Separate attention backends (#3005)

This commit is contained in:
Woosuk Kwon
2024-03-07 01:45:50 -08:00
committed by GitHub
parent cbf4c05b15
commit 2daf23ab0c
35 changed files with 561 additions and 271 deletions
+5 -5
View File
@@ -25,7 +25,7 @@ from transformers import PretrainedConfig
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import PagedAttention
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (LinearMethodBase,
MergedColumnParallelLinear,
QKVParallelLinear,
@@ -122,10 +122,10 @@ class StablelmAttention(nn.Module):
max_position=self.config.max_position_embeddings,
base=self.config.rope_theta,
)
self.attn = PagedAttention(self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_key_value_heads)
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_key_value_heads)
def forward(
self,