[0.11.0][Perf] Add padding vision tower for Qwen2_5_Omni (#4041)

### What this PR does / why we need it?
This PR repalce the vision tower in Qwen2.5-Omni-Thinker model,
Qwen2_5_VisionTransformer, with AscendQwen2_5_VisionTransformer, which
use QKV padding for padding performance.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: Ting FU <futing10@huawei.com>

vllm_ascend/models/__init__.py

@@ -23,6 +23,10 @@ def register_model():
             "Qwen2_5_VLForConditionalGeneration",
             "vllm_ascend.models.qwen2_5_vl:AscendQwen2_5_VLForConditionalGeneration"
         )
+        ModelRegistry.register_model(
+            "Qwen2_5OmniModel",
+            "vllm_ascend.models.qwen2_5_omni_thinker:AscendQwen2_5OmniThinkerForConditionalGeneration"
+        )
     else:
         ModelRegistry.register_model(
             "Qwen2_5_VLForConditionalGeneration",

vllm_ascend/models/qwen2_5_omni_thinker.py

@@ -0,0 +1,54 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Adapted from vllm/model_executor/models/qwen2_5_vl.py
+# Copyright 2023 The vLLM team.
+#
+# This file is a part of the vllm-ascend project.
+#
+# 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 transformers.models.qwen2_5_omni.configuration_qwen2_5_omni import \
+    Qwen2_5OmniThinkerConfig
+from vllm.config import VllmConfig
+from vllm.model_executor.models.qwen2_5_omni_thinker import (
+    Qwen2_5OmniThinkerDummyInputsBuilder,
+    Qwen2_5OmniThinkerForConditionalGeneration,
+    Qwen2_5OmniThinkerMultiModalProcessor, Qwen2_5OmniThinkerProcessingInfo)
+from vllm.model_executor.models.utils import maybe_prefix
+from vllm.multimodal import MULTIMODAL_REGISTRY
+
+from vllm_ascend.models.qwen2_5_vl import AscendQwen2_5_VisionTransformer
+
+
+@MULTIMODAL_REGISTRY.register_processor(
+    Qwen2_5OmniThinkerMultiModalProcessor,
+    info=Qwen2_5OmniThinkerProcessingInfo,
+    dummy_inputs=Qwen2_5OmniThinkerDummyInputsBuilder)
+class AscendQwen2_5OmniThinkerForConditionalGeneration(
+        Qwen2_5OmniThinkerForConditionalGeneration):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        config: Qwen2_5OmniThinkerConfig = vllm_config.model_config.hf_config.thinker_config
+        quant_config = vllm_config.quant_config
+        # The following code reuse AscendQwen2_5_VisionTransformer from Qwen2_5_VL,
+        # which does not import any model strcut difference. And will not impact
+        # the modeling files removing.
+        self.visual = AscendQwen2_5_VisionTransformer(
+            vision_config=config.vision_config,
+            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
+            quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "visual"),
+        )

vllm_ascend/models/qwen2_5_vl.py

@@ -349,6 +349,9 @@ class AscendQwen2_5_VisionTransformer(Qwen2_5_VisionTransformer):
             ("qkv_proj", "q_proj", "q"),
             ("qkv_proj", "k_proj", "k"),
             ("qkv_proj", "v_proj", "v"),
+            ("attn.qkv.", "attn.q.", "q"),
+            ("attn.qkv.", "attn.k.", "k"),
+            ("attn.qkv.", "attn.v.", "v"),
             ("mlp.gate_up_proj.", "mlp.gate_proj.", 0),
             ("mlp.gate_up_proj.", "mlp.up_proj.", 1),
         ]
@@ -363,6 +366,11 @@ class AscendQwen2_5_VisionTransformer(Qwen2_5_VisionTransformer):
                 param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
+                if self.enable_pad and shard_id == "v":
+                    if "attn.qkv.weight" in name:
+                        param.data = self.pad_qkv_weight(param.data)
+                    if "attn.qkv.bias" in name:
+                        param.data = self.pad_qkv_bias(param.data)
                 break
             else:
                 param = params_dict[name]