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

@ -17,7 +17,7 @@ impl LogitsProcessor {
pub fn sample(&mut self, logits: &Tensor) -> Result<u32> {
let logits = logits.to_dtype(DType::F32)?;
let next_token = if let Some(temperature) = self.temperature {
let prs = (&logits / temperature)?.softmax(D::Minus1)?;
let prs = candle_nn::ops::softmax(&(&logits / temperature)?, D::Minus1)?;
let prs: Vec<f32> = prs.to_vec1()?;
let distr = rand::distributions::WeightedIndex::new(prs).map_err(Error::wrap)?;
distr.sample(&mut self.rng) as u32