diff --git a/src/buffer.rs b/src/buffer.rs
index 2701295a..1fac9fda 100644
--- a/src/buffer.rs
+++ b/src/buffer.rs
@@ -461,7 +461,7 @@ impl<'a, T, D: Device> Buffer<'a, T, D> {
 }
 
 #[cfg(feature = "cpu")]
-impl<'a, Mods: OnDropBuffer, T, S: Shape> Buffer<'a, T, CPU<Mods>, S> {
+impl<'a, T, S: Shape> Buffer<'a, T, CPU<Base>, S> {
     /// Constructs a deviceless `Buffer` out of a host pointer and a length.
     /// # Example
     /// ```
@@ -484,15 +484,17 @@ impl<'a, Mods: OnDropBuffer, T, S: Shape> Buffer<'a, T, CPU<Mods>, S> {
     /// The pointer must be valid.
     /// The `Buffer` does not manage deallocation of the allocated memory.
     #[inline]
-    pub unsafe fn from_raw_host(ptr: *mut T, len: usize) -> Buffer<'a, T, CPU<Mods>, S> {
-        let x = 2;
-        todo!()
-        // Buffer {
-        //     data: self.wrap_in_base(CPUPtr::from_ptr(ptr, len, AllocFlag::Wrapper)),
-        //     device: None,
-        // }
+    pub unsafe fn from_raw_host(ptr: *mut T, len: usize) -> Buffer<'a, T, CPU<Base>, S> {
+        Buffer {
+            data: CPUPtr::from_ptr(ptr, len, AllocFlag::Wrapper),
+            device: None,
+        }
     }
 
+}
+
+#[cfg(feature = "cpu")]
+impl<'a, Mods: OnDropBuffer, T, S: Shape> Buffer<'a, T, CPU<Mods>, S> {
     /// Constructs a `Buffer` out of a host pointer and a length.
     /// The provided device can be used to shorten operation calls.
     ///
diff --git a/src/devices/opencl/cl_device.rs b/src/devices/opencl/cl_device.rs
index 9b889ab3..d7a44b41 100644
--- a/src/devices/opencl/cl_device.rs
+++ b/src/devices/opencl/cl_device.rs
@@ -8,7 +8,7 @@ use crate::{
     impl_buffer_hook_traits, impl_retriever, impl_wrapped_data, pass_down_grad_fn,
     pass_down_optimize_mem_graph, pass_down_tape_actions, pass_down_use_gpu_or_cpu, Alloc, Base,
     Buffer, Cached, CachedCPU, CloneBuf, Device, Module, OnDropBuffer, OnNewBuffer, Setup,
-    WrappedData, CPU,
+    WrappedData, CPU, pass_down_replace_buf,
 };
 use crate::{PtrConv, Shape};
 
@@ -309,6 +309,7 @@ impl<Mods> crate::LazyRun for OpenCL<Mods> {}
 
 pass_down_tape_actions!(OpenCL);
 pass_down_grad_fn!(OpenCL);
+pass_down_replace_buf!(OpenCL);
 
 #[cfg(test)]
 mod tests {
diff --git a/src/devices/opencl/unified.rs b/src/devices/opencl/unified.rs
index ea5fe0a0..5f2f3f32 100644
--- a/src/devices/opencl/unified.rs
+++ b/src/devices/opencl/unified.rs
@@ -56,7 +56,7 @@ impl<D: Device> UnifiedMemChain<D> for Base {
 /// This function is used in the `constuct_buffer()` function.
 /// # Safety
 /// The host pointer inside the no_drop `Buffer` must live as long as the resulting pointer.
-pub unsafe fn to_cached_unified<OclMods: OnDropBuffer, CpuMods: OnDropBuffer, T, S: Shape>(
+pub unsafe fn to_cached_unified<OclMods, CpuMods, T, S>(
     device: &OpenCL<OclMods>,
     no_drop: Buffer<T, CPU<CpuMods>, S>,
     cache: &mut HashMap<
@@ -65,7 +65,13 @@ pub unsafe fn to_cached_unified<OclMods: OnDropBuffer, CpuMods: OnDropBuffer, T,
         BuildHasherDefault<LocationHasher>,
     >,
     location: HashLocation<'static>,
-) -> crate::Result<*mut c_void> {
+) -> crate::Result<*mut c_void>
+where
+    OclMods: OnDropBuffer,
+    CpuMods: OnDropBuffer,
+    T: 'static,
+    S: Shape,
+{
     // use the host pointer to create an OpenCL buffer
     let cl_ptr = create_buffer(
         device.ctx(),
@@ -78,7 +84,7 @@ pub unsafe fn to_cached_unified<OclMods: OnDropBuffer, CpuMods: OnDropBuffer, T,
         location,
         Rc::new(CLPtr {
             ptr: cl_ptr,
-            host_ptr: no_drop.base().ptr as *mut u8,
+            host_ptr: no_drop.base().ptr as *mut T,
             len: no_drop.len(),
             flag: AllocFlag::None,
         }),
@@ -132,7 +138,9 @@ pub fn construct_buffer<'a, OclMods: OnDropBuffer, CpuMods: OnDropBuffer, T: 'st
 
     // if buffer was already converted, return the cache entry.
     if let Some(rawcl) = cache.get(&location) {
-        let rawcl = rawcl.downcast_ref::<<OpenCL::<OclMods> as Device>::Base<T, S>>().unwrap();
+        let rawcl = rawcl
+            .downcast_ref::<<OpenCL<OclMods> as Device>::Base<T, S>>()
+            .unwrap();
         let data = device.base_to_data::<T, S>(CLPtr {
             ptr: rawcl.ptr,
             host_ptr: rawcl.host_ptr as *mut T,
@@ -300,9 +308,8 @@ mod tests {
     #[cfg(unified_cl)]
     #[test]
     fn test_cpu_to_unified_is_reusing_converted_buf() -> crate::Result<()> {
-        use std::time::Instant;
-
         use crate::{Base, Cached, HashLocation, Retriever};
+        use std::time::Instant;
 
         let cl_dev = OpenCL::<Cached<Base>>::new(0)?;
         let device = CPU::<Cached<Base>>::new();
@@ -310,7 +317,7 @@ mod tests {
         let mut dur = 0.;
 
         for _ in 0..100 {
-            let mut buf: Buffer<i32,  _> = device.retrieve::<0>(6, ());
+            let mut buf: Buffer<i32, _> = device.retrieve::<0>(6, ());
 
             buf.copy_from_slice(&[1, 2, 3, 4, 5, 6]);
 
diff --git a/src/exec_on_cpu.rs b/src/exec_on_cpu.rs
index 0d9d9f5b..cbed7f53 100644
--- a/src/exec_on_cpu.rs
+++ b/src/exec_on_cpu.rs
@@ -219,9 +219,9 @@ macro_rules! to_cpu {
 /// The old `Buffer`s are shadowed.
 #[macro_export]
 macro_rules! to_raw_host {
-    ($cpu_ty:ty, $($t:ident),*) => {
+    ($cpu:expr, $($t:ident),*) => {
         $(
-            let $t = &unsafe { $crate::Buffer::<_, $cpu_ty, ()>::from_raw_host($t.data.host_ptr, $t.len()) };
+            let $t = &unsafe { $crate::Buffer::<_, _, ()>::from_raw_host_device($cpu, $t.data.host_ptr, $t.len()) };
         )*
     };
 }
@@ -230,10 +230,10 @@ macro_rules! to_raw_host {
 /// New names for the `CPU` `Buffer`s are provided by the user.
 #[macro_export]
 macro_rules! to_raw_host_mut {
-    ($cpu_ty:ty, $($t:ident, $cpu_name:ident),*) => {
+    ($cpu:expr, $($t:ident, $cpu_name:ident),*) => {
         $(
             #[allow(unused_mut)]
-            let mut $cpu_name = &mut unsafe { $crate::Buffer::<_, $cpu_ty, ()>::from_raw_host($t.data.host_ptr, $t.len()) };
+            let mut $cpu_name = &mut unsafe { $crate::Buffer::<_, _, ()>::from_raw_host_device($cpu, $t.data.host_ptr, $t.len()) };
         )*
     };
 }
diff --git a/src/exec_on_cpu/cl_may_unified.rs b/src/exec_on_cpu/cl_may_unified.rs
index 0921fec8..bf0f88da 100644
--- a/src/exec_on_cpu/cl_may_unified.rs
+++ b/src/exec_on_cpu/cl_may_unified.rs
@@ -31,7 +31,7 @@ where
 
         // host ptr buffer
         let no_drop = f(&device.cpu, &unsafe {
-            Buffer::from_raw_host(x.base().host_ptr, x.len())
+            Buffer::from_raw_host_device(&device.cpu, x.base().host_ptr, x.len())
         });
 
         // convert host ptr / CPU buffer into a host ptr + OpenCL ptr buffer
@@ -73,7 +73,7 @@ where
     if device.unified_mem() {
         return {
             f(&cpu, &mut unsafe {
-                Buffer::from_raw_host(lhs.base().host_ptr, lhs.len())
+                Buffer::from_raw_host_device(&cpu, lhs.base().host_ptr, lhs.len())
             });
             Ok(())
         };
@@ -87,12 +87,7 @@ where
 /// This is way faster than [cpu_exec_binary], as new memory is not allocated.
 ///
 /// `cpu_exec_binary_may_unified` can be used interchangeably with [cpu_exec_binary].
-pub fn cpu_exec_binary_may_unified<
-    'a,
-    T,
-    F,
-    Mods: OnDropBuffer + UnifiedMemChain<OpenCL<Mods>> + Retrieve<OpenCL<Mods>, T> + 'static,
->(
+pub fn cpu_exec_binary_may_unified<'a, T, F, Mods>(
     device: &'a OpenCL<Mods>,
     lhs: &Buffer<T, OpenCL<Mods>>,
     rhs: &Buffer<T, OpenCL<Mods>>,
@@ -105,6 +100,7 @@ where
         &Buffer<'_, T, CachedCPU>,
         &Buffer<'_, T, CachedCPU>,
     ) -> Buffer<'b, T, CachedCPU>,
+    Mods: UnifiedMemChain<OpenCL<Mods>> + Retrieve<OpenCL<Mods>, T> + 'static,
 {
     // TODO: use compile time unified_cl flag -> get from custos?
     #[cfg(not(feature = "realloc"))]
@@ -116,8 +112,8 @@ where
         // host ptr buffer
         let no_drop = f(
             &device.cpu,
-            &unsafe { Buffer::from_raw_host(lhs.base().host_ptr, lhs.len()) },
-            &unsafe { Buffer::from_raw_host(rhs.base().host_ptr, rhs.len()) },
+            &unsafe { Buffer::from_raw_host_device(&device.cpu, lhs.base().host_ptr, lhs.len()) },
+            &unsafe { Buffer::from_raw_host_device(&device.cpu, rhs.base().host_ptr, rhs.len()) },
         );
 
         // convert host ptr / CPU buffer into a host ptr + OpenCL ptr buffer
@@ -163,8 +159,8 @@ where
         return {
             f(
                 &cpu,
-                &mut unsafe { Buffer::from_raw_host(lhs.base().host_ptr, lhs.len()) },
-                &unsafe { Buffer::from_raw_host(rhs.base().host_ptr, rhs.len()) },
+                &mut unsafe { Buffer::from_raw_host_device(&cpu, lhs.base().host_ptr, lhs.len()) },
+                &unsafe { Buffer::from_raw_host_device(&cpu, rhs.base().host_ptr, rhs.len()) },
             );
             Ok(())
         };
@@ -242,8 +238,8 @@ macro_rules! cl_cpu_exec_unified_mut {
     ($device:ident, $($t:ident),* WRITE_TO<$($write_to:ident, $from:ident),*> $op:expr) => {{
         // TODO: add to graph?:     convert.node = device.graph().add(convert.len(), matrix.node.idx);
         if $device.unified_mem() {
-            $crate::to_raw_host!($crate::CPU::<$crate::CachedModule<$crate::Base, $crate::CPU>>, $($t),*);
-            $crate::to_raw_host_mut!($crate::CPU::<$crate::CachedModule<$crate::Base, $crate::CPU>>, $($write_to, $from),*);
+            $crate::to_raw_host!(&$device.cpu, $($t),*);
+            $crate::to_raw_host_mut!(&$device.cpu, $($write_to, $from),*);
             $op;
 
         } else {
diff --git a/src/modules/lazy.rs b/src/modules/lazy.rs
index e55d40ef..57927200 100644
--- a/src/modules/lazy.rs
+++ b/src/modules/lazy.rs
@@ -24,7 +24,7 @@ type Buffers = HashMap<UniqueId, Box<dyn Any>, BuildHasherDefault<NoHasher>>;
 #[derive(Default)]
 pub struct Lazy<Mods> {
     pub modules: Mods,
-    alloc_later: RefCell<Vec<(Id, fn(&mut Buffers, Id, &dyn Any))>>,
+    alloc_later: RefCell<Vec<(Id, fn(&mut Buffers, Id, &dyn Any))>>, // could use D generic instead of dyn Any (required LazyModule structure)
     allocated: Cell<bool>,
     buffers: RefCell<Buffers>,
     graph: RefCell<LazyGraph>,
@@ -199,6 +199,9 @@ where
         alloc_later.push((id, |buffers, id, device| {
             let device = device.downcast_ref::<D>().unwrap();
             // TODO: should be fixable - (lazy) -> either return error or fix
+            // creating buffers (with data) is not lazy - they are allocated instantly
+            // these are then added to `buffers` with their ID (which is the pointing address)
+            // new IDs start at 0. 1, 2, 3, ... till a collision with an address happens.
             assert!(
                 !buffers.contains_key(&id.id),
                 "IDs collided! Maybe pointing address already occupied this ID."
@@ -368,6 +371,24 @@ mod tests {
         assert_eq!(out.replace().read(), &[3; 10]);
     }
 
+    #[cfg(feature = "cpu")]
+    #[test]
+    fn test_lazy_alloc_later() {
+        use crate::Run;
+
+        let device = CPU::<Lazy<Base>>::new();
+
+        let buf = Buffer::<i32, _>::new(&device, 10);
+        let out = device.apply_fn(&buf, |x| x.add(3));
+
+        device.modules.alloc_later(&device);
+        device.modules.allocated.set(true);
+        assert_eq!(out.replace().read(), &[0; 10]);
+        unsafe { device.run().unwrap() };
+        assert_eq!(out.replace().read(), &[3; 10]);
+
+    }
+
     #[test]
     #[cfg(feature = "opencl")]
     fn test_lazy_apply_fn_with_run_cl() {
@@ -378,9 +399,8 @@ mod tests {
         let buf = Buffer::<i32, _>::new(&device, 10);
         let out = device.apply_fn(&buf, |x| x.add(3));
 
-        assert_eq!(out.read(), &[0; 10]);
         unsafe { device.run().unwrap() }
-        assert_eq!(out.read(), &[3; 10]);
+        assert_eq!(out.replace().read(), &[3; 10]);
     }
     #[test]
     #[cfg(feature = "cpu")]
diff --git a/src/shape.rs b/src/shape.rs
index 986a6e1f..8087fd9b 100644
--- a/src/shape.rs
+++ b/src/shape.rs
@@ -39,10 +39,10 @@ impl Shape for () {
 // TODO: impl for net device
 // this is used to
 /// If the [`Shape`] does not matter for a specific device [`Buffer`](crate::Buffer), than this trait should be implemented.
-pub trait IsShapeIndep: Device {}
+pub unsafe trait IsShapeIndep: Device {}
 
 #[cfg(not(feature = "no-std"))]
-impl<D: PtrConv + Device> IsShapeIndep for D {}
+unsafe impl<D: PtrConv + Device> IsShapeIndep for D {}
 
 /// If the [`Shape`] is provides a fixed size, than this trait should be implemented.
 /// Forgot how this is useful.