Use cat for faster MQA computation. (#2043)

* Use cat for faster MQA computation.

* Move the function to utils + use it in mistral.

* Use the shared repeat-kv in a few more models.

* Fix.

1 files changed, 2 insertions, 14 deletions

diff --git a/candle-transformers/src/models/quantized_mistral.rs b/candle-transformers/src/models/quantized_mistral.rs
index e37785de..0583810a 100644
--- a/candle-transformers/src/models/quantized_mistral.rs
+++ b/candle-transformers/src/models/quantized_mistral.rs
@@ -122,18 +122,6 @@ impl Attention {
         })
     }
 
-    fn repeat_kv(&self, xs: Tensor) -> Result<Tensor> {
-        let n_rep = self.num_kv_groups;
-        if n_rep == 1 {
-            Ok(xs)
-        } else {
-            let (b_sz, num_kv_heads, seq_len, head_dim) = xs.dims4()?;
-            xs.unsqueeze(2)?
-                .expand((b_sz, num_kv_heads, n_rep, seq_len, head_dim))?
-                .reshape((b_sz, num_kv_heads * n_rep, seq_len, head_dim))
-        }
-    }
-
     fn forward(
         &mut self,
         xs: &Tensor,
@@ -172,8 +160,8 @@ impl Attention {
         };
         self.kv_cache = Some((key_states.clone(), value_states.clone()));
 
-        let key_states = self.repeat_kv(key_states)?;
-        let value_states = self.repeat_kv(value_states)?;
+        let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?;
+        let value_states = crate::utils::repeat_kv(value_states, self.num_kv_groups)?;
 
         let attn_output = {
             let scale = 1f64 / f64::sqrt(self.head_dim as f64);