Separate attention backends (#3005)
This commit is contained in:
@@ -25,7 +25,7 @@ from transformers import OPTConfig
|
||||
|
||||
from vllm.model_executor.input_metadata import InputMetadata
|
||||
from vllm.model_executor.layers.activation import get_act_fn
|
||||
from vllm.model_executor.layers.attention import PagedAttention
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
LinearMethodBase,
|
||||
QKVParallelLinear,
|
||||
@@ -89,9 +89,9 @@ class OPTAttention(nn.Module):
|
||||
bias=bias,
|
||||
linear_method=linear_method,
|
||||
)
|
||||
self.attn = PagedAttention(self.num_heads,
|
||||
self.head_dim,
|
||||
scale=self.scaling)
|
||||
self.attn = Attention(self.num_heads,
|
||||
self.head_dim,
|
||||
scale=self.scaling)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user