[Attention] Refactor AttentionMetadata Preparation for Encoder-only Models (#23154)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>

vllm/model_executor/models/qwen2.py

@@ -32,6 +32,7 @@ from torch import nn
 from transformers import Qwen2Config
 
 from vllm.attention import Attention, AttentionType
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
@@ -159,7 +160,9 @@ class Qwen2Attention(nn.Module):
             rope_scaling=rope_scaling,
             dual_chunk_attention_config=dual_chunk_attention_config,
         )
-        self.attn = Attention(
+        attn_cls = (EncoderOnlyAttention
+                    if attn_type == AttentionType.ENCODER_ONLY else Attention)
+        self.attn = attn_cls(
             self.num_heads,
             self.head_dim,
             self.scaling,