[CI] cleanup single/multi-card test (#5623)

1. speed up e2e light test.
2. create `2-cards` and `4-cards` folder in multicard
3. move ops to nightly
4. run test in Alphabetical Order

- vLLM version: v0.13.0
- vLLM main:
8be6432bda

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

tests/e2e/multicard/4-cards/long_sequence/test_accuracy.py

@@ -0,0 +1,212 @@
+#
+# 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.
+#
+"""
+Compare the outputs of vLLM with and without context parallel.
+
+Run `pytest tests/e2e/multicard/long_sequence/test_accuracy.py`.
+"""
+
+import pytest
+
+from tests.e2e.conftest import VllmRunner
+from tests.e2e.model_utils import check_outputs_equal
+
+MODELS = [
+    "Qwen/Qwen3-8B",
+    "vllm-ascend/DeepSeek-V2-Lite-W8A8",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("max_tokens", [10])
+def test_models_long_sequence_output_between_tp_and_cp(
+    model: str,
+    max_tokens: int,
+) -> None:
+    prompts = [
+        "The president of the United States is", "The capital of France is"
+    ]
+
+    common_kwargs = {
+        "max_model_len": 1024,
+    }
+
+    if model == "vllm-ascend/DeepSeek-V2-Lite-W8A8":
+        cp_kwargs = {
+            "tensor_parallel_size": 2,
+            "decode_context_parallel_size": 2,
+            "prefill_context_parallel_size": 2,
+            "enable_expert_parallel": True,
+            "enforce_eager": True,
+            "quantization": "ascend",
+        }
+        tp_kwargs = {
+            "tensor_parallel_size": 4,
+            "enable_expert_parallel": True,
+            "enforce_eager": True,
+            "quantization": "ascend",
+        }
+
+    else:
+        cp_kwargs = {
+            "tensor_parallel_size": 1,
+            "decode_context_parallel_size": 1,
+            "prefill_context_parallel_size": 2,
+            "compilation_config": {
+                "cudagraph_mode": "FULL_DECODE_ONLY",
+                "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
+            },
+        }
+        tp_kwargs = {
+            "tensor_parallel_size": 2,
+            "enforce_eager": True,
+        }
+
+    cp_full_kwargs = {}
+    cp_full_kwargs.update(common_kwargs)  # type: ignore
+    cp_full_kwargs.update(cp_kwargs)  # type: ignore
+
+    tp_full_kwargs = {}
+    tp_full_kwargs.update(common_kwargs)  # type: ignore
+    tp_full_kwargs.update(tp_kwargs)  # type: ignore
+    with VllmRunner(model, **cp_full_kwargs) as runner:  # type: ignore
+        vllm_context_parallel_outputs = runner.generate_greedy(
+            prompts, max_tokens)
+
+    with VllmRunner(model, **tp_full_kwargs) as runner:  # type: ignore
+        vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=vllm_eager_outputs,
+        outputs_1_lst=vllm_context_parallel_outputs,
+        name_0="vllm_eager_outputs",
+        name_1="vllm_context_parallel_outputs",
+    )
+
+
+model = "vllm-ascend/DeepSeek-V2-Lite-W8A8"
+
+
+@pytest.mark.parametrize("max_tokens", [10])
+def test_accuracy_dcp_only_graph(max_tokens: int, ) -> None:
+    prompts = [
+        "The president of the United States is", "The capital of France is"
+    ]
+    cp_kwargs = {
+        "tensor_parallel_size": 2,
+        "decode_context_parallel_size": 2,
+        "prefill_context_parallel_size": 1,
+        "enable_expert_parallel": True,
+        "compilation_config": {
+            "cudagraph_mode": "FULL_DECODE_ONLY",
+            "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
+        },
+        "quantization": "ascend",
+        "max_model_len": 1024,
+    }
+    tp_kwargs = {
+        "tensor_parallel_size": 4,
+        "enable_expert_parallel": True,
+        "enforce_eager": True,
+        "quantization": "ascend",
+        "max_model_len": 1024,
+    }
+    with VllmRunner(model, **cp_kwargs) as runner:  # type: ignore
+        vllm_context_parallel_outputs = runner.generate_greedy(
+            prompts, max_tokens)
+
+    with VllmRunner(model, **tp_kwargs) as runner:  # type: ignore
+        vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=vllm_eager_outputs,
+        outputs_1_lst=vllm_context_parallel_outputs,
+        name_0="vllm_eager_outputs",
+        name_1="vllm_dcp_only_graph_outputs",
+    )
+
+
+@pytest.mark.parametrize("max_tokens", [10])
+def test_accuracy_dcp_only_eager(max_tokens: int, ) -> None:
+    prompts = [
+        "The president of the United States is", "The capital of France is"
+    ]
+    cp_kwargs = {
+        "tensor_parallel_size": 2,
+        "decode_context_parallel_size": 2,
+        "prefill_context_parallel_size": 1,
+        "enable_expert_parallel": True,
+        "enforce_eager": True,
+        "quantization": "ascend",
+        "max_model_len": 1024,
+    }
+    tp_kwargs = {
+        "tensor_parallel_size": 4,
+        "enable_expert_parallel": True,
+        "enforce_eager": True,
+        "quantization": "ascend",
+        "max_model_len": 1024,
+    }
+    with VllmRunner(model, **cp_kwargs) as runner:  # type: ignore
+        vllm_context_parallel_outputs = runner.generate_greedy(
+            prompts, max_tokens)
+
+    with VllmRunner(model, **tp_kwargs) as runner:  # type: ignore
+        vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=vllm_eager_outputs,
+        outputs_1_lst=vllm_context_parallel_outputs,
+        name_0="vllm_eager_outputs",
+        name_1="vllm_dcp_only_eager_outputs",
+    )
+
+
+@pytest.mark.parametrize("max_tokens", [10])
+def test_accuracy_pcp_only(max_tokens: int, ) -> None:
+    prompts = [
+        "The president of the United States is", "The capital of France is"
+    ]
+    cp_kwargs = {
+        "tensor_parallel_size": 2,
+        "decode_context_parallel_size": 1,
+        "prefill_context_parallel_size": 2,
+        "enable_expert_parallel": True,
+        "enforce_eager": True,
+        "quantization": "ascend",
+        "max_model_len": 1024,
+    }
+    tp_kwargs = {
+        "tensor_parallel_size": 4,
+        "enable_expert_parallel": True,
+        "enforce_eager": True,
+        "quantization": "ascend",
+        "max_model_len": 1024,
+    }
+    with VllmRunner(model, **cp_kwargs) as runner:  # type: ignore
+        vllm_context_parallel_outputs = runner.generate_greedy(
+            prompts, max_tokens)
+
+    with VllmRunner(model, **tp_kwargs) as runner:  # type: ignore
+        vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=vllm_eager_outputs,
+        outputs_1_lst=vllm_context_parallel_outputs,
+        name_0="vllm_eager_outputs",
+        name_1="vllm_pcp_only_outputs",
+    )

tests/e2e/multicard/4-cards/long_sequence/test_basic.py

@@ -0,0 +1,248 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+import os
+
+from vllm import SamplingParams
+
+from tests.e2e.conftest import VllmRunner
+
+os.environ["HCCL_BUFFSIZE"] = "768"
+
+
+def test_models_pcp_dcp_basic():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=True,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=2,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+    model = "vllm-ascend/Qwen3-30B-A3B-W8A8"
+    with VllmRunner(
+            model,
+            enforce_eager=True,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            prefill_context_parallel_size=2,
+            decode_context_parallel_size=1,
+            enable_expert_parallel=True,
+            block_size=128,
+            quantization="ascend",
+    ) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_models_pcp_dcp_full_graph():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=2,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128,
+                    compilation_config={
+                        "cudagraph_mode": "FULL_DECODE_ONLY",
+                        "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
+                    }) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+    model = "vllm-ascend/Qwen3-30B-A3B-W8A8"
+    with VllmRunner(model,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=1,
+                    enable_expert_parallel=True,
+                    block_size=128,
+                    quantization="ascend",
+                    compilation_config={
+                        "cudagraph_mode": "FULL_DECODE_ONLY",
+                        "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
+                    }) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_models_pcp_dcp_piece_wise():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=2,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    cudagraph_capture_sizes=[1, 2, 4, 8],
+                    block_size=128) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+    model = "vllm-ascend/Qwen3-30B-A3B-W8A8"
+    with VllmRunner(model,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=1,
+                    enable_expert_parallel=True,
+                    cudagraph_capture_sizes=[1, 2, 4, 8],
+                    block_size=128,
+                    quantization="ascend") as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_pcp_basic():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=True,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=1,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_pcp_full_graph():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=False,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=1,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128,
+                    compilation_config={
+                        "cudagraph_mode": "FULL_DECODE_ONLY",
+                        "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
+                    }) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_pcp_piece_wise():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=False,
+                    max_model_len=1024,
+                    tensor_parallel_size=2,
+                    prefill_context_parallel_size=2,
+                    decode_context_parallel_size=1,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_dcp_basic():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=True,
+                    max_model_len=1024,
+                    tensor_parallel_size=4,
+                    prefill_context_parallel_size=1,
+                    decode_context_parallel_size=2,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_dcp_full_graph():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=False,
+                    max_model_len=1024,
+                    tensor_parallel_size=4,
+                    prefill_context_parallel_size=1,
+                    decode_context_parallel_size=2,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128,
+                    compilation_config={
+                        "cudagraph_mode": "FULL_DECODE_ONLY",
+                        "cudagraph_capture_sizes": [4, 8, 24, 48, 60]
+                    }) as runner:
+        runner.model.generate(prompts, sampling_params)
+
+
+def test_dcp_piece_wise():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "deepseek-ai/DeepSeek-V2-Lite-Chat"
+    sampling_params = SamplingParams(max_tokens=32, temperature=0.0)
+    with VllmRunner(model,
+                    enforce_eager=False,
+                    max_model_len=1024,
+                    tensor_parallel_size=4,
+                    prefill_context_parallel_size=1,
+                    decode_context_parallel_size=2,
+                    max_num_batched_tokens=1024,
+                    enable_expert_parallel=True,
+                    block_size=128) as runner:
+        runner.model.generate(prompts, sampling_params)

tests/e2e/multicard/4-cards/long_sequence/test_chunked_prefill.py

@@ -0,0 +1,73 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+import os
+import random
+import string
+from unittest.mock import patch
+
+from vllm import SamplingParams
+
+from tests.e2e.conftest import VllmRunner
+
+
+def generate_prompts(input_len, batchsize):
+    prompts = [
+        " ".join([
+            f"{random.choice(string.ascii_letters)}" for _ in range(input_len)
+        ]) for _ in range(batchsize)
+    ]
+    return prompts
+
+
+@patch.dict(
+    os.environ, {
+        "HCCL_BUFFSIZE": "768",
+        "VLLM_ASCEND_ENABLE_FLASHCOMM1": "1",
+        "VLLM_ALLOW_LONG_MAX_MODEL_LEN": "1"
+    })
+def test_models_chunked_prefill_mixed_length_prompts_including_1_token():
+    TEST_ROPE_PARAMETERS = {
+        "rope_theta": 1000000,
+        "rope_type": "yarn",
+        "factor": 4,
+        "original_max_position_embeddings": 32768
+    }
+    prompts = [
+        generate_prompts(128 * 1024, 1)[0],
+        generate_prompts(1, 1)[0],
+        generate_prompts(9104, 1)[0],
+    ]
+    sampling_params = SamplingParams(max_tokens=1, temperature=0.0)
+
+    model = "vllm-ascend/Qwen3-30B-A3B-W8A8"
+    with VllmRunner(
+            model,
+            enforce_eager=True,
+            max_num_seqs=2,
+            max_num_batched_tokens=131000,
+            max_model_len=132000,
+            tensor_parallel_size=2,
+            prefill_context_parallel_size=2,
+            decode_context_parallel_size=1,
+            enable_expert_parallel=True,
+            block_size=128,
+            quantization="ascend",
+            hf_overrides={"rope_parameters": TEST_ROPE_PARAMETERS},
+    ) as runner:
+        runner.model.generate(prompts, sampling_params)

tests/e2e/multicard/4-cards/long_sequence/test_mtp.py

@@ -0,0 +1,152 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+
+import os
+
+from tests.e2e.conftest import VllmRunner
+
+os.environ["HCCL_BUFFSIZE"] = "512"
+
+
+def test_pcp_dcp_mtp1_eager():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "wemaster/deepseek_mtp_main_random_bf16"
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            prefill_context_parallel_size=2,
+            decode_context_parallel_size=2,
+            max_num_batched_tokens=1024,
+            enable_expert_parallel=True,
+            block_size=128,
+            speculative_config={
+                "num_speculative_tokens": 1,
+                "method": "deepseek_mtp",
+            },
+            enforce_eager=True,
+    ) as runner:
+        runner.generate_greedy(prompts, 32)
+
+
+def test_pcp_dcp_mtp3_eager():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "wemaster/deepseek_mtp_main_random_bf16"
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            prefill_context_parallel_size=2,
+            decode_context_parallel_size=2,
+            max_num_batched_tokens=1024,
+            enable_expert_parallel=True,
+            block_size=128,
+            speculative_config={
+                "num_speculative_tokens": 3,
+                "method": "deepseek_mtp",
+            },
+            enforce_eager=True,
+    ) as runner:
+        runner.generate_greedy(prompts, 32)
+
+
+def test_pcp_dcp_mtp3_piecewise_graph():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "wemaster/deepseek_mtp_main_random_bf16"
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            prefill_context_parallel_size=2,
+            decode_context_parallel_size=2,
+            max_num_batched_tokens=1024,
+            enable_expert_parallel=True,
+            block_size=128,
+            speculative_config={
+                "num_speculative_tokens": 3,
+                "method": "deepseek_mtp",
+            },
+            compilation_config={
+                "cudagraph_mode": "PIECEWISE",
+                "cudagraph_capture_sizes": [4, 8, 16],
+            },
+    ) as runner:
+        runner.generate_greedy(prompts, 32)
+
+
+def test_pcp_dcp_mtp3_full_graph():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "wemaster/deepseek_mtp_main_random_bf16"
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            prefill_context_parallel_size=2,
+            decode_context_parallel_size=2,
+            max_num_batched_tokens=1024,
+            enable_expert_parallel=True,
+            block_size=128,
+            speculative_config={
+                "num_speculative_tokens": 3,
+                "method": "deepseek_mtp",
+            },
+            compilation_config={
+                "cudagraph_mode": "FULL_DECODE_ONLY",
+                "cudagraph_capture_sizes": [4, 8, 16],
+            },
+    ) as runner:
+        runner.generate_greedy(prompts, 32)
+
+
+def test_dcp_mtp3_full_graph():
+    prompts = [
+        "The capital of France is", "Hello, my name is Tom, I am",
+        "The president of United States is", "AI future is"
+    ]
+    model = "wemaster/deepseek_mtp_main_random_bf16"
+    with VllmRunner(
+            model,
+            max_model_len=1024,
+            tensor_parallel_size=2,
+            decode_context_parallel_size=2,
+            max_num_batched_tokens=1024,
+            enable_expert_parallel=True,
+            block_size=128,
+            speculative_config={
+                "num_speculative_tokens": 3,
+                "method": "deepseek_mtp",
+            },
+            compilation_config={
+                "cudagraph_mode": "FULL_DECODE_ONLY",
+                "cudagraph_capture_sizes": [4, 8, 16],
+            },
+    ) as runner:
+        runner.generate_greedy(prompts, 32)

tests/e2e/multicard/4-cards/spec_decode/test_mtp_qwen3_next.py

@@ -0,0 +1,154 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+"""Compare the short outputs of HF and vLLM when using greedy sampling.
+
+Run `pytest tests/e2e/multicard/spec_decode/test_mtp_qwen3_next.py`.
+"""
+
+import os
+
+import pytest
+from vllm.config import CompilationConfig
+from vllm.v1.metrics.reader import Counter, Vector
+
+from tests.e2e.conftest import VllmRunner, cleanup_dist_env_and_memory
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+MODELS = ["Qwen/Qwen3-Next-80B-A3B-Instruct"]
+
+
+@pytest.mark.parametrize("model_name", MODELS)
+def test_qwen3_next_mtp_acceptance_tp4(model_name):
+    golden = [0.85, 0.46, 0.19]
+
+    example_prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+
+    max_tokens = 1024
+
+    with VllmRunner(model_name,
+                    tensor_parallel_size=4,
+                    max_model_len=4096,
+                    gpu_memory_utilization=0.8,
+                    distributed_executor_backend="mp",
+                    disable_log_stats=False,
+                    speculative_config={
+                        "cudagraph_mode": "FULL_DECODE_ONLY",
+                        "method": "qwen3_next_mtp",
+                        "num_speculative_tokens": 3,
+                    },
+                    compilation_config=CompilationConfig(
+                        cudagraph_capture_sizes=[20])) as spec_vllm_model:
+        _ = spec_vllm_model.generate_greedy(example_prompts, max_tokens)
+        metrics = spec_vllm_model.model.get_metrics()
+
+    num_drafts = 0
+    num_accepted_tokens_per_pos = [0] * 3
+    for metric in metrics:
+        if metric.name == "vllm:spec_decode_num_drafts":
+            assert isinstance(metric, Counter)
+            num_drafts += metric.value
+        elif metric.name == "vllm:spec_decode_num_accepted_tokens_per_pos":
+            assert isinstance(metric, Vector)
+            for pos in range(len(metric.values)):
+                num_accepted_tokens_per_pos[pos] += metric.values[pos]
+
+    acceptance_per_pos = [
+        num_accepted_tokens / num_drafts
+        for num_accepted_tokens in num_accepted_tokens_per_pos
+    ]
+
+    match = all(abs(a - b) < 0.05 for a, b in zip(acceptance_per_pos, golden))
+    if not match:
+        print(f"acceptance_per_pos: {acceptance_per_pos}")
+        print(f"golden: {golden}")
+
+    assert match
+    cleanup_dist_env_and_memory()
+
+
+# FIXME: When applying `FULL_DECODE_ONLY` in this e2e, ci will fail.
+# The failure can not be reproduced locally.
+@pytest.mark.parametrize("model_name", MODELS)
+@pytest.mark.parametrize("num_speculative_tokens", [1])
+@pytest.mark.parametrize("disable_padded_drafter_batch", [True, False])
+def test_qwen3_next_mtp_correctness_tp4(model_name: str,
+                                        num_speculative_tokens: int,
+                                        disable_padded_drafter_batch: bool):
+    example_prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+
+    max_tokens = 20
+    '''
+    Compare the outputs of a original LLM and a speculative LLM
+    should be the same when using mtp speculative decoding.
+    '''
+    with VllmRunner(model_name,
+                    tensor_parallel_size=4,
+                    max_model_len=4096,
+                    gpu_memory_utilization=0.8,
+                    distributed_executor_backend="mp",
+                    speculative_config={
+                        "method":
+                        "mtp",
+                        "num_speculative_tokens":
+                        num_speculative_tokens,
+                        "disable_padded_drafter_batch":
+                        disable_padded_drafter_batch,
+                    },
+                    compilation_config=CompilationConfig(
+                        cudagraph_capture_sizes=[20])) as spec_llm:
+        spec_outputs = spec_llm.generate_greedy(example_prompts, max_tokens)
+    del spec_llm
+
+    with VllmRunner(model_name,
+                    tensor_parallel_size=4,
+                    max_model_len=4096,
+                    gpu_memory_utilization=0.8,
+                    distributed_executor_backend="mp",
+                    compilation_config=CompilationConfig(
+                        cudagraph_capture_sizes=[20])) as ref_llm:
+        ref_outputs = ref_llm.generate_greedy(example_prompts, max_tokens)
+    del ref_llm
+
+    matches = 0
+    misses = 0
+    for ref_output, spec_output in zip(ref_outputs, spec_outputs):
+        ref_token_ids = ref_output[0]
+        spec_token_ids = spec_output[0]
+        if ref_token_ids == spec_token_ids[:len(ref_token_ids)]:
+            matches += 1
+        else:
+            misses += 1
+            print(f"ref_output: {ref_output[1]}")
+            print(f"spec_output: {spec_output[1]}")
+
+    # Heuristic: expect at least 66% of the prompts to match exactly
+    # Upon failure, inspect the outputs to check for inaccuracy.
+    assert matches > int(0.66 * len(ref_outputs))
+    cleanup_dist_env_and_memory()

tests/e2e/multicard/4-cards/test_data_parallel_tp2.py

@@ -0,0 +1,48 @@
+import os
+import subprocess
+import sys
+from unittest.mock import patch
+
+import pytest
+
+MODELS = ["Qwen/Qwen3-30B-A3B"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("max_tokens", [32])
+@patch.dict(os.environ, {"ASCEND_RT_VISIBLE_DEVICES": "0,1,2,3"})
+def test_qwen3_inference_dp2_tp2(model, max_tokens):
+    script = "examples/offline_data_parallel.py"
+
+    env = os.environ.copy()
+
+    cmd = [
+        sys.executable,
+        script,
+        "--model",
+        model,
+        "--dp-size",
+        "2",
+        "--tp-size",
+        "2",
+        "--node-size",
+        "1",
+        "--node-rank",
+        "0",
+        "--trust-remote-code",
+    ]
+
+    print(f"Running subprocess: {' '.join(cmd)}")
+    proc = subprocess.run(cmd,
+                          env=env,
+                          stdout=subprocess.PIPE,
+                          stderr=subprocess.STDOUT,
+                          timeout=600)
+    output = proc.stdout.decode(errors='ignore')
+
+    print(output)
+
+    assert "DP rank 0 needs to process" in output
+    assert "DP rank 1 needs to process" in output
+    assert "Generated text:" in output
+    assert proc.returncode == 0

tests/e2e/multicard/4-cards/test_kimi_k2.py

@@ -0,0 +1,44 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+import os
+
+from tests.e2e.conftest import VllmRunner
+
+os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+
+def test_kimi_k2_thinking_w4a16_tp4():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+
+    with VllmRunner(
+            "vllm-ascend/Kimi-K2-Thinking-Pruning",
+            max_model_len=8192,
+            dtype="auto",
+            tensor_parallel_size=4,
+            enable_expert_parallel=True,
+            compilation_config={
+                "cudagraph_mode": "FULL_DECODE_ONLY",
+                "cudagraph_capture_sizes": [1],
+            },
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)

tests/e2e/multicard/4-cards/test_qwen3_next.py

@@ -0,0 +1,77 @@
+#
+# 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.
+# Adapted from vllm/tests/basic_correctness/test_basic_correctness.py
+#
+import os
+from unittest.mock import patch
+
+from modelscope import snapshot_download  # type: ignore
+
+from tests.e2e.conftest import VllmRunner
+
+
+def test_qwen3_next_distributed_mp_tp4():
+    example_prompts = [
+        "Hello, my name is",
+    ] * 4
+    max_tokens = 5
+    with VllmRunner("Qwen/Qwen3-Next-80B-A3B-Instruct",
+                    tensor_parallel_size=4,
+                    cudagraph_capture_sizes=[1, 2, 4, 8],
+                    max_model_len=4096,
+                    gpu_memory_utilization=0.8,
+                    distributed_executor_backend="mp") as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+        del vllm_model
+
+
+def test_qwen3_next_distributed_mp_full_decode_only_tp4():
+    example_prompts = [
+        "Hello, my name is",
+    ] * 4
+    max_tokens = 5
+    with VllmRunner("Qwen/Qwen3-Next-80B-A3B-Instruct",
+                    tensor_parallel_size=4,
+                    max_model_len=4096,
+                    gpu_memory_utilization=0.8,
+                    distributed_executor_backend="mp",
+                    compilation_config={
+                        "cudagraph_mode": "FULL_DECODE_ONLY",
+                        "cudagraph_capture_sizes": [1, 8, 24, 48, 60]
+                    }) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+        del vllm_model
+
+
+# TODO: will conduct accuracy verification after the subsequent version becomes stable
+@patch.dict(os.environ, {"HCCL_BUFFSIZE": "1024"})
+def test_qwen3_next_w8a8dynamic_distributed_tp4_ep():
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+    with VllmRunner(
+            snapshot_download("vllm-ascend/Qwen3-Next-80B-A3B-Instruct-W8A8"),
+            max_model_len=4096,
+            tensor_parallel_size=4,
+            gpu_memory_utilization=0.4,
+            max_num_seqs=1,
+            enable_expert_parallel=True,
+            cudagraph_capture_sizes=[1, 2, 4, 8],
+            quantization="ascend",
+    ) as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)