Use softmax-last-dim where possible. (#1057)

2025-06-17 02:58:50 +00:00 · 2023-10-08 13:16:42 +01:00
parent 9abeddd750
commit 783735cf22
5 changed files with 5 additions and 5 deletions
--- a/candle-transformers/src/models/t5.rs
+++ b/candle-transformers/src/models/t5.rs
@ -441,7 +441,7 @@ impl T5Attention {

        let attn_weights = {
            let _enter = self.span_sm.enter();
-            candle_nn::ops::softmax(&scores, D::Minus1)?
+            candle_nn::ops::softmax_last_dim(&scores)?
        };
        let attn_output = attn_weights.matmul(&v)?;
        let attn_output = attn_output