feat(moe): Add patch for Shared Fused MoE integration with DeepSeek-V2 and Llama4

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>

Author: yiz-liu

vllm_ascend/models/deepseek_v2.py

@@ -55,7 +55,6 @@
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import (
     default_weight_loader, maybe_remap_kv_scale_name)
-from vllm.model_executor.models import deepseek_v2
 from vllm.model_executor.models.deepseek_v2 import \
     DeepseekV2ForCausalLM  # noqa: E501
 from vllm.model_executor.models.deepseek_v2 import \
@@ -69,14 +68,11 @@
 from vllm.sequence import IntermediateTensors
 
 from vllm_ascend.ascend_config import get_ascend_config
-from vllm_ascend.ops.common_fused_moe import AscendSharedFusedMoE
 from vllm_ascend.ops.fused_moe import AscendFusedMoE
 from vllm_ascend.quantization.quant_config import AscendLinearMethod
 from vllm_ascend.quantization.w8a8_dynamic import AscendW8A8DynamicLinearMethod
 from vllm_ascend.utils import dispose_tensor
 
-deepseek_v2.SharedFusedMoE = AscendSharedFusedMoE
-
 
 class CustomDeepseekV2SiluAndMul(SiluAndMul):
 

vllm_ascend/patch/platform/patch_common/__init__.py

@@ -16,3 +16,4 @@
 #
 
 import vllm_ascend.patch.platform.patch_common.patch_distributed  # noqa
+import vllm_ascend.patch.platform.patch_common.patch_shared_fused_moe  # noqa

vllm_ascend/patch/platform/patch_common/patch_shared_fused_moe.py

@@ -0,0 +1,21 @@
+# 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.
+
+from vllm.model_executor.models import deepseek_v2, llama4
+
+from vllm_ascend.ops.common_fused_moe import AscendSharedFusedMoE
+
+deepseek_v2.SharedFusedMoE = AscendSharedFusedMoE
+llama4.SharedFusedMoE = AscendSharedFusedMoE