vllm-ascend support chunked prefill (#1172)

### What this PR does / why we need it?
vllm-ascend support chunked prefill for MLA


---------

Signed-off-by: fems14 <1804143737@qq.com>

Author: fems14

docs/source/user_guide/additional_config.md

@@ -31,6 +31,7 @@ The following table lists the additional configuration options available in vLLM
 | `expert_tensor_parallel_size` | str | `0`  | Expert tensor parallel size the model to use.                                                 |
 | `refresh`                     | bool | `false` | Whether to refresh global ascend config content. This value is usually used by rlhf case.     |
 | `expert_map_path`             | str | None | When using expert load balancing for the MOE model, an expert map path needs to be passed in. |
+| `chunked_prefill_for_mla` | bool | `False` | Whether to enable the fused operator-like chunked_prefill. |
 
 The details of each config option are as follows:
 

tests/singlecard/test_chunked.py

@@ -0,0 +1,74 @@
+#
+# 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 aclgraph.
+
+Run `pytest tests/compile/test_aclgraph.py`.
+"""
+
+import os
+
+import pytest
+import torch
+from vllm import LLM, SamplingParams
+
+MODELS = ["deepseek-ai/DeepSeek-V2-Lite"]
+
+
+@pytest.mark.skipif(os.getenv("VLLM_USE_V1") == "0",
+                    reason="new chunked only support on v1")
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("max_tokens", [1])
+def test_models(
+    model: str,
+    max_tokens: int,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    return
+    with monkeypatch.context() as m:
+        prompts = "The president of the United States is"
+
+        m.setenv("VLLM_USE_V1", "1")
+
+        sampling_params = SamplingParams(
+            max_tokens=max_tokens,
+            temperature=0.0,
+        )
+
+        vllm_model = LLM(model,
+                         long_prefill_token_threshold=4,
+                         enforce_eager=True)
+        output_chunked = vllm_model.generate(prompts, sampling_params)
+        logprobs_chunked = output_chunked.outputs[0].logprobs
+        del vllm_model
+        torch.npu.empty_cache()
+
+        vllm_model = LLM(model,
+                         enforce_eager=True,
+                         additional_config={
+                             'ascend_scheduler_config': {
+                                 'enabled': True
+                             },
+                         })
+        output = vllm_model.generate(prompts, sampling_params)
+        logprobs = output.outputs[0].logprobs
+        del vllm_model
+        torch.npu.empty_cache()
+
+        logprobs_similarity = torch.cosine_similarity(
+            logprobs_chunked.flatten(), logprobs.flatten(), dim=0)
+        assert logprobs_similarity > 0.95

vllm_ascend/ascend_config.py

@@ -39,6 +39,8 @@ def __init__(self, vllm_config):
         self.expert_tensor_parallel_size = int(
             additional_config.get("expert_tensor_parallel_size", 0))
         self.expert_map_path = additional_config.get("expert_map_path", None)
+        self.chunked_prefill_for_mla = additional_config.get(
+            "chunked_prefill_for_mla", False)
 
 
 class TorchairGraphConfig: