Separate attention backends (#3005)

vllm/model_executor/models/deepseek.py

@@ -29,7 +29,7 @@ from transformers import PretrainedConfig
 
 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.fused_moe import fused_moe
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (LinearMethodBase,
@@ -229,10 +229,10 @@ class DeepseekAttention(nn.Module):
             base=rope_theta,
             rope_scaling=rope_scaling,
         )
-        self.attn = PagedAttention(self.num_heads,
-                                   self.head_dim,
-                                   self.scaling,
-                                   num_kv_heads=self.num_kv_heads)
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads)
 
     def forward(
         self,