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
+2 -2
View File
@@ -29,7 +29,7 @@ from transformers import MixtralConfig
from vllm.config import LoRAConfig
from vllm.model_executor.input_metadata import InputMetadata
from vllm.model_executor.layers.attention import PagedAttention
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (LinearMethodBase,
@@ -197,7 +197,7 @@ class MixtralAttention(nn.Module):
base=int(self.rope_theta),
is_neox_style=True,
)
self.attn = PagedAttention(
self.attn = Attention(
self.num_heads,
self.head_dim,
self.scaling,