[bugfix] fix Error 'ValueError: Duplicate layer name' (#5280)

### What this PR does / why we need it?
When matmul_and_reduce is enabled, the prefix attribute is required.
However, in some models, the prefix is not passed correctly, causing
errors when starting the service.
The issue of incorrect prefix passing will be fixed in vLLM in the
future.

- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: Wang Kunpeng <1289706727@qq.com>

.github/workflows/_e2e_test.yaml

@@ -115,6 +115,7 @@ jobs:
           pytest -sv --durations=0 tests/e2e/singlecard/test_sampler.py
           pytest -sv --durations=0 tests/e2e/singlecard/test_vlm.py
           pytest -sv --durations=0 tests/e2e/singlecard/test_xlite.py
+          pytest -sv --durations=0 tests/e2e/singlecard/test_models.py
           pytest -sv --durations=0 tests/e2e/singlecard/pooling/
           pytest -sv --durations=0 tests/e2e/singlecard/compile/test_norm_quant_fusion.py
           pytest -sv --durations=0 tests/e2e/singlecard/test_multistream_overlap_shared_expert.py

tests/e2e/singlecard/test_models.py

@@ -0,0 +1,36 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+# Adapted from vllm/tests/entrypoints/llm/test_guided_generate.py
+# 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.
+#
+
+import os
+
+from tests.e2e.conftest import VllmRunner
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+
+def test_minicpm_2b() -> None:
+    example_prompts = [
+        "Hello, my name is",
+    ]
+    max_tokens = 5
+
+    with VllmRunner("openbmb/MiniCPM-2B-sft-bf16",
+                    max_model_len=512,
+                    gpu_memory_utilization=0.7) as runner:
+        runner.generate_greedy(example_prompts, max_tokens)

vllm_ascend/ops/linear.py

@@ -36,7 +36,7 @@ from vllm.model_executor.layers.quantization.base_config import \
 from vllm.model_executor.utils import set_weight_attrs
 
 from vllm_ascend.ops.linear_op import get_parallel_op, get_replicated_op
-from vllm_ascend.utils import maybe_trans_nz
+from vllm_ascend.utils import enable_sp, maybe_trans_nz
 
 
 class AscendUnquantizedLinearMethod(UnquantizedLinearMethod):
@@ -219,6 +219,9 @@ class AscendRowParallelLinear(RowParallelLinear):
     and the original TP group in other modules.
     """
 
+    # NOTE: Globally unique prefix identifier used in SP scenarios
+    unique_prefix_idx = 0
+
     def __init__(
         self,
         input_size: int,
@@ -234,14 +237,15 @@ class AscendRowParallelLinear(RowParallelLinear):
         return_bias: bool = True,
         disable_tp: bool = False,
     ):
-        compilation_config = get_current_vllm_config().compilation_config
+        # TODO(kunpengW-code): Specifying the prefix in linear layers of some models in the vLLM.
-        # TODO(shaopeng-666): Remove the visual check after the mm model reconstruction is complete.
+        if enable_sp():
-        # TODO(MengqingCao): Remove the empty string check, after specifying the prefix in linear layers of some models in the vLLM.
+            compilation_config = get_current_vllm_config().compilation_config
-        if prefix in compilation_config.static_forward_context and \
+            unique_prefix = prefix
-            prefix != "" and \
+            if prefix in compilation_config.static_forward_context:
-            "visual" not in prefix:
+                unique_prefix = f"{prefix}.unique_prefix{AscendRowParallelLinear.unique_prefix_idx}"
-            raise ValueError(f"Duplicate layer name: {prefix}")
+                AscendRowParallelLinear.unique_prefix_idx += 1
-        compilation_config.static_forward_context[prefix] = self
+            self.unique_prefix = unique_prefix
+            compilation_config.static_forward_context[unique_prefix] = self
 
         self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(
             disable_tp, prefix, self, "row")

vllm_ascend/ops/linear_op.py

@@ -484,6 +484,10 @@ class SequenceColumnParallelOp(CustomColumnParallelOp):
 
 class SequenceRowParallelOp(CustomRowParallelOp):
 
+    def __init__(self, layer):
+        super().__init__(layer)
+        self.unique_prefix = None
+
     def apply_impl(
         self, input_: torch.Tensor
     ) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
@@ -509,7 +513,7 @@ class SequenceRowParallelOp(CustomRowParallelOp):
                                              bias=bias_)
         else:
             output = torch.ops.vllm.matmul_and_reduce(input_parallel,
-                                                      self.prefix)
+                                                      self.unique_prefix)
 
         output_bias = self.bias if self.skip_bias_add else None
         return output, output_bias
@@ -602,6 +606,7 @@ class SequenceRowParallelOp(CustomRowParallelOp):
         super().update_attrs()
         self.input_is_parallel = self.layer.input_is_parallel
         self.reduce_results = self.layer.reduce_results
+        self.unique_prefix = self.layer.unique_prefix
 
 
 def _get_column_parallel_op(