[Core] Refactor Attention Take 2 (#3462)

vllm/model_executor/models/deepseek.py

@@ -21,15 +21,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Deepseek model."""
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, 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 SiluAndMul
-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,
@@ -51,8 +50,6 @@ 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 DeepseekMLP(nn.Module):
 
@@ -239,14 +236,13 @@ class DeepseekAttention(nn.Module):
         self,
         positions: 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.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
         q, k = self.rotary_emb(positions, 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.o_proj(attn_output)
         return output
 
@@ -294,8 +290,8 @@ class DeepseekDecoderLayer(nn.Module):
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
-        kv_cache: KVCache,
-        input_metadata: InputMetadata,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
         residual: Optional[torch.Tensor],
     ) -> torch.Tensor:
         # Self Attention
@@ -309,7 +305,7 @@ class DeepseekDecoderLayer(nn.Module):
             positions=positions,
             hidden_states=hidden_states,
             kv_cache=kv_cache,
-            input_metadata=input_metadata,
+            attn_metadata=attn_metadata,
         )
 
         # Fully Connected
@@ -346,15 +342,15 @@ class DeepseekModel(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)
         residual = None
         for i in range(len(self.layers)):
             layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states,
-                                            kv_caches[i], input_metadata,
+                                            kv_caches[i], attn_metadata,
                                             residual)
         hidden_states, _ = self.norm(hidden_states, residual)
         return hidden_states
@@ -379,11 +375,11 @@ class DeepseekForCausalLM(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
 
     def compute_logits(self, hidden_states: torch.Tensor,