diff --git a/tensor/src/lib.rs b/tensor/src/lib.rs
index 6736c3b8..cb6be893 100644
--- a/tensor/src/lib.rs
+++ b/tensor/src/lib.rs
@@ -2,6 +2,7 @@ mod broadcast;
 mod data_type;
 mod fmt;
 mod pattern;
+mod physical_cell;
 mod reshape;
 mod slice;
 mod split;
@@ -17,8 +18,9 @@ pub type idim = i32;
 pub use data_type::DataType;
 pub use nalgebra::DVector;
 pub use pattern::{expand_indices, idx_strides, Affine, Shape};
+pub use physical_cell::PhysicalCell;
 pub use slice::SliceDim;
-pub use tensor::{Storage, Tensor};
+pub use tensor::Tensor;
 
 use std::mem::{align_of, size_of, size_of_val};
 
diff --git a/tensor/src/physical_cell.rs b/tensor/src/physical_cell.rs
new file mode 100644
index 00000000..9530b4b7
--- /dev/null
+++ b/tensor/src/physical_cell.rs
@@ -0,0 +1,59 @@
+use crate::Tensor;
+use std::ops::{Deref, DerefMut};
+
+pub trait PhysicalCell {
+    type Raw: ?Sized;
+    type Access<'a>: Deref<Target = Self::Raw>
+    where
+        Self: 'a;
+    type AccessMut<'a>: DerefMut<Target = Self::Raw>
+    where
+        Self: 'a;
+
+    unsafe fn get_unchecked(&self) -> &Self::Raw;
+    unsafe fn get_unchecked_mut(&mut self) -> &mut Self::Raw;
+    fn access(&self) -> Self::Access<'_>;
+    fn access_mut(&mut self) -> Self::AccessMut<'_>;
+}
+
+impl<Physical: PhysicalCell> Tensor<Physical> {
+    #[inline]
+    pub unsafe fn access_unchecked(&self) -> Tensor<&Physical::Raw> {
+        Tensor {
+            data_type: self.data_type,
+            shape: self.shape.clone(),
+            pattern: self.pattern.clone(),
+            physical: self.physical.get_unchecked(),
+        }
+    }
+
+    #[inline]
+    pub unsafe fn access_unchecked_mut(&mut self) -> Tensor<&mut Physical::Raw> {
+        Tensor {
+            data_type: self.data_type,
+            shape: self.shape.clone(),
+            pattern: self.pattern.clone(),
+            physical: self.physical.get_unchecked_mut(),
+        }
+    }
+
+    #[inline]
+    pub fn access(&self) -> Tensor<Physical::Access<'_>> {
+        Tensor {
+            data_type: self.data_type,
+            shape: self.shape.clone(),
+            pattern: self.pattern.clone(),
+            physical: self.physical.access(),
+        }
+    }
+
+    #[inline]
+    pub fn access_mut(&mut self) -> Tensor<Physical::AccessMut<'_>> {
+        Tensor {
+            data_type: self.data_type,
+            shape: self.shape.clone(),
+            pattern: self.pattern.clone(),
+            physical: self.physical.access_mut(),
+        }
+    }
+}
diff --git a/tensor/src/tensor.rs b/tensor/src/tensor.rs
index 7b668d9c..845c094d 100644
--- a/tensor/src/tensor.rs
+++ b/tensor/src/tensor.rs
@@ -1,7 +1,11 @@
 use crate::{expand_indices, idim, idx_strides, pattern::Pattern, udim, DataType, Shape};
 use nalgebra::{DVector, DVectorView};
 use rayon::iter::*;
-use std::ops::{Deref, DerefMut};
+use std::{
+    iter::zip,
+    ops::{Deref, DerefMut},
+    panic,
+};
 
 #[derive(Clone, Debug)]
 pub struct Tensor<Physical> {
@@ -120,66 +124,48 @@ impl<Physical> Tensor<Physical> {
             physical: f(&self.physical),
         }
     }
-
-    #[inline]
-    fn byte_offset(&self) -> usize {
-        self.pattern.offset() as usize * self.data_type.size()
-    }
 }
 
-pub trait Storage {
-    type Raw: ?Sized;
-    type Access<'a>: Deref<Target = Self::Raw>
-    where
-        Self: 'a;
-    type AccessMut<'a>: DerefMut<Target = Self::Raw>
-    where
-        Self: 'a;
-
-    unsafe fn get_unchecked(&self) -> &Self::Raw;
-    unsafe fn get_unchecked_mut(&mut self) -> &mut Self::Raw;
-    fn access(&self) -> Self::Access<'_>;
-    fn access_mut(&mut self) -> Self::AccessMut<'_>;
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+#[repr(u8)]
+pub enum Compatibility {
+    Same,
+    Squeeze,
+    Reform,
+    None,
 }
 
-impl<Physical: Storage> Tensor<Physical> {
-    #[inline]
-    pub unsafe fn access_unchecked(&self) -> Tensor<&Physical::Raw> {
-        Tensor {
-            data_type: self.data_type,
-            shape: self.shape.clone(),
-            pattern: self.pattern.clone(),
-            physical: self.physical.get_unchecked(),
+impl Compatibility {
+    pub fn between<T, U>(a: &Tensor<T>, b: &Tensor<U>) -> Self {
+        if a.data_type != b.data_type {
+            return Self::None;
         }
-    }
 
-    #[inline]
-    pub unsafe fn access_unchecked_mut(&mut self) -> Tensor<&mut Physical::Raw> {
-        Tensor {
-            data_type: self.data_type,
-            shape: self.shape.clone(),
-            pattern: self.pattern.clone(),
-            physical: self.physical.get_unchecked_mut(),
-        }
-    }
-
-    #[inline]
-    pub fn access(&self) -> Tensor<Physical::Access<'_>> {
-        Tensor {
-            data_type: self.data_type,
-            shape: self.shape.clone(),
-            pattern: self.pattern.clone(),
-            physical: self.physical.access(),
+        let mut actual_a = zip(&a.shape, a.pattern.0.as_slice()).filter(|(&d, _)| d > 1);
+        let mut actual_b = zip(&b.shape, b.pattern.0.as_slice()).filter(|(&d, _)| d > 1);
+        let mut squeeze = true;
+        loop {
+            match (actual_a.next(), actual_b.next()) {
+                (Some((da, sa)), Some((db, sb))) => {
+                    if da != db {
+                        return Self::None;
+                    }
+                    if sa != sb {
+                        squeeze = false;
+                    }
+                }
+                (Some(_), None) | (None, Some(_)) => return Self::None,
+                (None, None) => break,
+            }
         }
-    }
-
-    #[inline]
-    pub fn access_mut(&mut self) -> Tensor<Physical::AccessMut<'_>> {
-        Tensor {
-            data_type: self.data_type,
-            shape: self.shape.clone(),
-            pattern: self.pattern.clone(),
-            physical: self.physical.access_mut(),
+        if squeeze {
+            if a.shape == b.shape {
+                Self::Same
+            } else {
+                Self::Squeeze
+            }
+        } else {
+            Self::Reform
         }
     }
 }
@@ -188,14 +174,14 @@ impl<Physical: Deref<Target = [u8]>> Tensor<Physical> {
     #[inline]
     pub fn as_slice(&self) -> &[u8] {
         debug_assert!(self.is_contiguous());
-        let off = self.byte_offset();
+        let off = self.bytes_offset();
         let len = self.bytes_size();
-        &self.physical[off..][..len]
+        &self.physical[off as usize..][..len]
     }
 
     pub fn locate_start(&self) -> *const u8 {
-        let off = self.byte_offset();
-        (&self.physical[off]) as _
+        let off = self.bytes_offset();
+        (&self.physical[off as usize]) as _
     }
 
     pub fn locate(&self, indices: &DVectorView<idim>) -> Option<*const u8> {
@@ -203,18 +189,11 @@ impl<Physical: Deref<Target = [u8]>> Tensor<Physical> {
         self.physical.get(i).map(|r| r as _)
     }
 
-    #[inline]
-    pub fn as_ptr(&self) -> *const u8 {
-        let ptr = self.physical.as_ptr();
-        let offset = self.byte_offset();
-        unsafe { ptr.add(offset) }
-    }
-
     /// # Safety
     ///
     /// The caller must ensure that the `dst` can be a valid tensor physical.
     pub unsafe fn reform_to_raw(&self, dst: &mut [u8]) {
-        let src = &self.physical[self.byte_offset()..];
+        let src = &self.physical[self.bytes_offset() as usize..];
         // 计算结尾连续维度数量
         let contiguous = self.contiguous_len();
         if contiguous == self.shape.len() {
@@ -240,27 +219,29 @@ impl<Physical: Deref<Target = [u8]>> Tensor<Physical> {
     where
         U: DerefMut<Target = [u8]>,
     {
-        assert_eq!(self.data_type, dst.data_type);
-        assert_eq!(self.shape, dst.shape);
-        let contiguous = self.contiguous_len().min(dst.contiguous_len());
-        if contiguous == self.shape.len() {
-            dst.as_mut_slice().copy_from_slice(self.as_slice());
-        } else {
-            let dt = self.data_type.size();
-            // 一部分维度连续，迭代不连续的部分
-            let (iter, contiguous) = self.shape.split_at(self.shape.len() - contiguous);
-            let (n, idx_strides) = idx_strides(iter);
-            let src_pattern = self.pattern.0.view_range(..iter.len(), ..);
-            let dst_pattern = dst.pattern.0.view_range(..iter.len(), ..);
-            let src = self.locate_start() as usize;
-            let dst = dst.locate_start() as usize;
-            let count = contiguous.iter().product::<udim>() as usize * dt;
-            (0..n).into_par_iter().for_each(|i| {
-                let indices = expand_indices(i, &idx_strides, &[]);
-                let src = (src + src_pattern.dot(&indices) as usize * dt) as *const u8;
-                let dst = (dst + dst_pattern.dot(&indices) as usize * dt) as *mut u8;
-                unsafe { std::ptr::copy_nonoverlapping(src, dst, count) };
-            });
+        match Compatibility::between(self, dst) {
+            Compatibility::Same | Compatibility::Squeeze => {
+                dst.as_mut_slice().copy_from_slice(self.as_slice());
+            }
+            Compatibility::Reform => {
+                let contiguous = self.contiguous_len().min(dst.contiguous_len());
+                let dt = self.data_type.size();
+                // 一部分维度连续，迭代不连续的部分
+                let (iter, contiguous) = self.shape.split_at(self.shape.len() - contiguous);
+                let (n, idx_strides) = idx_strides(iter);
+                let src_pattern = self.pattern.0.view_range(..iter.len(), ..);
+                let dst_pattern = dst.pattern.0.view_range(..iter.len(), ..);
+                let src = self.locate_start() as usize;
+                let dst = dst.locate_start() as usize;
+                let count = contiguous.iter().product::<udim>() as usize * dt;
+                (0..n).into_par_iter().for_each(|i| {
+                    let indices = expand_indices(i, &idx_strides, &[]);
+                    let src = (src + src_pattern.dot(&indices) as usize * dt) as *const u8;
+                    let dst = (dst + dst_pattern.dot(&indices) as usize * dt) as *mut u8;
+                    unsafe { std::ptr::copy_nonoverlapping(src, dst, count) };
+                });
+            }
+            Compatibility::None => panic!("Incompatible tensors"),
         }
     }
 }
@@ -269,14 +250,14 @@ impl<Physical: DerefMut<Target = [u8]>> Tensor<Physical> {
     #[inline]
     pub fn as_mut_slice(&mut self) -> &mut [u8] {
         debug_assert!(self.is_contiguous());
-        let off = self.byte_offset();
+        let off = self.bytes_offset();
         let len = self.bytes_size();
-        &mut self.physical[off..][..len]
+        &mut self.physical[off as usize..][..len]
     }
 
     pub fn locate_start_mut(&mut self) -> *mut u8 {
-        let off = self.byte_offset();
-        (&mut self.physical[off]) as _
+        let off = self.bytes_offset();
+        (&mut self.physical[off as usize]) as _
     }
 
     pub fn locate_mut(&mut self, indices: &DVectorView<idim>) -> Option<*mut u8> {
diff --git a/transformer-cpu/src/storage.rs b/transformer-cpu/src/storage.rs
index 256a06f7..cbed0834 100644
--- a/transformer-cpu/src/storage.rs
+++ b/transformer-cpu/src/storage.rs
@@ -3,13 +3,14 @@
     ops::{Deref, DerefMut},
     rc::Rc,
 };
+use tensor::PhysicalCell;
 
 #[derive(Clone, Debug)]
 pub struct Storage(Rc<RefCell<Vec<u8>>>);
 pub struct VecRef<'a>(Ref<'a, Vec<u8>>);
 pub struct VecRefMut<'a>(RefMut<'a, Vec<u8>>);
 
-impl tensor::Storage for Storage {
+impl PhysicalCell for Storage {
     type Raw = [u8];
     type Access<'a> = VecRef<'a>;
     type AccessMut<'a> = VecRefMut<'a>;