Use flashinfer for Gemma 2.
This commit is contained in:
parent
cf04a43fb1
commit
ce7e356561
|
@ -94,7 +94,7 @@ fn resolve_attention(config: &Option<Config>, lora_adapters: &Option<String>) ->
|
||||||
prefix_caching = Some("0".to_string());
|
prefix_caching = Some("0".to_string());
|
||||||
}
|
}
|
||||||
match config.model_type.as_deref() {
|
match config.model_type.as_deref() {
|
||||||
Some("gemma2") | Some("falcon") | Some("deepseek_v2") => {
|
Some("falcon") | Some("deepseek_v2") => {
|
||||||
// Required because gemma2 needs bfloat16 which is not supported by
|
// Required because gemma2 needs bfloat16 which is not supported by
|
||||||
// flashinfer ?
|
// flashinfer ?
|
||||||
if attention.is_none() {
|
if attention.is_none() {
|
||||||
|
|
Loading…
Reference in New Issue