[Core] Refactor Attention Take 2 (#3462)

vllm/model_executor/models/baichuan.py

@@ -25,9 +25,8 @@ 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.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (LinearMethodBase,
                                                MergedColumnParallelLinear,
@@ -45,8 +44,6 @@ 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]
-
 
 def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
     closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
@@ -170,15 +167,14 @@ class BaiChuanAttention(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.W_pack(hidden_states)
         q, k, v = qkv.chunk(chunks=3, dim=-1)
         if self.postion_embedding != "ALIBI":
             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
 
@@ -217,8 +213,8 @@ class BaiChuanDecoderLayer(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],
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
@@ -232,7 +228,7 @@ class BaiChuanDecoderLayer(nn.Module):
             positions=positions,
             hidden_states=hidden_states,
             kv_cache=kv_cache,
-            input_metadata=input_metadata,
+            attn_metadata=attn_metadata,
         )
 
         # Fully Connected
@@ -267,8 +263,8 @@ class BaiChuanModel(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
@@ -278,7 +274,7 @@ class BaiChuanModel(nn.Module):
                 positions,
                 hidden_states,
                 kv_caches[i],
-                input_metadata,
+                attn_metadata,
                 residual,
             )
         hidden_states, _ = self.norm(hidden_states, residual)
@@ -303,11 +299,11 @@ class BaiChuanBaseForCausalLM(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,