A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 09:15 UTC
Fix attention
Woosuk Kwon committed
Feb 23, 2023 at 22:29 UTC
ba84b8728a8d0a766a636b30661836c30b17fbe6
1 file changed
+5
-1
cacheflow/models/attention.py
+5
-1
@@ -1,4 +1,4 @@
1
-from typing import Optional, Tuple
1
+from typing import Optional
2
3
import torch
4
import torch.nn as nn
@@ -24,8 +24,12 @@ class OPTCacheFlowAttention(nn.Module):
24
key: torch.Tensor,
25
value: torch.Tensor,
26
) -> None:
27
+ query = query.unsqueeze(0)
28
+ key = key.unsqueeze(0)
29
+ value = value.unsqueeze(0)
30
out = xops.memory_efficient_attention(
31
query, key, value, attn_bias=self.attention_mask, scale=self.scale)
32
+ out = out.squeeze(0)
33
# FIXME(woosuk): Directly write the attention output.
34
output.copy_(out, non_blocking=True)
35