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
@@ -23,7 +23,7 @@ from transformers import GemmaConfig
from vllm.config import LoRAConfig
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.attention import PagedAttention
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,
@@ -123,10 +123,10 @@ class GemmaAttention(nn.Module):
base=self.rope_theta,
is_neox_style=True,
)
self.attn = PagedAttention(self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads)
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
num_kv_heads=self.num_kv_heads)
def forward(
self,