Separate attention backends (#3005)

vllm/model_executor/models/qwen2.py

@@ -30,7 +30,7 @@ from transformers import Qwen2Config
 
 from vllm.model_executor.input_metadata import InputMetadata
 from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.attention import Attention
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (LinearMethodBase,
                                                MergedColumnParallelLinear,
@@ -135,11 +135,11 @@ class Qwen2Attention(nn.Module):
             max_position=max_position,
             base=self.rope_theta,
         )
-        self.attn = PagedAttention(self.num_heads,
-                                   self.head_dim,
-                                   self.scaling,
-                                   num_kv_heads=self.num_kv_heads,
-                                   sliding_window=self.sliding_window)
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads,
+                              sliding_window=self.sliding_window)
 
     def forward(
         self,