[1/N] Refactor nightly test structure (#5479)

### What this PR does / why we need it?
This patch is a series of refactoring actions, including clarifying the
directory structure of nightly tests, refactoring the config retrieval
logic, and optimizing the workflow, etc. This is the first step:
refactoring the directory structure of nightly to make it more readable
and logical.

- vLLM version: v0.13.0
- vLLM main:
5326c89803

Signed-off-by: wangli <wangli858794774@gmail.com>

tests/e2e/nightly/single_node/models/test_deepseek_r1_0528_w8a8.py

@@ -0,0 +1,117 @@
+# 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.network_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 = [
+    "single",
+    "aclgraph",
+    "aclgraph_mlapo",
+]
+
+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": "mtp"}
+    additional_config = {"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")
+    if mode == "aclgraph_mlapo":
+        env_dict["VLLM_ASCEND_ENABLE_MLAPO"] = "1"
+    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"]:
+            return
+        # aisbench test
+        run_aisbench_cases(model,
+                           port,
+                           aisbench_cases,
+                           server_args=server_args)

tests/e2e/nightly/single_node/models/test_deepseek_r1_0528_w8a8_eplb.py

@@ -0,0 +1,115 @@
+# 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.network_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",
+]
+
+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
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+async def test_models(model: str) -> None:
+    port = get_open_port()
+    env_dict = {
+        "OMP_NUM_THREADS": "100",
+        "OMP_PROC_BIND": "false",
+        "HCCL_BUFFSIZE": "200",
+        "VLLM_ASCEND_ENABLE_MLAPO": "1",
+        "VLLM_RPC_TIMEOUT": "3600000",
+        "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000",
+        "DISABLE_L2_CACHE": "1",
+        "DYNAMIC_EPLB": "true",
+    }
+    speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
+    compilation_config = {
+        "cudagraph_capture_sizes": [24],
+        "cudagraph_mode": "FULL_DECODE_ONLY"
+    }
+    additional_config: dict[str, Any] = {
+        "enable_shared_expert_dp": False,
+        "multistream_overlap_shared_expert": False,
+        "dynamic_eplb": True,
+        "num_iterations_eplb_update": 14000,
+        "num_wait_worker_iterations": 30,
+        "init_redundancy_expert": 0,
+        "gate_eplb": False
+    }
+    server_args = [
+        "--quantization", "ascend", "--seed", "1024",
+        "--no-enable-prefix-caching", "--data-parallel-size", "4",
+        "--tensor-parallel-size", "4", "--enable-expert-parallel", "--port",
+        str(port), "--max-model-len", "40000", "--max-num-batched-tokens",
+        "4096", "--max-num-seqs", "12", "--trust-remote-code",
+        "--gpu-memory-utilization", "0.92"
+    ]
+    server_args.extend(
+        ["--speculative-config",
+         json.dumps(speculative_config)])
+    server_args.extend(
+        ["--compilation-config",
+         json.dumps(compilation_config)])
+    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)
+        # aisbench test
+        run_aisbench_cases(model,
+                           port,
+                           aisbench_cases,
+                           server_args=server_args)

tests/e2e/nightly/single_node/models/test_deepseek_v3_2_exp_w8a8.py

@@ -0,0 +1,105 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/DeepSeek-V3.2-Exp-W8A8",
+]
+
+TENSOR_PARALLELS = [8]
+DATA_PARALLELS = [2]
+FULL_GRAPH = [True, False]
+
+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": 4096,
+    "batch_size": 8,
+    "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": 16,
+    "max_out_len": 1500,
+    "batch_size": 8,
+    "request_rate": 0,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
+@pytest.mark.parametrize("dp_size", DATA_PARALLELS)
+@pytest.mark.parametrize("full_graph", FULL_GRAPH)
+async def test_models(model: str, tp_size: int, dp_size: int,
+                      full_graph: bool) -> None:
+    port = get_open_port()
+    env_dict = {"HCCL_BUFFSIZE": "1024", "VLLM_ASCEND_ENABLE_MLAPO": "0"}
+    server_args = [
+        "--no-enable-prefix-caching", "--enable-expert-parallel",
+        "--tensor-parallel-size",
+        str(tp_size), "--data-parallel-size",
+        str(dp_size), "--port",
+        str(port), "--max-model-len", "16384", "--max-num-batched-tokens",
+        "16384", "--block-size", "16", "--trust-remote-code", "--quantization",
+        "ascend", "--gpu-memory-utilization", "0.9"
+    ]
+    if full_graph:
+        server_args += [
+            "--compilation-config",
+            '{"cudagraph_capture": [16], "cudagraph_model":"FULL_DECODE_ONLY"}'
+        ]
+    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)

tests/e2e/nightly/single_node/models/test_glm4_5.py

@@ -0,0 +1,111 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "ZhipuAI/GLM-4.5",
+]
+
+TENSOR_PARALLELS = [8]
+DATA_PARALLELS = [2]
+
+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": 4096,
+    "batch_size": 8,
+    "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": 16,
+    "max_out_len": 1500,
+    "batch_size": 8,
+    "request_rate": 0,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("tp_size", TENSOR_PARALLELS)
+@pytest.mark.parametrize("dp_size", DATA_PARALLELS)
+async def test_models(
+    model: str,
+    tp_size: int,
+    dp_size: int,
+) -> None:
+    port = get_open_port()
+    env_dict = {"HCCL_BUFFSIZE": "1024"}
+    server_args = [
+        "--no-enable-prefix-caching",
+        "--enable-expert-parallel",
+        "--tensor-parallel-size",
+        str(tp_size),
+        "--data-parallel-size",
+        str(dp_size),
+        "--port",
+        str(port),
+        "--max-model-len",
+        "8192",
+        "--max-num-batched-tokens",
+        "8192",
+        "--block-size",
+        "16",
+        "--trust-remote-code",
+        "--gpu-memory-utilization",
+        "0.9",
+    ]
+    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)

tests/e2e/nightly/single_node/models/test_mtpx_deepseek_r1_0528_w8a8.py

@@ -0,0 +1,140 @@
+# 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.network_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 = ["mtp2", "mtp3"]
+
+prompts = [
+    "San Francisco is a",
+]
+
+api_keyword_args = {
+    "max_tokens": 10,
+}
+
+aisbench_gsm8k = [{
+    "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
+}]
+
+aisbench_aime = [{
+    "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": 86.67,
+    "threshold": 7
+}]
+
+
+@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": "100",
+        "OMP_PROC_BIND": "false",
+        "HCCL_BUFFSIZE": "1024",
+        "VLLM_RPC_TIMEOUT": "3600000",
+        "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000"
+    }
+    speculative_config = {"num_speculative_tokens": 2, "method": "mtp"}
+    compilation_config = {
+        "cudagraph_capture_sizes": [56],
+        "cudagraph_mode": "FULL_DECODE_ONLY"
+    }
+    server_args = [
+        "--quantization",
+        "ascend",
+        "--seed",
+        "1024",
+        "--no-enable-prefix-caching",
+        "--data-parallel-size",
+        "2",
+        "--tensor-parallel-size",
+        "8",
+        "--enable-expert-parallel",
+        "--port",
+        str(port),
+        "--max-model-len",
+        "40960",
+        "--max-num-seqs",
+        "14",
+        "--trust-remote-code",
+    ]
+    if mode == "mtp2":
+        server_args.extend(["--max-num-batched-tokens", "4096"])
+        server_args.extend(
+            ["--speculative-config",
+             json.dumps(speculative_config)])
+        server_args.extend(["--gpu-memory-utilization", "0.92"])
+        aisbench_cases = aisbench_gsm8k
+    if mode == "mtp3":
+        env_dict["HCCL_OP_EXPANSION_MODE"] = "AIV"
+        server_args.extend(["--max-num-batched-tokens", "2048"])
+        speculative_config["num_speculative_tokens"] = 3
+        server_args.extend(
+            ["--speculative-config",
+             json.dumps(speculative_config)])
+        server_args.extend(["--gpu-memory-utilization", "0.9"])
+        server_args.extend(
+            ["--compilation-config",
+             json.dumps(compilation_config)])
+        aisbench_cases = aisbench_aime
+    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)
+        # aisbench test
+        run_aisbench_cases(model,
+                           port,
+                           aisbench_cases,
+                           server_args=server_args)

tests/e2e/nightly/single_node/models/test_prefix_cache_deepseek_r1_0528_w8a8.py

@@ -0,0 +1,107 @@
+# 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
+
+import pytest
+from vllm.utils.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import get_TTFT, run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/DeepSeek-R1-0528-W8A8",
+]
+
+aisbench_warm_up = [{
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/GSM8K-in1024-bs210",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 210,
+    "max_out_len": 2,
+    "batch_size": 1000,
+    "baseline": 0,
+    "threshold": 0.97
+}]
+
+aisbench_cases0 = [{
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/prefix0-in3500-bs210",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 210,
+    "max_out_len": 1500,
+    "batch_size": 18,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+aisbench_cases75 = [{
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/prefix75-in3500-bs210",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 210,
+    "max_out_len": 1500,
+    "batch_size": 18,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+async def test_models(model: 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",
+    }
+    additional_config = {"enable_weight_nz_layout": True}
+    speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
+    server_args = [
+        "--quantization", "ascend", "--data-parallel-size", "2",
+        "--tensor-parallel-size", "8", "--enable-expert-parallel", "--port",
+        str(port), "--seed", "1024", "--max-model-len", "5200",
+        "--max-num-batched-tokens", "4096", "--max-num-seqs", "16",
+        "--trust-remote-code", "--gpu-memory-utilization", "0.9",
+        "--additional-config",
+        json.dumps(additional_config), "--speculative-config",
+        json.dumps(speculative_config)
+    ]
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False):
+        run_aisbench_cases(model, port, aisbench_warm_up)
+        result = run_aisbench_cases(model, port, aisbench_cases0)
+        TTFT0 = get_TTFT(result)
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False):
+        run_aisbench_cases(model, port, aisbench_warm_up)
+        result = run_aisbench_cases(model, port, aisbench_cases75)
+        TTFT75 = get_TTFT(result)
+    assert TTFT75 < 0.8 * TTFT0, f"The TTFT for prefix75 {TTFT75} is not less than 0.8*TTFT for prefix0 {TTFT0}."
+    print(
+        f"The TTFT for prefix75 {TTFT75} is less than 0.8*TTFT for prefix0 {TTFT0}."
+    )

tests/e2e/nightly/single_node/models/test_prefix_cache_qwen3_32b_int8.py

@@ -0,0 +1,99 @@
+# 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
+
+import pytest
+from vllm.utils.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import get_TTFT, run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/Qwen3-32B-W8A8",
+]
+
+aisbench_warm_up = [{
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/GSM8K-in1024-bs210",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 210,
+    "max_out_len": 2,
+    "batch_size": 1000,
+    "baseline": 0,
+    "threshold": 0.97
+}]
+
+aisbench_cases0 = [{
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/prefix0-in3500-bs210",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 210,
+    "max_out_len": 1500,
+    "batch_size": 48,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+aisbench_cases75 = [{
+    "case_type": "performance",
+    "dataset_path": "vllm-ascend/prefix75-in3500-bs210",
+    "request_conf": "vllm_api_stream_chat",
+    "dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_str_perf",
+    "num_prompts": 210,
+    "max_out_len": 1500,
+    "batch_size": 48,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+async def test_models(model: str) -> None:
+    port = get_open_port()
+    env_dict = {"TASK_QUEUE_ENABLE": "1", "HCCL_OP_EXPANSION_MODE": "AIV"}
+    additional_config = {"enable_weight_nz_layout": True}
+    server_args = [
+        "--quantization", "ascend", "--reasoning-parser", "qwen3",
+        "--tensor-parallel-size", "4", "--port",
+        str(port), "--max-model-len", "8192", "--max-num-batched-tokens",
+        "8192", "--max-num-seqs", "256", "--trust-remote-code",
+        "--gpu-memory-utilization", "0.9", "--additional-config",
+        json.dumps(additional_config)
+    ]
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False):
+        run_aisbench_cases(model, port, aisbench_warm_up)
+        result = run_aisbench_cases(model, port, aisbench_cases0)
+        TTFT0 = get_TTFT(result)
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False):
+        run_aisbench_cases(model, port, aisbench_warm_up)
+        result = run_aisbench_cases(model, port, aisbench_cases75)
+        TTFT75 = get_TTFT(result)
+    assert TTFT75 < 0.8 * TTFT0, f"The TTFT for prefix75 {TTFT75} is not less than 0.8*TTFT for prefix0 {TTFT0}."
+    print(
+        f"The TTFT for prefix75 {TTFT75} is less than 0.8*TTFT for prefix0 {TTFT0}."
+    )

tests/e2e/nightly/single_node/models/test_qwen2_5_vl_32b.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.network_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-32B-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": 76.22,
+    "temperature": 0,
+    "top_k": -1,
+    "top_p": 1,
+    "repetition_penalty": 1,
+    "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,
+    "temperature": 0,
+    "top_k": -1,
+    "top_p": 1,
+    "repetition_penalty": 1,
+    "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", "--mm-processor-cache-gb", "0",
+        "--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", "--compilation_config",
+        '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+    ]
+    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/single_node/models/test_qwen2_5_vl_7b.py

@@ -0,0 +1,102 @@
+# 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.network_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": 82.05,
+    "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", "--mm-processor-cache-gb", "0",
+        "--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", "--compilation_config",
+        '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+    ]
+    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/single_node/models/test_qwen3_235b_a22b_w8a8_eplb.py

@@ -0,0 +1,105 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/Qwen3-235B-A22B-W8A8",
+]
+
+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,
+    "top_k": 20,
+    "baseline": 95,
+    "threshold": 5
+}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model", MODELS)
+async def test_models(model: 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",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM1": "1"
+    }
+    additional_config: dict[str, Any] = {}
+    compilation_config = {"cudagraph_mode": "FULL_DECODE_ONLY"}
+    server_args = [
+        "--quantization", "ascend", "--async-scheduling",
+        "--data-parallel-size", "4", "--tensor-parallel-size", "4",
+        "--enable-expert-parallel", "--port",
+        str(port), "--max-model-len", "40960", "--max-num-batched-tokens",
+        "8192", "--max-num-seqs", "12", "--trust-remote-code",
+        "--gpu-memory-utilization", "0.9"
+    ]
+    env_dict["EXPERT_MAP_RECORD"] = "true"
+    env_dict["DYNAMIC_EPLB"] = "true"
+    additional_config["dynamic_eplb"] = True
+    additional_config["num_iterations_eplb_update"] = 14000
+    additional_config["num_wait_worker_iterations"] = 30
+    additional_config["init_redundancy_expert"] = 0
+    additional_config["gate_eplb"] = False
+    server_args.extend(
+        ["--compilation-config",
+         json.dumps(compilation_config)])
+    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)
+        # aisbench test
+        run_aisbench_cases(model,
+                           port,
+                           aisbench_cases,
+                           server_args=server_args)

tests/e2e/nightly/single_node/models/test_qwen3_235b_w8a8.py

@@ -0,0 +1,101 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/Qwen3-235B-A22B-W8A8",
+]
+
+MODES = ["full_graph", "piecewise"]
+
+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,
+    "top_k": 20,
+    "baseline": 95,
+    "threshold": 5
+}]
+
+
+@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",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM1": "1"
+    }
+    compilation_config = {"cudagraph_mode": "FULL_DECODE_ONLY"}
+    server_args = [
+        "--quantization", "ascend", "--async-scheduling",
+        "--data-parallel-size", "4", "--tensor-parallel-size", "4",
+        "--enable-expert-parallel", "--port",
+        str(port), "--max-model-len", "40960", "--max-num-batched-tokens",
+        "8192", "--max-num-seqs", "12", "--trust-remote-code",
+        "--gpu-memory-utilization", "0.9"
+    ]
+    if mode == "piecewise":
+        compilation_config["cudagraph_mode"] = "PIECEWISE"
+    server_args.extend(
+        ["--compilation-config",
+         json.dumps(compilation_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)
+        # aisbench test
+        run_aisbench_cases(model,
+                           port,
+                           aisbench_cases,
+                           server_args=server_args)

tests/e2e/nightly/single_node/models/test_qwen3_30b_w8a8.py

@@ -0,0 +1,92 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "vllm-ascend/Qwen3-30B-A3B-W8A8",
+]
+
+TENSOR_PARALLELS = [1]
+
+prompts = [
+    "San Francisco is a",
+]
+
+api_keyword_args = {
+    "max_tokens": 10,
+}
+
+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": 180,
+    "max_out_len": 1500,
+    "batch_size": 45,
+    "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 = {
+        "OMP_PROC_BIND": "false",
+        "OMP_NUM_THREADS": "10",
+        "HCCL_BUFFSIZE": "1024",
+        "HCCL_OP_EXPANSION_MODE": "AIV",
+        "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True"
+    }
+    server_args = [
+        "--quantization", "ascend", "--async-scheduling",
+        "--no-enable-prefix-caching", "--tensor-parallel-size",
+        str(tp_size), "--port",
+        str(port), "--max-model-len", "5600", "--max-num-batched-tokens",
+        "16384", "--max-num-seqs", "100", "--trust-remote-code",
+        "--gpu-memory-utilization", "0.9", "--compilation-config",
+        '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+    ]
+    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)

tests/e2e/nightly/single_node/models/test_qwen3_32b.py

@@ -0,0 +1,99 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "Qwen/Qwen3-32B",
+]
+
+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_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": 80,
+    "max_out_len": 1500,
+    "batch_size": 20,
+    "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",
+        "OMP_PROC_BIND": "false",
+        "HCCL_OP_EXPANSION_MODE": "AIV",
+        "PAGED_ATTENTION_MASK_LEN": "5500"
+    }
+    server_args = [
+        "--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}'
+    ]
+    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)

tests/e2e/nightly/single_node/models/test_qwen3_32b_int8.py

@@ -0,0 +1,129 @@
+# 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
+import os
+from typing import Any
+
+import openai
+import pytest
+from vllm.utils.network_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": 72,
+    "linux-aarch64-a3-4": 76,
+}
+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": 7
+}, {
+    "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
+}]
+
+
+@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",
+        "HCCL_OP_EXPANSION_MODE": "AIV",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM": "1",
+        "VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"
+    }
+    compilation_config = {
+        "cudagraph_mode":
+        "FULL_DECODE_ONLY",
+        "cudagraph_capture_sizes":
+        [1, 12, 16, 20, 24, 32, 48, 60, 64, 68, 72, 76, 80]
+    }
+    server_args = [
+        "--quantization", "ascend", "--no-enable-prefix-caching",
+        "--tensor-parallel-size",
+        str(tp_size), "--port",
+        str(port), "--max-model-len", "40960", "--max-num-batched-tokens",
+        "40960", "--block-size", "128", "--trust-remote-code",
+        "--reasoning-parser", "qwen3", "--gpu-memory-utilization", "0.9",
+        "--async-scheduling"
+    ]
+    if mode == "single":
+        server_args.append("--enforce-eager")
+    if mode == "aclgraph":
+        server_args.extend(
+            ["--compilation-config",
+             json.dumps(compilation_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 == "single":
+            return
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)

tests/e2e/nightly/single_node/models/test_qwen3_32b_int8_a3_feature_stack3.py

@@ -0,0 +1,98 @@
+# 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 pytest
+from vllm.utils.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+from tools.send_request import send_v1_chat_completions
+
+MODELS = [
+    "vllm-ascend/Qwen3-32B-W8A8",
+]
+
+TENSOR_PARALLELS = [4]
+
+prompts = [
+    "9.11 and 9.8, which is greater?",
+]
+
+api_keyword_args = {
+    "chat_template_kwargs": {
+        "enable_thinking": True
+    },
+}
+
+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_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",
+        '{"enable_weight_nz_layout":true}', "--compilation-config",
+        '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[1,8,24,48,60]}'
+    ]
+    with RemoteOpenAIServer(model,
+                            server_args,
+                            server_port=port,
+                            env_dict=env_dict,
+                            auto_port=False) as server:
+        send_v1_chat_completions(prompts[0],
+                                 model,
+                                 server,
+                                 request_args=api_keyword_args)
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)

tests/e2e/nightly/single_node/models/test_qwq_32b.py

@@ -0,0 +1,116 @@
+# 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.network_utils import get_open_port
+
+from tests.e2e.conftest import RemoteOpenAIServer
+from tools.aisbench import run_aisbench_cases
+
+MODELS = [
+    "Qwen/QwQ-32B",
+]
+
+MODES = [
+    "aclgraph",
+    "single",
+]
+
+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_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": 240,
+    "max_out_len": 1500,
+    "batch_size": 60,
+    "baseline": 1,
+    "threshold": 0.97
+}]
+
+
+@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",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM": "1",
+        "VLLM_ASCEND_ENABLE_DEBSE_OPTIMIZE": "1",
+        "VLLM_ASCEND_ENABLE_PREFETCH_MLP": "1"
+    }
+    server_args = [
+        "--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", "--compilation_config",
+        '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes": [1, 8, 24, 48, 60]}',
+        "--reasoning-parser", "deepseek_r1", "--distributed_executor_backend",
+        "mp"
+    ]
+    if mode == "single":
+        server_args.remove("--compilation_config")
+        server_args.remove(
+            '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes": [1, 8, 24, 48, 60]}'
+        )
+        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"
+        if mode == "single":
+            return
+        # aisbench test
+        run_aisbench_cases(model, port, aisbench_cases)