[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)