[Core] Refactor Attention Take 2 (#3462)

vllm/model_executor/models/phi.py

@@ -35,15 +35,14 @@
 # OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 """Inference-only Phi-1.5 model compatible with HuggingFace weights."""
-from typing import List, Optional, Tuple
+from typing import List, Optional
 
 import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.model_executor.input_metadata import InputMetadata
+from vllm.attention import Attention, AttentionMetadata
 from vllm.model_executor.layers.activation import get_act_fn
-from vllm.model_executor.layers.attention import Attention
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                LinearMethodBase,
                                                QKVParallelLinear,
@@ -60,8 +59,6 @@ from vllm.model_executor.weight_utils import (default_weight_loader,
                                               hf_model_weights_iterator)
 from vllm.sequence import SamplerOutput
 
-KVCache = Tuple[torch.Tensor, torch.Tensor]
-
 
 class PhiAttention(nn.Module):
 
@@ -115,14 +112,13 @@ class PhiAttention(nn.Module):
         self,
         position_ids: torch.Tensor,
         hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.chunk(chunks=3, dim=-1)
         q, k = self.rotary_emb(position_ids, q, k)
-        k_cache, v_cache = kv_cache
-        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
         output, _ = self.dense(attn_output)
         return output
 
@@ -172,8 +168,8 @@ class PhiLayer(nn.Module):
         self,
         position_ids: torch.Tensor,
         hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         residual = hidden_states
         hidden_states = self.input_layernorm(hidden_states)
@@ -181,7 +177,7 @@ class PhiLayer(nn.Module):
             position_ids=position_ids,
             hidden_states=hidden_states,
             kv_cache=kv_cache,
-            input_metadata=input_metadata,
+            attn_metadata=attn_metadata,
         )
         feed_forward_hidden_states = self.mlp(hidden_states)
         hidden_states = attn_outputs + feed_forward_hidden_states + residual
@@ -209,8 +205,8 @@ class PhiModel(nn.Module):
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         hidden_states = self.embed_tokens(input_ids)
         for i in range(self.config.num_hidden_layers):
@@ -219,7 +215,7 @@ class PhiModel(nn.Module):
                 positions,
                 hidden_states,
                 kv_caches[i],
-                input_metadata,
+                attn_metadata,
             )
 
         hidden_states = self.final_layernorm(hidden_states)
@@ -248,11 +244,11 @@ class PhiForCausalLM(nn.Module):
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, kv_caches,
-                                   input_metadata)
+                                   attn_metadata)
 
         return hidden_states