diff --git a/model-parameters/src/memory/inside_memory.rs b/model-parameters/src/memory/inside_memory.rs
index 83420d6c..bc98991d 100644
--- a/model-parameters/src/memory/inside_memory.rs
+++ b/model-parameters/src/memory/inside_memory.rs
@@ -2,7 +2,7 @@
 use half::{bf16, f16};
 
 impl Memory<Vec<u8>> {
-    pub fn cast(src: &impl Llama2, new_dtype: DataType) -> Self {
+    pub fn cast(src: &dyn Llama2, new_dtype: DataType) -> Self {
         let from = src.data_type();
         let mut blob = Vec::with_capacity(src.size() * new_dtype.size() / from.size());
         let mut append = |src: &[u8]| {
diff --git a/transformer-cpu/src/cache.rs b/transformer-cpu/src/cache.rs
new file mode 100644
index 00000000..c90a5374
--- /dev/null
+++ b/transformer-cpu/src/cache.rs
@@ -0,0 +1,19 @@
+use model_parameters::Llama2;
+
+pub(super) struct LayerCache(Vec<u8>);
+
+impl LayerCache {
+    pub fn new(model: &dyn Llama2, batch: usize) -> Self {
+        let n = batch;
+        let dkv = model.num_key_value_heads();
+        let ds = model.max_position_embeddings();
+        let dh = model.hidden_size() / model.num_attention_heads();
+        Self(vec![0; 2 * n * dkv * ds * dh])
+    }
+
+    #[inline]
+    pub fn get(&mut self) -> (&mut [u8], &mut [u8]) {
+        let mid = self.0.len() / 2;
+        self.0.split_at_mut(mid)
+    }
+}
diff --git a/transformer-cpu/src/lib.rs b/transformer-cpu/src/lib.rs
index b02cad0c..698dfd65 100644
--- a/transformer-cpu/src/lib.rs
+++ b/transformer-cpu/src/lib.rs
@@ -1,19 +1,44 @@
+mod cache;
+
+use cache::LayerCache;
 use model_parameters::{DataType, Llama2, Memory};
 
 pub struct Transformer {
     model: Box<dyn Llama2>,
+    cache: Vec<LayerCache>,
 }
 
-impl<T> From<T> for Transformer
-where
-    T: 'static + Llama2,
-{
-    fn from(value: T) -> Self {
-        let model: Box<dyn Llama2> = if value.data_type() == DataType::BF16 {
-            Box::new(Memory::cast(&value, DataType::F32))
-        } else {
-            Box::new(value)
+impl Transformer {
+    pub fn new(model: Box<dyn Llama2>, batch: usize) -> Self {
+        let model = match model.data_type() {
+            DataType::BF16 => Box::new(Memory::cast(&*model, DataType::F32)),
+            _ => model,
         };
-        Self { model }
+        let cache = (0..model.num_hidden_layers())
+            .map(|_| LayerCache::new(&*model, batch))
+            .collect();
+        Self { model, cache }
     }
 }
+
+#[test]
+fn test_build() {
+    use model_parameters::SafeTensorError;
+    use std::time::Instant;
+
+    let t0 = Instant::now();
+    let safetensors = Memory::load_safetensors("../../TinyLlama-1.1B-Chat-v1.0");
+    let t1 = Instant::now();
+    println!("mmap {:?}", t1 - t0);
+
+    let safetensors = match safetensors {
+        Ok(m) => m,
+        Err(SafeTensorError::Io(e)) if e.kind() == std::io::ErrorKind::NotFound => return,
+        Err(e) => panic!("{e:?}"),
+    };
+
+    let t0 = Instant::now();
+    let _transformer = Transformer::new(Box::new(safetensors), 1);
+    let t1 = Instant::now();
+    println!("build transformer {:?}", t1 - t0);
+}