[Core] Refactor Attention Take 2 (#3462)

vllm/model_executor/models/gpt2.py

@@ -17,15 +17,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only GPT-2 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 GPT2Config
 
-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,
@@ -41,8 +40,6 @@ 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 GPT2Attention(nn.Module):
 
@@ -79,14 +76,12 @@ class GPT2Attention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         qkv, _ = self.c_attn(hidden_states)
         q, k, v = qkv.chunk(chunks=3, dim=-1)
-        key_cache, value_cache = kv_cache
-        attn_output = self.attn(q, k, v, key_cache, value_cache,
-                                input_metadata)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
         attn_output, _ = self.c_proj(attn_output)
         return attn_output
 
@@ -144,15 +139,15 @@ class GPT2Block(nn.Module):
     def forward(
         self,
         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.ln_1(hidden_states)
         attn_output = self.attn(
             hidden_states=hidden_states,
             kv_cache=kv_cache,
-            input_metadata=input_metadata,
+            attn_metadata=attn_metadata,
         )
         # residual connection
         hidden_states = attn_output + residual
@@ -190,8 +185,8 @@ class GPT2Model(nn.Module):
         self,
         input_ids: torch.Tensor,
         position_ids: torch.Tensor,
-        kv_caches: List[KVCache],
-        input_metadata: InputMetadata,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         inputs_embeds = self.wte(input_ids)
         position_embeds = self.wpe(position_ids)
@@ -199,7 +194,7 @@ class GPT2Model(nn.Module):
 
         for i in range(len(self.h)):
             layer = self.h[i]
-            hidden_states = layer(hidden_states, kv_caches[i], input_metadata)
+            hidden_states = layer(hidden_states, kv_caches[i], attn_metadata)
 
         hidden_states = self.ln_f(hidden_states)
         return hidden_states
@@ -224,11 +219,11 @@ class GPT2LMHeadModel(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.transformer(input_ids, positions, kv_caches,
-                                         input_metadata)
+                                         attn_metadata)
         return hidden_states
 
     def compute_logits(self, hidden_states: torch.Tensor,