[TEST]Add initial multi modal cases for nightly test and deepseek-r1 tests (#3631)

### What this PR does / why we need it?
This PR adds the initial multi modal model for nightly test, including 3
cases for Qwen2.5-vl-7b acc/perf test on A3, we need test them daily. It
also inclues 8 cases for deepseek-r1-0528-w8a8 func, acc and perf tests
### 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>

.github/workflows/vllm_ascend_test_nightly.yaml

@@ -26,7 +26,7 @@ on:
     branches:
       - 'main'
       - '*-dev'
-    types: [labeled]
+    types: [labeled,opened,synchronize]
 
 # Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
 # declared as "shell: bash -el {0}" on steps that need to be properly activated.
@@ -80,10 +80,7 @@ jobs:
     if: contains(github.event.pull_request.labels.*.name, 'run-nightly')
     strategy:
       matrix:
-        # should add A3 chip runner when available
         os: [ linux-aarch64-a3-16 ]
-    # Note (yikun): If CI resource are limited we can split job into two chain jobs
-    # only trigger e2e test after lint passed and the change is e2e related with pull request.
     uses: ./.github/workflows/_e2e_nightly.yaml
     with:
       vllm: v0.11.0
@@ -94,15 +91,32 @@ jobs:
     if: contains(github.event.pull_request.labels.*.name, 'run-nightly')
     strategy:
       matrix:
-        # should add A3 chip runner when available
         os: [ linux-aarch64-a3-16 ]
-    # Note (yikun): If CI resource are limited we can split job into two chain jobs
-    # only trigger e2e test after lint passed and the change is e2e related with pull request.
     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_deepseek_r1_w8a8_eplb.py
-
-
+  qwen2-5-vl-7b:
+    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/models/test_qwen2_5_vl_7b.py
+  deepseek-r1-0528-w8a8:
+    if: contains(github.event.pull_request.labels.*.name, 'run-nightly')
+    strategy:
+      matrix:
+        os: [ linux-aarch64-a3-16 ]
+    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_deepseek_r1_0528_w8a8.py

tests/e2e/nightly/models/test_deepseek_r1_0528_w8a8.py

@@ -0,0 +1,136 @@
+# 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 json
+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/DeepSeek-R1-0528-W8A8",
+]
+
+MODES = [
+    "torchair",
+    "single",
+    "aclgraph",
+    "no_chunkprefill",
+]
+
+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_cot_chat_prompt",
+    "max_out_len": 32768,
+    "batch_size": 32,
+    "baseline": 95,
+    "threshold": 5
+}, {
+    "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": 400,
+    "max_out_len": 1500,
+    "batch_size": 1000,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("mode", MODES)
+async def test_models(model: str, mode: str) -> None:
+    port = get_open_port()
+    env_dict = {
+        "OMP_NUM_THREADS": "10",
+        "OMP_PROC_BIND": "false",
+        "HCCL_BUFFSIZE": "1024",
+        "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True"
+    }
+    speculative_config = {
+        "num_speculative_tokens": 1,
+        "method": "deepseek_mtp"
+    }
+    additional_config = {
+        "ascend_scheduler_config": {
+            "enabled": False
+        },
+        "torchair_graph_config": {
+            "enabled": True,
+            "enable_multistream_moe": False,
+            "enable_multistream_mla": True,
+            "graph_batch_sizes": [16],
+            "use_cached_graph": True
+        },
+        "chunked_prefill_for_mla": True,
+        "enable_weight_nz_layout": True
+    }
+    server_args = [
+        "--quantization", "ascend", "--data-parallel-size", "2",
+        "--tensor-parallel-size", "8", "--enable-expert-parallel", "--port",
+        str(port), "--seed", "1024", "--max-model-len", "36864",
+        "--max-num-batched-tokens", "4096", "--max-num-seqs", "16",
+        "--trust-remote-code", "--gpu-memory-utilization", "0.9",
+        "--speculative-config",
+        json.dumps(speculative_config)
+    ]
+    if mode == "single":
+        server_args.append("--enforce-eager")
+        additional_config["torchair_graph_config"] = {"enabled": False}
+    if mode == "aclgraph":
+        additional_config["torchair_graph_config"] = {"enabled": False}
+    if mode == "no_chunkprefill":
+        additional_config["ascend_scheduler_config"] = {"enabled": True}
+        i = server_args.index("--max-num-batched-tokens") + 1
+        server_args[i] = "36864"
+    server_args.extend(["--additional-config", json.dumps(additional_config)])
+    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 in ["single", "no_chunkprefill"]:
+            return
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)

tests/e2e/nightly/models/test_qwen3_32b_int8.py

@@ -51,6 +51,15 @@ 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": "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
+}, {
     "case_type": "performance",
     "dataset_path": "vllm-ascend/GSM8K-in3500-bs400",
     "request_conf": "vllm_api_stream_chat",
@@ -60,15 +69,6 @@ aisbench_cases = [{
     "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
 }]
 
 

tools/aisbench.py

@@ -43,13 +43,12 @@ class AisbenchRunner:
         if self.task_type == "accuracy":
             aisbench_cmd = [
                 'ais_bench', '--models', f'{self.request_conf}_custom',
-                '--datasets', f'{dataset_conf}', '--debug'
+                '--datasets', f'{dataset_conf}'
             ]
         if self.task_type == "performance":
             aisbench_cmd = [
                 'ais_bench', '--models', f'{self.request_conf}_custom',
-                '--datasets', f'{dataset_conf}_custom', '--debug', '--mode',
-                'perf'
+                '--datasets', f'{dataset_conf}_custom', '--mode', 'perf'
             ]
             if self.num_prompts:
                 aisbench_cmd.extend(['--num-prompts', str(self.num_prompts)])
@@ -64,9 +63,11 @@ class AisbenchRunner:
                  port: int,
                  aisbench_config: dict,
                  verify=True):
-        self.result_line = None
         self.dataset_path = snapshot_download(aisbench_config["dataset_path"],
                                               repo_type='dataset')
+        self.model = model
+        self.model_path = snapshot_download(model)
+        self.port = port
         self.task_type = aisbench_config["case_type"]
         self.request_conf = aisbench_config["request_conf"]
         self.dataset_conf = aisbench_config.get("dataset_conf")
@@ -74,10 +75,13 @@ class AisbenchRunner:
         self.max_out_len = aisbench_config["max_out_len"]
         self.batch_size = aisbench_config["batch_size"]
         self.request_rate = aisbench_config.get("request_rate", 0)
-        self.model = model
-        self.model_path = snapshot_download(model)
-        self.port = port
+        self.temperature = aisbench_config.get("temperature")
+        self.top_k = aisbench_config.get("top_k")
+        self.top_p = aisbench_config.get("top_p")
+        self.seed = aisbench_config.get("seed")
+        self.repetition_penalty = aisbench_config.get("repetition_penalty")
         self.exp_folder = None
+        self.result_line = None
         self._init_dataset_conf()
         self._init_request_conf()
         self._run_aisbench_task()
@@ -138,6 +142,19 @@ class AisbenchRunner:
             content = re.sub(
                 r"temperature.*",
                 "temperature = 0.6,\n            ignore_eos = False,", content)
+        if self.temperature:
+            content = re.sub(r"temperature.*",
+                             f"temperature = {self.temperature}", content)
+        if self.top_p:
+            content = re.sub(r"#?top_p.*", f"top_p = {self.top_p}", content)
+        if self.top_k:
+            content = re.sub(r"#top_k.*", f"top_k = {self.top_k}", content)
+        if self.seed:
+            content = re.sub(r"#seed.*", f"seed = {self.seed}", content)
+        if self.repetition_penalty:
+            content = re.sub(
+                r"#repetition_penalty.*",
+                f"repetition_penalty = {self.repetition_penalty}", content)
         conf_path_new = os.path.join(REQUEST_CONF_DIR,
                                      f'{self.request_conf}_custom.py')
         with open(conf_path_new, 'w', encoding='utf-8') as f: