Softmax numerical stability. (#267)

* Softmax numerical stability.

* Fix the flash-attn test.
This commit is contained in:
Laurent Mazare
2023-07-28 13:13:01 +01:00
committed by GitHub
parent 68eab38de6
commit 3eb2bc6d07
28 changed files with 117 additions and 188 deletions

View File

@ -200,7 +200,7 @@ impl MultiHeadAttention {
}
let w = {
let _timer = crate::Timer::new("qk::softmax");
qk.softmax(candle::D::Minus1)?
candle_nn::ops::softmax(&qk, candle::D::Minus1)?
};
let wv = {
let _timer = crate::Timer::new("wv::matmul");