Merge pull request #2937 from ljt019/fix-phi3-kv-cache-reset

fix kv cache issue with quantized_phi3 implementation

Author: ivarflakstad

candle-transformers/src/models/quantized_phi3.rs

@@ -136,6 +136,9 @@ impl LayerWeights {
         let q = self.apply_rotary_emb(&q, index_pos)?.contiguous()?;
         let k = self.apply_rotary_emb(&k, index_pos)?;
 
+        if index_pos == 0 {
+            self.kv_cache.reset();
+        }
         let (k, v) = self.kv_cache.append(&k.contiguous()?, &v.contiguous()?)?;
 
         let k = crate::utils::repeat_kv(k, self.n_head / self.n_kv_head)?;