[Core] Refactor Attention Take 2 (#3462)

vllm/model_executor/models/qwen.py

@@ -10,9 +10,8 @@ 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,
@@ -30,8 +29,6 @@ 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 QWenMLP(nn.Module):
 
@@ -111,15 +108,13 @@ class QWenAttention(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.c_attn(hidden_states)
         q, k, v = qkv.chunk(chunks=3, 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.c_proj(attn_output)
         return output
 
@@ -153,8 +148,8 @@ class QWenBlock(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
@@ -167,7 +162,7 @@ class QWenBlock(nn.Module):
             positions=positions,
             hidden_states=hidden_states,
             kv_cache=kv_cache,
-            input_metadata=input_metadata,
+            attn_metadata=attn_metadata,
         )
 
         # Fully Connected
@@ -201,8 +196,8 @@ class QWenModel(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.wte(input_ids)
         residual = None
@@ -212,7 +207,7 @@ class QWenModel(nn.Module):
                 positions,
                 hidden_states,
                 kv_caches[i],
-                input_metadata,
+                attn_metadata,
                 residual,
             )
         hidden_states, _ = self.ln_f(hidden_states, residual)
@@ -238,11 +233,11 @@ class QWenLMHeadModel(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,