.join(mm_dir, "test_mm2.jpg")
+with open(image_path, 'rb') as image_file:
+    image_data = base64.b64encode(image_file.read()).decode('utf-8')
+
+data = {
+    "messages": [{
+        "role":
+        "user",
+        "content": [{
+            "type": "text",
+            "text": "What is the content of this image?"
+        }, {
+            "type": "image_url",
+            "image_url": {
+                "url": f"data:image/jpeg;base64,{image_data}"
+            }
+        }]
+    }],
+    "eos_token_id": [1, 106],
+    "pad_token_id":
+    0,
+    "top_k":
+    64,
+    "top_p":
+    0.95,
+    "max_tokens":
+    8192,
+    "stream":
+    False
+}
+
+headers = {'Accept': 'application/json', 'Content-Type': 'application/json'}
+
+
+def send_image_request(model, server):
+    data["model"] = model
+    url = server.url_for("v1", "chat", "completions")
+    response = requests.post(url, headers=headers, json=data)
+    print("Status Code:", response.status_code)
+    response_json = response.json()
+    print("Response:", response_json)
+    assert response_json["choices"][0]["message"]["content"], "empty response"