Compare commits

...

1 Commits

Author SHA1 Message Date
jmorganca
d8b3e09fb7 llm: enable flash attention by default 2024-06-08 22:55:22 -07:00

View File

@ -90,6 +90,7 @@ func init() {
NumParallel = 1 NumParallel = 1
MaxRunners = 1 MaxRunners = 1
MaxQueuedRequests = 512 MaxQueuedRequests = 512
FlashAttention = true
LoadConfig() LoadConfig()
} }